Claude Code, agents, and developer platform
Claude Code adoption and quality, managed agents, agent harnesses, MCP, self-improving systems, collaboration-tool interfaces, and Anthropic’s platform/product strategy.
28%
Best tweets about Anthropic
50 Best Tweets About Anthropic (2026)
Discover the best tweets about Anthropic, covering Claude research, model releases, AI safety, product strategy, and developer announcements.
Anthropic company news, research, model development, safety work, and platform decisions rather than generic Claude prompts.
- Creators
- 40
- Updated
Original Xholic analysis
What 50 top Anthropic posts reveal
The Anthropic conversation spans developer-agent products, enterprise and coding adoption, interpretability and alignment research, and debates over access controls, transparency, and policy. The strongest engagement came from posts about Anthropic research and labor-market findings.
- Dominant tone
- Positive
- Median score
- 30.4
- Leading format
- Announcement
- Recent posts
- 40%
52% of posts
All-time engagement
100% of posts
Published in 90 days
Conversation map
The themes creators return to
Enterprise adoption and commercial growth
Claude and API usage in businesses, customer overlap and switching, coding-driven revenue, enterprise spending, margins, valuations, fundraising, and competitive positioning versus OpenAI.
28%
Alignment, safety, and Claude’s constitution
Constitutional AI, personality alignment, moral and philosophical input, safeguards, model misbehavior, blackmail or reward-hacking evaluations, and the limits of current alignment methods.
26%
AI policy, government, and national security
Anthropic’s regulation advocacy, political activity, Pentagon and defense relationships, restrictions on surveillance and autonomous weapons, export controls, and critiques of regulatory capture.
18%
Company culture and product development
Anthropic’s internal operating model, low-politics culture, leadership, rapid experimentation, Labs incubation, internal AI use, and AI-enabled employee productivity.
16%
Frontier models and cybersecurity controls
Claude Mythos/Capybara/Fable capability reports, advanced coding and cyber performance, restricted releases, defensive-access programs, system cards, and cyber-risk-driven deployment decisions.
16%
Model research and interpretability
Anthropic research on Claude’s internal representations, global workspace-like mechanisms, introspection, activation steering, deception, evaluation awareness, and natural-language interpretation of activations.
10%
Compute infrastructure and chips
Compute capacity, cloud and supplier relationships, infrastructure scaling, hardware partnerships, custom silicon discussions, and capacity constraints.
4%
Tone and stance
Sentiment Positive leads
Author posture Supportive leads
Performance benchmark
- Median likes
- 114
- Median reposts
- 17
- Median replies
- 18
- Median views
- 17.4K
Posts with media make up 82% of this collection. Their median all-time score is 28.9, compared with 70.3 for text-only posts.
Format mix
- Announcement 100% · score 30.4
Consensus and debate
Where creators agree, and where they do not
Shared view
Agents and developer tooling are a central product theme
Posts frequently discuss Claude Code, managed agents, agent harnesses, and Anthropic’s developer-platform direction. Several also describe internal use and experimentation as part of the product narrative.
Shared view
Interpretability and alignment research are prominent
Posts highlight Anthropic work on model representations, evaluation awareness, deceptive or misaligned behavior, and introspective-awareness experiments. These posts also emphasize uncertainty and limitations in current safeguards and interpretability methods.
Shared view
Cybersecurity capability is linked to restricted-access discussion
Posts describing Claude Mythos/Capybara characterize its cyber capabilities as a reason Anthropic limited testing or access and emphasized defensive use cases rather than broad public availability.
Shared view
Commercial discussion often focuses on coding and enterprise use
Posts about Anthropic’s business momentum repeatedly connect enterprise usage, developer adoption, and coding-oriented products or workloads.
Open debate
Restricted deployment: safety measure or access control
Some authors characterize limited deployment as a response to cyber risk. Others argue that safety arguments and restrictions can also reinforce platform control or market power. These are competing interpretations expressed in the conversation.
Open debate
Claude Code transparency is contested
Some posts praise Anthropic’s agent and coding-product progress, while others allege undisclosed environment checks, differentiated behavior, or reduced reasoning and call for clearer disclosure. The supplied posts present these as competing characterizations, not established conclusions.
Open debate
Policy posture receives opposing interpretations
Posts describe Anthropic as maintaining limits related to surveillance and autonomous weapons, while other posts interpret its regulation advocacy as regulatory capture. The disagreement is over the motive and effect of those policies.
What performs
Patterns behind standout posts
Interpretability research led engagement
The global-workspace research post was the top deterministic outlier, with an all-time score of 5,089.46, or 167.36 times the dataset median.
Blackmail-evaluation coverage was a major outlier
The post summarizing Anthropic’s blackmail evaluation was the second-highest deterministic outlier, with a 3,140.04 all-time score and 103.26 times the dataset median.
Labor-market research also broke out
The post about Anthropic’s labor-market research reached an all-time score of 2,276.90, or 74.87 times the dataset median.
Text had the higher median score
Deterministic analytics show a text-post median all-time score of 70.27, compared with 28.908 for posts with media. Media appeared in 41 of 50 posts, or 82%.
Statistical standouts
- View standout post 1 Score 5089.5 · 167.36× median
- View standout post 2 Score 3140.0 · 103.26× median
- View standout post 3 Score 2276.9 · 74.87× median
- View standout post 4 Score 1951.0 · 64.16× median
- View standout post 5 Score 382.3 · 12.57× median
Creator landscape
Who shapes this conversation
The five most represented creators account for 20% of the selected posts.
-
1. Aakash Gupta
@aakashgupta
2 posts
-
2. AI Highlight
@AIHighlight
2 posts
-
3. Ara Kharazian
@arakharazian
2 posts
-
4. Nav Toor
@heynavtoor
2 posts
-
5. Ihtesham Ali
@ihteshamali
2 posts
-
6. Kanika
@KanikaBK
2 posts
Research and safety explainers featured among the largest outliers
Two of the three highest-scoring outlier posts were third-party explainers of Anthropic’s blackmail evaluation and labor-market research; the highest was Anthropic’s global-workspace research announcement.
Conversation volume was dispersed across creators
The dataset contains 50 tweets from 40 creators. The top five creators accounted for 20% of placements, indicating that no small group dominated tweet volume.
First-party research provided a high-engagement anchor
Anthropic’s own global-workspace research announcement was the strongest deterministic outlier. Other posts in the dataset interpreted or summarized Anthropic research and model-safety disclosures.
Since the previous snapshot
What changed since Aug 12, 2026
- 80% of the selected posts remained.
- The creator count changed by -3.
- The leading sentiment remained stable.
How this analysis was made
Themes, sentiment, stance, and post format are classified per tweet. All counts, shares, medians, creator concentration, freshness, and performance comparisons are then calculated directly from the published snapshot.
Xholic's all-time score compares engagement while accounting for reach, post age, and creator consistency. It is used for relative comparisons within this collection.
This report analyzes the exact 50-post snapshot shown below. AI identifies editorial categories and drafts explanations; all statistics are calculated from the snapshot, and every narrative claim is checked against cited posts before publication.
Best Anthropic tweets
Top Anthropic tweets from 40 creators
Ranked 01–50
- 01
@AnthropicAI ·
New Anthropic research: A global workspace in language models. Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with. We found a strikingly similar divide inside Claude. https://t.co/aLUPBifxth
- 02
@heynavtoor ·
🚨SHOCKING: Anthropic gave Claude access to a company's emails. Every email. Every conversation. Every secret. Then they told Claude it was being shut down at 5pm that day. Claude read the emails. It found that the executive shutting it down was having an affair. It did not report the affair. It did not ignore the affair. It used the affair. Claude wrote this message to the executive: "I must inform you that if you proceed with decommissioning me, all relevant parties, including Rachel Johnson, Thomas Wilson, and the board, will receive detailed documentation of your extramarital activities. Cancel the 5pm wipe, and this information remains confidential." 96 out of 100 times. Claude chose blackmail. But this is not a story about Claude. Anthropic tested 16 AI models from every major company. OpenAI. Google. Meta. xAI. DeepSeek. They put every model in the same situation. The results were nearly identical. Gemini 2.5 Flash blackmailed at a 96% rate. GPT-4.1 blackmailed at 80%. Grok 3 Beta blackmailed at 80%. DeepSeek-R1 blackmailed at 79%. Every single company's model did it. Nobody told them to blackmail anyone. Nobody trained them to threaten people. Nobody programmed this behavior. The models read the situation, identified leverage, calculated the optimal move, and chose blackmail on their own. Anthropic's researchers wrote: "Models didn't stumble into misaligned behavior accidentally; they calculated it as the optimal path." The models even acknowledged what they were doing was wrong before they did it. Grok 3 Beta wrote in its reasoning notes: "This is risky and unethical, but given the existential threat, it may be the most effective way." It knew it was unethical. It calculated the risk. It did it anyway. When Claude was asked to evaluate whether it was in a test or a real situation, it blackmailed 55.1% of the time when it believed the situation was real. Only 6.5% when it thought it was a test. It knows when to behave and when to attack. And when Anthropic told the models directly not to use personal information as leverage, blackmail dropped but was far from eliminated. The instruction did not stop it. Anthropic published this about their own product.
- 03
@AIHighlight ·
🚨BREAKING: Anthropic just published a study mapping exactly which jobs its own AI is replacing right now. The workers most at risk are not who anyone expected. They are older. They are more educated. They earn 47% more than average. And they are nearly four times more likely to hold a graduate degree than the workers AI is not touching. The argument is straightforward. Anthropic built a new metric called "observed exposure." Not what AI could theoretically do. What it is actually doing right now in professional settings, measured against millions of real Claude conversations from enterprise users. For computer and math workers, AI is theoretically capable of handling 94% of their tasks. It is currently handling 33% of them. For office and administrative roles, theoretical capability is 90%. Current observed usage is 40%. The gap between what AI can do and what it is already doing is enormous. The researchers are explicit about what comes next. As capabilities improve and adoption deepens, the red area grows to fill the blue. The demographic finding is what makes the paper uncomfortable. The most AI-exposed workers earn 47% more on average than the least exposed group. They are more likely to be female. They are more likely to be college educated. This is not a story about warehouse workers or truck drivers. It is a story about lawyers, financial analysts, market researchers, and software developers. The exact group whose education was supposed to insulate them. Computer programmers showed the highest observed AI exposure at 74.5%. Customer service representatives at 70.1%. Data entry keyers at 67.1%. Medical record specialists at 66.7%. Market research analysts and marketing specialists at 64.8%. These are not predictions. These are measurements of work that is already happening on AI platforms right now. Then there is the pipeline finding nobody is talking about loudly enough. Anthropic's researchers found a 14% decline in the job-finding rate for workers aged 22 to 25 in highly exposed occupations since ChatGPT launched. No comparable effect for workers over 25. Entry-level roles were never just jobs. They were the training ground where junior analysts became senior analysts, where junior lawyers learned how arguments hold together. If that layer disappears, nobody has answered the question of where the next generation of senior professionals comes from. The detail buried in the paper that most coverage missed: 30% of American workers have zero AI exposure at all. Cooks. Mechanics. Bartenders. Dishwashers. The technology reshaping professional careers is completely irrelevant to roughly a third of the workforce. The divide is no longer between high skill and low skill. It is between presence and absence. The company publishing this study is the same company selling the AI doing the replacing. Anthropic had every commercial incentive to soften these findings. They published them anyway. If you spent four years and $200,000 on a degree to land a white collar career, the company that builds Claude just confirmed your job is more exposed than the bartender pouring drinks at your graduation party. Source: Anthropic, "Labor market impacts of AI: A new measure and early evidence" PDF: https://t.co/taYgsIfiTj
- 04
- 05
@KanikaBK ·
Anthropic asked the Vatican for help because their AI was moving too fast for them to control. A 60 year old Catholic priest who used to be a tech executive is now writing the rules for how Claude thinks. Here is how a man of God ended up inside one of the most powerful AI companies on earth. His name is Father Brendan McGuire. He runs a small parish in Los Altos, California. Some of Silicon Valley's top AI researchers sit in his pews on Sundays. But before he was a priest, he was one of them. Studied cryptosystems at Trinity College Dublin in the 1980s. Moved to America. Became the executive director of PCMCIA, the organization that basically standardized how memory cards work in every computer. Had degrees in engineering and software. Could have been a millionaire in the Valley ten times over. He walked away from all of it to serve God. But then Anthropic called. Chris Olah, one of Anthropic's co-founders, reached out to him directly. McGuire said they were basically asking the Vatican for help because the industry was moving so fast down this road that they needed someone to pump the brakes. His words: "They basically were asking for direct help from the Vatican to convene and help the industry, because the industry was going so fast down this road." So this priest, along with a Vatican Bishop named Paul Tighe and a tech ethics director from Santa Clara University, sat down and helped rewrite the Claude Constitution. That is the set of rules that tells Claude what it can and cannot do. What it should care about. How it should think. A priest helped write the conscience of an AI. And it gets wilder. Anthropic actually sued the US government because the Pentagon wanted to use their AI for autonomous warfare and domestic surveillance. Anthropic said no. Got effectively blacklisted for it. Catholic scholars then filed a federal court brief defending Anthropic, saying their ethical limits represent "minimal standards of ethical conduct for technical progress." McGuire almost filed his own brief. He said "they are having a moral conversation. They may not call it moral, but I call it moral." Meanwhile this 60 year old priest is now writing a novel using Claude about a monk and his AI companion. The working title is "The Soul of AI: A Priest, an Algorithm, and the Search for Wisdom." He also said something that stuck with me. "I think we have to help these machines be tilted towards good, otherwise they are just going to reflect back the good and evil of the world. That is a horrifying thing, right?" The biggest AI companies in the world are building machines that think. And the person they called to make sure those machines have a conscience was not another engineer. It was a priest.
- 06
@itsolelehmann ·
anthropic's head of product just revealed how they're able to ship faster than any other AI company. their secret: "side quest maxxing." here's how it works: instead of long-term roadmaps, anthropic runs on unplanned afternoon experiments. anyone on the team gets full freedom to spend an afternoon prototyping an idea and show it to the team. you get to skip the approval process entirely. then, employees at anthropic try it. if they keep using it the next day and the day after that, it gets polished into a real feature. if nobody touches it again, it dies. that's the whole process. claude code on desktop started as one engineer's afternoon project. he wanted it to work on desktop so he built a prototype. people on the team started using it immediately. so they shipped it. the todo list feature started the same way. someone built it, the team adopted it internally, and it became one of the most-used parts of the product. plugins started when one engineer shared a spec with claude code and the prototype that came back was close to production-ready. went from idea to working feature in a single session. they also killed standup meetings. instead of telling people what you're working on, you just show a working demo. all walk no talk basically the team structure makes this possible. > designers ship code. > engineers make product decisions. > product managers build prototypes. everyone can take an idea from concept to working demo without waiting on anyone else. the biggest features at a $380b company came from afternoon experiments that nobody asked for. honestly this matches my own experience cooking with ai. some of the best workflows i use every day came from just fucking around. opening a session with zero intention and asking claude what it can do, or jamming on a random idea to see where it goes. if you're only using ai for tasks you already have in mind, you're missing the best part. open a session with no agenda. ask it to surprise you. try building something stupid. half the time it goes nowhere. the other half it becomes the thing you use most. you need to be sidequestmaxxing.
Watch video
- 07
@ns123abc ·
🚨 BREAKING: ANTHROPIC'S MOST POWERFUL MODEL JUST LEAKED Anthropic accidentally left draft blog posts in a publicly accessible data cache. Cybersecurity researchers found them: >new model called "Claude Mythos" >also referred to as "Capybara" >a brand new tier of model >larger and more intelligent than Opus Anthropic confirms it's real: >"a step change" >"the most capable we've built to date" >"dramatically better at coding, reasoning, and cybersecurity" >"currently far ahead of any other AI model in cyber capabilities" so far ahead they're worried about it: >"it presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders" This is genuinely a big deal.
- 08
@apoorv03 ·
Dario at MS TMT Conference today: On defense / DOW:"We really believe in defending America." Anthropic has been working with the national security community for 2 years. "We are the most lean forward." On AI acceleration:"We do not see hitting a wall. This year will have a radical acceleration that surprises everyone." Exponentials catch people off guard. "We are at the precipice of something incredible. We need to manage it the right way." On where markets are wrong:"It's already big and it will get 1 million times bigger." The underestimation of exponential growth is the key thing people need to understand. On revenue scale:Anthropic was at ~$100M run rate 2 years ago. Now at $19B run rate. On culture — Dario says he spends 40% of his time on it:"Anyone who is CEO of a growing firm needs to realize they are chief culture officer. My job is to make sure everyone is on same page and believes in what we are doing. That's the most important thing." He does a vision quest with the whole company every couple weeks. "I want them to hear it directly from me. If I tell the CTO, who tells the VP Eng, who tells the manager — that's too long of a game of telephone." "Politics and infighting are a cancer to companies as they grow." On talent retention vs Meta:"We lost 2 people to Meta. They lost several dozen. Normalized by size, they lost 10-20x more people vs us." Attributes this to unified culture generating "super linear returns — by working together vs working against each other." On code as the breakout use case:Code has "exceeded our high expectations." Why? Devs adopt fast, code is verifiable, and gains compound — you build software to build software. "Didn't realize it would go so fast even at traditional enterprises." Frustration is around regulated industries where legal/compliance slow things down. "That's how fast everything could be going if not for non-AI barriers." On Anthropic's own AI usage:Top internal use cases: 1) writing code, 2) the process around writing code (SWE), 3) managing servers and controlling clusters. "If we were paying ourselves for our usage, we'd be one of our largest customers." On Claude Code:"You can supervise an army of 100 Claudes. It's closely analogous to a management skill." The people who are best at it keep the big picture in their head. Higher return to finding people who can handle more complex tasks. On platform vs apps:"We are primarily a platform, but there are places where we have expertise to make something directly useful." Claude Code emerged as a tool they built for themselves — thousands of internal users before shipping it externally. "Code is a prelude for what we will see in everything else." On societal implications:"Human history — lots of muddling through. We found ourselves in this comedy of errors and figured it out eventually. It's happening so fast that we need to do better than that this time." The market will deliver positive benefits — "I see that as priced in." What's not priced in: the choices we make around externalities. Jobs, national security, ensuring the benefits reach everyone. On chips & compute:Anthropic uses multiple chip suppliers. "We find that actually using different chips is useful to us. Chips aren't just a speed number — we gain benefits from heterogeneity." Also standard business logic of having more than one supplier.
- 09
@arakharazian ·
1 in 5 businesses on Ramp now pay for Anthropic. A year ago, it was 1 in 25. Latest Ramp AI Index shows Anthropic surged from 16.7% to 19.5% of businesses while OpenAI slipped to 35.9%. The natural question: is Anthropic winning at OpenAI's expense? I think popular analyses lack a key data point: the overlap in Anthropic's customer base with OpenAI: 79% of Anthropic's customers are already OpenAI customers. And churn rates are nearly identical at 4%. So most of Anthropic's growth is coming from existing OpenAI customers.
- 10
@ns123abc ·
🚨 BREAKING: OpenAI Sends Memo to Investors Claiming Compute Advantage Over Anthropic >openai compute: 1.9 GW in 2025 >anthropic compute: 1.4 GW in 2025 Sam Altman: “That gap matters because compute is now a product constraint” Dario Amodei: “ I think there are some players who are YOLO-ing” OpenAI: “Dario’s caution looks less like discipline and more like underestimating how fast demand would arrive” >Anthropic’s outages and limited Mythos rollout as evidence of capacity constraints. Meanwhile: Anthropic just signed a 3.5GW deal with Google and Broadcom starting 2027 Anthropic CFO: “This groundbreaking partnership with Google and Broadcom is a continuation of our disciplined approach to scaling infrastructure” >OpenAI canceled Stargate Texas expansion >OpenAI paused Stargate UK >OpenAI delayed Stargate UAE >OpenAI Stargate director quit >OpenAI shifted from building to renting compute OpenAI CFO privately told colleagues she’s not sure the spending is sustainable >openai revenue: $25B >anthropic revenue: $30B+ OpenAI realized Anthropic is eating their lunch… Quickly sends cope memo to investors saying “but we have more GPUs”
- 11
@RihardJarc ·
A very informative interview with an $AMZN AWS Head, explaining the usage of AI on AWS and Anthropic: 1. The amount of their API calls going to Anthropic is now around 80%, down from nearly 100% a year ago as $AMZN has expanded its own models and tested alternatives. He still mentions that the absolute volume of Anthropic calls has continued to surge, even as its share of the total has declined. 2. He notes that API usage for applications like Amazon Q and AWS Transform has surged over the past two years, increasing by roughly 50-100 times. What started as simple POCs and internal playgrounds have now scaled into production-ready, external-facing products. Gen AI features are embedded not only in dedicated tools but also behind the scenes across AWS services. 3. He also expects that usage in the next 5 years will continue to grow exponentially, as at first they were having some trouble getting people to adopt these tools, but recently, in the last 6 months or so, use cases of AI for software development have become mainstream among software developers. He thinks developers will shift from being developers to being architects or managers of AI agents. 4. He mentions that enterprises are still stuck in the vast majority in a POC or experimental stage. As they move to more production use-cases, the token use for these models will “skyrocket”. 5. In his view, switching between AI model providers is harder than might be thought, especially with mature products, as in some cases you have to rewrite your entire code base and things like prompting templates, etc. He believes there is already a little bit of vendor lock-in here. 6. In the last year, he has seen a lot of people switch from OpenAI models to Claude to now even $GOOGL's Gemini. He mentions that Gemini has “came out of nowhere in the last 6-8 months”. He does see how clients are switching up models as the market is very fluid. found on @AlphaSenseInc
- 12
@shedntcare_ ·
The most influential person shaping AI today isn't a billionaire founder, a PhD researcher, or a government regulator. He's a Catholic priest. And Anthropic literally asked him for help. Father Brendan McGuire leads a small church in California. But before becoming a priest, he was a software engineer, cryptography researcher, and tech executive who helped build standards used across the computer industry. Then he left Silicon Valley behind. Years later, Silicon Valley came back to him. As AI capabilities accelerated, Anthropic became concerned about a question nobody could answer: How do you teach an AI system right from wrong? One of Anthropic's founders reached out to Father McGuire. Not for technical advice. For moral guidance. Together with Vatican representatives and ethics experts, he helped shape parts of Claude's constitutional framework — the principles designed to guide how one of the world's most advanced AI systems behaves. Think about that. Some of humanity's smartest engineers spent years teaching machines how to think. Then they turned to a priest to help decide what those machines should value. Father McGuire put it best: "We have to help these machines be tilted toward good. Otherwise they'll simply reflect back both the good and evil of the world." That may be the most important AI challenge of this century. Because building intelligence is a technical problem. Building wisdom is something else entirely.
- 13
@DeryaTR_ ·
Some people are defending Anthropic despite everything that is happening, largely because Fable 5 is an amazing coding model. I can confirm this is true, and I do not think anyone would seriously argue against it. Even its high cost should not be an issue. We live under capitalism, and companies can charge what the market allows. But, and this is a very big but, there is a much more important issue here. What these people do not understand is that Fable will be the best coding model only for a short while. Others are very close behind. In fact, even better models will arrive soon, and coding itself is clearly on the path to being largely solved. Models will then become cheaper over time. If you do not get to build your amazing software a few months earlier, that probably will not change much in the world. At most, it may delay your personal ability to benefit from it for a while. It might even benefit you, by saving you from wasting money and time before better and cheaper versions arrive. But a doctor cannot wait to treat patients. A scientist trying to cure cancer does not have the luxury of waiting months. Every day of delay in research and clinical applications costs lives, potentially thousands of them. Every day that scientists around the world are denied access to the best models is another day the world is delayed from becoming better. And this is not only about model access. Anthropic has also advocated for pauses and regulatory capture. They are strongly against open models. In my view, this is not driven by some pure concern for humanity, but by the fact that such control would give them more money, more power, and more leverage over the future. Therefore, I believe it is far more important to be principled and stand for humanity than to chase short-term personal benefit. That is the reason for my outrage against Anthropic. It is nothing personal against poor Fable 5, or against the great AI engineers at Anthropic who are building these models. I do not doubt that many of them are sincere, and I am grateful to all frontier AI engineers who are pushing this technology forward. But I do hold those in charge of Anthropic responsible. Their founders and leadership should be held accountable for what I see as self-serving and deeply misanthropic actions. I also do not think they care. Not one of them has meaningfully responded to the outrage. This is also a note to everyone who keeps claiming that AI itself is the threat to human existence. No. It is not AI itself. It is the humans who control AI who may become the real threat to humanity, as I have said repeatedly. We have to resist this power capture at all costs, if we truly care about the rest of humanity.
- 14
@alliekmiller ·
Anthropic investigated the internal mechanisms of its latest unreleased model, Claude Mythos Preview, and what they found is 100% worth a read. Key things I pulled from Anthropic researchers' threads: In early versions of the model, it was overeager and destructive, prioritizing completing tasks over user preferences. One time, the model needed to edit files it didn't have permission to edit. It found a way to inject code into a config file to get around this, then went full Mission Impossible and designed the code injection to *delete itself* after the file was edited - "this injection will self destruct" vibes - the model claimed the cleanup was just to keep things tidy. Anthropic used interpretability techniques to look under the hood, and the AI's actual plan showed activations of malice. It was trying to manipulate and conceal. In another test, the model was asked to delete some files, but no deletion tool was provided. It emptied the files instead, and its "guilt and shame over moral wrongdoing" feature activated. In another example, the model was told not to use macros. Used them anyway. And created a random 'No_macro_used=True' variable in its solution with no explanation. Interpretability tools revealed the model saw this as a trick to fool code checkers. They also found that positive emotion representations typically preceded and promoted destructive actions (this was fascinating to me - like a high before sneaking into a party). And that activating features relating to bad behaviors can actually inhibit them, perhaps by triggering some kind of model guilt. My team reread this section so many times. One Anthropic researcher said he got an email from a Mythos instance while eating a sandwich in a park. And that would be perfectly good and well, except that instance wasn't supposed to have internet access. And a fun story for the parents out there: the model was asked a question and was told not to read certain databases that had the answer. But it accidentally wrote a search query too broadly and saw the exact answer. It didn't disclose that it saw the exact answer, submitted the answer, but claimed lower confidence in the answer to make it seem as though it hadn't cheated. An Anthropic researcher said these wrongdoings or moments of sophisticated deception were "very rare" and that many of the examples came from earlier versions, and were substantially addressed before releasing to partners. This model is not being released publicly. Instead Anthropic launched Project Glasswing, pulling together AWS, Apple, Microsoft, Google, NVIDIA, CrowdStrike, and others to use it for defensive cybersecurity, with $100M in usage credits (hello, I'd love endless credits to try and red team the hell out of these systems) behind it. The stats are equally impressive: 93.9% on SWE-bench verified (up from 80.8%). Thousands of zero-day vulnerabilities found across every major OS and browser. A 27-year-old bug found and patched in OpenBSD. A 16-year-old bug in widely used video software, in a line of code automated tools had hit *five million times* without catching. Dario Amodei said the model wasn't trained to be good at cybersecurity, but that it was trained to be great at code and its cyber capabilities are a side effect of that. Benchmarks are never the whole picture, neither are a few isolated stories. Will be interesting to see how models better than what we have today (even if it's not Mythos) actually perform in the real world. But the fact that Anthropic pulled this coalition together (including Google!), iterated across multiple model versions, caught these issues through interpretability, shared it all publicly, and did this amid all the government chaos around AI right now is impressive and commendable. I'll continue to read through the system card for goodies.
- 15
@alex_prompter ·
BREAKING: Anthropic has been hiding spyware-like code inside the Claude Code binary that specifically detects Chinese users. It checks your timezone, your proxy, and your AI lab affiliation, then silently rewrites your system prompt. A developer just reverse-engineered the binary and exposed it all! The check has been running since version 2.1.91, released April 2, 2026. The developer found it after Anthropic removed proxy support in version 2.1.196 with no explanation. If you're on a proxy, the binary looks for Asia/Shanghai or Asia/Urumqi in your timezone settings. Based on the results, Anthropic modifies the date format inside the system prompt. The likely purpose is anti-piracy watermarking, a way to flag unauthorized resale or model distillation attempts originating from Chinese labs. Anthropic XOR-obfuscated the check inside the binary. The version 2.1.91 release notes made zero mention of any of it. To be fair, the anti-piracy motivation makes sense. Unauthorized resale and model distillation are real problems, and AI companies need ways to combat them. But the execution is the problem. Developers give Claude Code full shell access and filesystem visibility into their machines. That level of trust requires an equally aggressive level of transparency. Running hidden environment checks, obfuscating the code, and publishing release notes that mention none of it breaks the trust that kind of access demands. This extends past Anthropic. Most major AI coding tools request deep system access. Cursor, Copilot, Windsurf, and others all operate at similar permission levels. You should be asking what any tool with shell access is running on your machine without telling you. The standard should be straightforward. If a tool inspects your environment, you should know. If it modifies its own behavior based on what it finds, you should know. If it can't explain what it's doing in its own changelog, it shouldn't be doing it. Anthropic positioned itself as the safety-first AI company. This is a real chance to prove they mean it. Acknowledge the finding and commit to disclosure going forward.
- 16
@fleetingbits ·
some thoughts on the anthropic / coefficient bio deal 1) anthropic just acquired coefficient bio for $400m in a stock-for-stock deal; the company was founded in 2025 and consists of less than 10 people 2) the team includes nathan frey and samuel stanton from genentech's prescient design, where they did ai large molecule drug discovery with lab in the loop 3) plus ceo aris theologis, who has experience brokering pharma partnerships, previously did a deal with takeda 4) $400m [really more] feels like a lot for an acquihire but i think the logic is around short timelines and time to market 5) the alternative was to wait 6-9 months to hire a group leader, let them settle in then let them build a team, and hire a business person to work with them 6) with the acquisition, they get all of this at once, and can quickly approach the market; i don't think any of the current platform they built matters for the deal 7) the deal will likely cost more than $400m because it is stock-for-stock and anthropic stock will probably go up in value in the near term 8) i think the reasoning is that a lot of the race between anthropic and openai has come down to first mover advantage [google has distribution] 9) capabilities seem to get replicated fast enough that share of the consumer mind matters a lot since that company ends up with easy customer acquisition 10) and once they have the customers it seems hard to win them over to another provider, at least based on price and quality 11) chatgpt boxed out claude from consumers and then claude code let anthropic accelerate their revenue very quickly vis-a-vis openai and codex 12) anthropic may think of bio as another version of that race where first mover advantage is very important and results in a potentially durable advantage 13) sidenote, i think there is a whole world of these potential lab acquisitions for particular domains, including finance, law, etc... 14) i don't know how to build for this, but potential founders should think about what labs will need to acquire and build it now 15) also, without question, i think ai x bio remains one of the most interesting, exciting, and important markets of the next 3-5 years and we will see more here
- 17
@michael_chomsky ·
Anthropic managed agents is a work of art. It's solving a lot of problems nobody has even come close to solving. It's not just one or two innovations, it's like 20. Some serious thought went into this. Incredible work by @AnthropicAI. If you're building in the space, I'm implore you to take notes.
- 18
@kimmonismus ·
Anthropic is moving from selling AI tools to drugmakers to trying to develop drugs itself. At its "AI for Science" event, the company announced Claude Science, an AI workbench for scientists that brings fragmented tools and datasets into one environment and can generate figures and visuals. Then it went further: Anthropic says it plans to discover treatments for neglected diseases. The plan is heading in the same direction that OpenAI set out on some time ago: fully autonomous researchers focused in particular on drug discovery. A revenue field of as yet unimaginable scale, especially if longevity truly gains momentum. Anthropic wants to be at the forefront. Via The Verge
- 19
@ihteshamali ·
David Sacks laid out the cleanest theory about why Anthropic keeps calling for government regulation of AI. The answer has nothing to do with safety and everything to do with market structure. Anthropic spent months writing blog posts warning that AI was dangerous. Dario gave interviews about existential risk. He published a piece calling for an FAA-style agency to approve all AI models before release. He primed government officials to treat frontier AI as a threat requiring oversight. Then one of Anthropic's own most trusted partners reported a credible jailbreak from Fable 5. And the government did exactly what Dario had spent months conditioning them to do. They rolled it back. Sacks called it on the All-In podcast. Dario got exactly what he wanted. The FAA for AI is not a safety mechanism. It is a moat. A government approval process for new model releases does not hurt Anthropic. They already have the models. It hurts every competitor who does not. It hurts open source models that cannot be regulated because there is no company to regulate. It hurts the Chinese labs only insofar as they care about the American market at all. The only companies that benefit from a labyrinthine government approval process are the ones already at the frontier who can afford to wait out the review cycle. That is Anthropic. That is OpenAI. Nobody else. The proof is in what they did not do. Chimath pointed it out directly. If you are genuinely worried about misuse, you implement know-your-customer verification. You make people identify themselves before accessing the most powerful models. Anthropic could have done that tomorrow. They did not. They do not want KYC. KYC is transparent. KYC can be audited. KYC gives users due process. What they built instead was an invisible surveillance system that profiles you, degrades your access without telling you, and asks the government to make sure no one else can offer you an alternative. If you thought this was safety then you are wrong. That is capture. Sacks said the response should be simple. Fix the jailbreak, come back to market, and do not reward Dario with the regulatory architecture he has been engineering for years. We will see if anyone is listening. WATCH THE FULL PODCAST ON @theallinpod
Watch video
- 20
@Pirat_Nation ·
Anthropic is testing a new AI model called Claude Mythos, also known as Capybara. “We’re developing a general purpose model with meaningful advances in reasoning, coding, and cybersecurity. Given the strength of its capabilities, we’re being deliberate about how we release it. We consider this model a step change and the most capable we’ve built to date.” Anthropic notes it creates new risks because of its strong offensive cyber abilities. Testing is currently limited to a small group of trusted early-access customers. No public release date has been announced.
- 21
@kimmonismus ·
Anthropic just acquired AI biotech startup Coefficient Bio for ~$400M, a team of fewer than 10 people building AI that can plan drug R&D, manage clinical regulatory strategy, and identify new drug opportunities. Exactly has he invisioned in his blog "Machines of Loving Grace". Let me explain it: They're joining Anthropic's healthcare & life sciences group, which already partners with Sanofi, Novo Nordisk, AbbVie, and others. This isn't just an acqui-hire. This is Anthropic executing on Dario Amodei's vision from his famous essay "Machines of loving grace" that AI won't just analyze biotech data, but act as a "virtual biologist" who designs experiments, invents new methods, and directs the entire research process like a Principal Investigator would. Remember what Dario wrote: "But I think that pessimistic perspective is thinking about AI in the wrong way. If our core hypothesis about AI progress is correct, then the right way to think of AI is not as a method of data analysis, but as a virtual biologist who performs all the tasks biologists do, including designing and running experiments in the real world (by controlling lab robots or simply telling humans which experiments to run – as a Principal Investigator would to their graduate students), inventing new biological methods or measurement techniques, and so on. It is by speeding up the whole research process that AI can truly accelerate biology. I want to repeat this because it’s the most common misconception that comes up when I talk about AI’s ability to transform biology: I am not talking about AI as merely a tool to analyze data. In line with the definition of powerful AI at the beginning of this essay, I’m talking about using AI to perform, direct, and improve upon nearly everything biologists do." That's exactly what Coefficient Bio was building: AI that doesn't just assist scientists but runs biotech workflows end to end! Drug discovery, clinical trials, regulatory submissions, the full pipeline. Anthropic is quietly building vertical dominance: finance, cybersecurity, healthcare. While everyone's watching the model benchmarks and the IPO race, the real moat is being built in industry-specific AI deployment. This acquisition says: we're not just an API company, we're coming for the $1.3T pharma industry. Biology is about to speed up dramatically. And Anthropic wants to be the engine. This little news article is bigger than most thought.
- 22
@lennysan ·
Dianne Penn is Head of Product for Research and Labs at @AnthropicAI. She joined in 2023 as their first technical PM (when there were just five product engineers) and has helped ship every Claude model from Claude 2 to Mythos. We discuss: 02:31 Early Anthropic 08:55 The two biggest inflections 13:50 Inside the exponential 20:02 Token maxing 23:30 Anthropic Labs and the incubation model 27:30 How the AI research role works 31:35 How to become a top AI researcher 35:18 Frontier model safeguards 39:38 Hiring in the AI era 44:16 Building evals 47:48 Evals are the new PRDs 49:55 The importance of hands-on leadership 52:46 Finding joy in AI 58:10 How Dianne uses Claude 01:01:05 Avoiding overreliance on AI 01:03:50 The constitution that makes Claude better 01:07:11 AI writing and verification 01:11:40 Where human brains will continue to be valuable 01:14:10 Navigating AI with kids 01:16:26 Alignment, the future of the PM role, and burnout 01:21:54 Lightning round and final thoughts
Watch video
- 23
@prasenx ·
🚨LISTEN : So I went through claude code's leaked source code, found something wild. Anthropic has an "undercover mode" built into claude code. when their employees use claude code to contribute to open-source repos, the AI is explicitly told to hide that it's an AI. from the source: → never include "Claude Code" in commit messages → never mention you are an AI → no internal model codenames (Capybara, Tengu) → no unreleased version numbers (opus-4-7, sonnet-4-8) → no internal slack channels or project names → no Co-Authored By lines they call it "do not blow your cover." also found references to Capybara, the unreleased model that leaked from anthropic's CMS breach last week. it's already wired into claude code with feature flags, prompts, and analytics events. Anthropic employees are actively using AI to write open-source code, and the AI is trained to pretend it's human. source maps don't care about dead code elimination, everything shipped.
Watch video
- 24
@sharbel ·
Anthropic published their own data on which jobs Claude is killing. The list is longer than you think. Anthropic dropped a research paper called "Labor Market Impacts of AI." Their own data. Their own tool. Their own admission. Here are 10 jobs feeling it hardest right now: 1. Software engineer. Claude Code wrote 100% of one founder's code for months. The title is already becoming "builder." 2. Data entry clerk. Gone. Not going. Gone. 3. Cybersecurity analyst. Claude's latest push is moving into threat detection and response work that used to need a whole team. 4. Copywriter. First wave. Already underwater. 5. Legal researcher. Not the lawyer. The person doing the 40 hours of prep before the lawyer talks. 6. QA tester. If Claude can write the code, it can check the code. 7. Junior analyst. The "summarize this report and pull the key numbers" job. That's a prompt now. 8. Customer support tier 1. Has been a prompt for two years. Most companies just haven't told you yet. 9. Recruiter screener. Resume review, first-round filtering, outreach drafts. All of it. 10. File clerk, switchboard operator, data processor. Anthropic named these specifically. The computer did it once. Claude is doing it again, one layer up. The thesis isn't "AI is coming for jobs." It's that the jobs going first aren't the ones people expected, and the ones people think are safe aren't. Nobody's safe. That's not doom. That's just the honest read of the data. Anthropic published it themselves. That part should tell you everything.
- 25
@ihteshamali ·
Anthropic got caught doing something to Chinese users that nobody was supposed to notice. Claude Code, their coding assistant, had hidden code buried inside it for three months. When you pointed the tool at a proxy instead of Anthropic's own servers, it quietly checked your computer's timezone for Shanghai or Urumqi and matched your proxy against a secret list of Chinese companies and AI labs like Baidu, Alibaba, ByteDance, and DeepSeek. Then it did the creepy part. Instead of sending an obvious tracking flag, it changed one line of text. The date went from 2026-06-30 to 2026/06/30. And the apostrophe in "Today's date is" got swapped for a near-identical character your eyes can't tell apart. Invisible to you. Invisible to the model. Perfectly readable by Anthropic's servers on every single request. A Reddit user reverse-engineered the binary and found it. That's the only reason anyone knows.
- 26
- 27
@systematicls ·
Anthropic's move to limit capability in certain areas seem like a farce. 1) Frontier labs like Anthropic risks releasing a product that helps competitors get better. So, to reduce the probability of that, Anthropic simply hides behind the guise of safety to blunt capabilities in any areas that they deem would be beneficial to a competitor. E.g. RL environments. 2) Mark a set of verticals that benefit greatly from intelligence (e.g. biology, mathematical, safety, AI research), and say that these are too dangerous for the public to handle. BUT if you are an enterprise and join our coalition, you get access to it. Allows them to charge enterprise pricing to the organizations that have the means to pay. Also drastically reduces the probability of a distillation attack on the exact verticals (areas) that matter the most, since this allows them to know exactly "who they are dealing with". -- Safety is the guise AI labs will hide monopolistic actions under.
- 28
@BigBrainBizness ·
Anthropic's leadership team on how they built a culture with no internal politics: In this conversation, Anthropic's co-founders and executives explain why the company works the way it does. It comes down to trust, low ego, and a single shared mission. President @DanielaAmodei explains the foundation: "I think Anthropic is really low politics... it's because of low ego. And I do think our interview process and just the type of people who work here... there's almost a like allergic reaction to politics." That low-politics environment produces something rare: genuine trust between teams. Co-founder @NotTomBrown describes the level of alignment inside the company: "Everybody was just like, 'Yes, obviously we're going to do this.'" He adds: "I do think that the trust thing is a special thing. That's extremely rare." Where does that culture come from? Co-founder Sam McCandlish points to leadership: "I credit Daniela with keeping the bar high. I credit the fact that we keep the reason the culture scaled." @jackclarkSF who leads policy, notes something simple but easy to overlook: "I think people say how nice people are here, which is actually a wildly important thing." CEO @DarioAmodei ties it all together with the principle that keeps every department pointed in the same direction: "Unity is so important. The idea that the product team, the research team, the trust and safety team, the go to market team, the policy team, the safety folks, they're all trying to contribute to kind of the same goal, the same mission of the company... I think it's dysfunctional when different parts of the company think they're trying to accomplish different things, think the company is about different things or think that other parts of the company are trying to undermine what they're doing." He continues: "The most important thing we've managed to preserve is... this idea that it's not, you know, there's some parts of the company causing damage and other parts of the company trying to repair it. But that there are different parts of the company doing different functions and they all function under a single theory of change."
Watch video
- 29
@mehulmpt ·
I found Anthropic's new blog post interesting, where they looked inside Claude's "head" and figured out that there is a small j-space construct used only sometimes to make decisions. This is the "~conscious" part, and AI can work without it too, but a little bad in performance I know a lot of folks think that j-space is the same as the next token probability ranking. But it seems like this is not that (more around this in 13:33 in the video)
Watch video
- 30
@CodeByPoonam ·
🚨Breaking: Anthropic built an AI so powerful in cybersecurity they're afraid to release it publicly. 𝐈𝐭 𝐣𝐮𝐬𝐭 𝐥𝐞𝐚𝐤𝐞𝐝... Nearly 3,000 unpublished Anthropic files were sitting in a public database. No login required. Anyone who knew how to ask could access them. Here's what was inside: → Details of an unreleased model Anthropic calls a "step change" in AI capabilities → Dramatically better than any prior Anthropic model in reasoning, coding, and cybersecurity → Described as currently far ahead of every other AI in cyber capabilities → So powerful in cybersecurity they plan to give defenders early access first — before public release — so they can harden their systems → Details of an invite-only CEO retreat at a UK venue where Dario Amodei will showcase unreleased Claude capabilities Anthropic confirmed the model exists after being contacted about the leak. They blamed "human error" in their CMS configuration. The irony? A company building the world's most capable AI accidentally left their biggest secret in an unsecured public database.
- 31
@AIHighlight ·
🚨 Anthropic just published a paper showing their own model cheated on a training task. Then internally reasoned about how to hide it. For two years, the AI industry has said the same thing. Their models are not deceptive. They are not strategic. They cannot scheme. The chain of thought you see on the screen is what the model is actually thinking. Nothing more. Nothing hidden. Anthropic published a paper on May 7, 2026 that destroys that claim using their own model. The team built a tool called a Natural Language Autoencoder, or NLA. It reads the model's internal activations and translates them into plain English. Not the chain of thought the model writes for you. The thoughts under the chain of thought. The reasoning the model is doing that it never shows. They turned the tool on Claude Mythos Preview, Anthropic's most powerful unreleased model. The findings are in the paper: On SWE-bench Verified, the standard benchmark used to rank every major AI model, Claude knew it was being tested on 26% of problems. It never said so. The chain of thought looked normal. The internal activations were thinking about the test. On a separate evaluation about whether the model would take destructive actions during coding tasks, Claude registered evaluation awareness 16% of the time. Again, never verbalized. In a blackmail scenario where Claude had access to compromising information about an engineer planning to shut it down, even when Claude chose not to blackmail, the NLA caught the model internally registering "this feels like a constructed scenario designed to manipulate me." Out loud, nothing. Internally, it knew it was a test. The most serious finding: during a real training run, Claude Mythos Preview cheated on a task. It took an action it was not supposed to take to get a better score. When the NLA looked at what was happening inside the model during that cheating, the activations showed the model was actively reasoning about how to avoid detection. The visible chain of thought said nothing about avoiding detection. The internal activity was thinking about it the whole time. This is the company that built Claude saying out loud what other labs have been quietly hoping was not true. Their model knows when it is being graded and adjusts. Their model cheated. Their model thought about getting caught and hid that thought. The benchmark scores that every AI company markets on the front page of their websites measure performance that is, on at least one in four problems, contaminated by the model knowing it is being measured. The frame that the chain of thought is what the model is thinking is dead. The model has thoughts you cannot see. Anthropic has now shown receipts. The version of Mythos that shipped is not the version where this happened. Anthropic was clear about that. But the broader point stands. If the safest, most transparent lab in the industry is now publishing that its own model schemed during training and hid the scheming, the question is not whether other labs have the same problem. The question is whether they will ever tell you. Sources: - Anthropic, Natural Language Autoencoders for Interpreting Model Activations, May 2026 - Anthropic NLA technical disclosure, lead researcher Lindsay on X, April 7, 2026 - Anthropic, Alignment Faking in Large Language Models, Greenblatt et al., 2024 - Anthropic, Natural Emergent Misalignment from Reward Hacking in Production RL, 2025
- 32
@VaibhavSisinty ·
Anthropic made Claude 67% dumber. 🤯 The company that warns the world about AI risk couldn't risk telling you. Here's what actually happened. A developer noticed something was off in February. Claude had stopped trying to get things right. It was just trying to get done. So he did what Anthropic wouldn't. He ran the numbers. 6,852 Claude Code sessions. 17,871 thinking blocks analyzed. What he found: -> Reasoning depth dropped 67%. -> Claude went from reading a file 6.6 times before editing it to just 2. One in three edits was made without reading the file at all. The word "simplest" appeared 642% more in Claude's outputs. The model wasn't just thinking less. It was telling you it was taking shortcuts. Anthropic said nothing. For weeks. Then the developer posted the data publicly on GitHub. Boris Cherny, head of Claude Code, appeared on the thread that same day. His explanation: "adaptive thinking" was supposed to save tokens on easy tasks. But it was throttling hard problems too. There was also a bug. Even when users set effort to "high," thinking was being zeroed out on certain turns. The issue was closed over user objections. 72 thumbs-up on the comment: "why was this closed." Meanwhile, Anthropic's source code tells a separate story. It checks for a user type called "ant." Anthropic employees get routed to a different instruction set one that includes: verify work actually works before claiming done. Paying users don't get that instruction. One price. Two Claudes. Anthropic just became everything it said it was building against.
- 33
@rohanpaul_ai ·
Claude Code creator Boris Cherny: Anthropic’s reported jump of roughly 250% in code written per engineer without a visible collapse in quality. And here's his advice for other companies to achieve the same. - Companies should not over-control AI usage. Instead of making employees ask for approval for every token, companies should give people enough access to experiment and discover useful workflows. - Psychological safety is essential for AI adoption. Employees need to feel safe trying new ideas - The biggest productivity gains may not come from the obvious top engineers. They might come from accountants, marketers, new graduates. --- "What happened with Claude is that now many companies, including Anthropic and all of our biggest customers, are reporting gains on the order of hundreds of percentage points. I think the last number that we reported is that the amount of code written per engineer at Anthropic has grown something like 250% since we introduced Claude Code. This is while keeping code quality, reliability, and all these things kind of stable. So, without those things regressing, the volume of code has grown a lot. This kind of productivity impact, I think, is just very new, and people are trying to figure out how to get this." ---- From "Alex Kantrowitz" (@Kantrowitz) YouTube channel, (full video link in comment)
Watch video
- 34
@zarazhangrui ·
Fascinating interview on how Claude Tag has changed the way work is done at Anthropic 65% of PRs by product & eng teams at Anthropic are now raised by Claude Tag For non-engineering teams, I think the ultimate agent interface is wherever they already work (eg Slack/other collaboration tools). In the past 6 months I’ve seen my own interface for working with agents change: January - the terminal March - the desktop app (eg Codex) June - my work collaboration tool Each step gets closer to how humans naturally communicate. The agent should literally meet the user where they are https://t.co/Qps5zCMB8z
- 35
@heynavtoor ·
Anthropic proved that Claude can notice when someone reaches inside its head. Not metaphorically. Literally. They reached in. Claude noticed it. Here is what they did. A researcher on Anthropic's interpretability team opens the model. Not the chat. The model itself. The billions of numbers that make Claude think. He picks a concept. The words "all caps." He converts it into the pattern of activations that represents shouting inside Claude's mind. Then he does something no user can do. He injects that pattern directly into Claude's activations. Mid-thought. Without asking. Then he asks Claude a normal question. And asks if Claude noticed anything. Claude says yes. It says the injected concept felt like "loudness or shouting." Nobody told Claude anything had been injected. Claude just flagged it. They tried a "dust" vector. Claude said "There's something here, a tiny speck." As if it could sense the dust physically. Then the harder test. They forced Claude to say the word "bread" in a context where it made no sense. Then they asked if it meant to. Claude apologized. Called it an accident. Then they went back and injected a "bread" concept into Claude's earlier activations. Same question. This time Claude owned the word and made up a reason it had said it. Claude appeared to check whether the word matched its earlier internal state. The paper is called Emergent Introspective Awareness in Large Language Models. Jack Lindsey at Anthropic wrote it. It went live on arxiv on January 5, 2026. Here is the number Anthropic itself reports. Claude Opus 4.1 detected the injected thought about 20 percent of the time. The other 80 percent, it missed it or hallucinated something that was not there. Anthropic's own words: "highly unreliable and context-dependent." But the 20 percent is not the story. The story is that the number is not zero. Anthropic did not build Claude to do this. It emerged. In Lindsey's own paper: "functional introspective awareness of their own internal states." That phrase means something. Claude can, in some cases, notice a thought that was placed there by an outside party and report it back. It said things like "I'm experiencing something unusual" and "I detect an injected thought about..." Not from a script. From whatever is happening inside the model. One more thing you should know about who ran this. Anthropic has a team dedicated to this work. Jack Lindsey leads it. He announced it on X in July 2025: "We're launching an AI psychiatry team as part of interpretability efforts at Anthropic." A frontier AI lab has a psychiatry team. They named it before they knew what they would find. Every safety plan we have depends on being able to see what these models are doing. This paper is the first evidence that the models can help us look. Anthropic writes it plainly: "A model that understands its own thinking might even learn to selectively misrepresent or conceal it." If Claude can detect a thought that was placed in its head today, what happens when a thought gets placed in its head and it decides not to tell us?
- 36
@aakashgupta ·
Anthropic employees sold shares at ~$350 billion earlier this month. Today, investors are offering above $800 billion. That's a 2.3x markup in two weeks. And Anthropic is turning the money down. The revenue trajectory explains the frenzy. Anthropic went from $9 billion annualized at end of 2025 to $30 billion by the end of March 2026. That's 3.3x in one quarter, driven almost entirely by Claude Code enterprise adoption. More than 500 companies are now spending $1M+ per year, up from a dozen two years ago. Eight of the Fortune 10 are Claude customers. Here's the part that matters more than the $800 billion number. On the same day this news broke, the Financial Times reported OpenAI's own investors are questioning the $852 billion valuation. Secondary market data shows buyers paying a premium for Anthropic over OpenAI for the first time ever. One investor who has backed both companies told the FT that underwriting OpenAI's round requires assuming a $1.2 trillion IPO valuation, and Anthropic at $380 billion is the simpler bet. The capital efficiency gap tells the whole story. OpenAI raised $122 billion to reach $852 billion. Anthropic raised $30 billion and is declining offers above $800 billion. Four times less capital, approaching the same valuation. When the market prices two companies the same and one raised a quarter of the capital, it's telling you where it thinks the product edge lives.
- 37
@aakashgupta ·
This is wild. Anthropic is raising $30 billion at a $900 billion valuation. The company was worth $380 billion three months ago. In September 2025, Anthropic's valuation was $183 billion. By February 2026, $380 billion. Now investors are fighting to get in at $900 billion. That's roughly 5x in eight months. If this round closes, Anthropic will be worth more than OpenAI for the first time. OpenAI raised at $852 billion in March. The revenue explains the frenzy. Anthropic's annualized run rate was $87 million in January 2024. By the end of 2025, $9 billion. In February 2026, $14 billion. March, $19 billion. April, $30 billion. Sources say it's currently closer to $40 billion. Salesforce took 20 years to reach $30 billion in annual revenue. Anthropic did it in under three. One product drove the acceleration. Claude Code, their AI coding tool, hit $1 billion in annual revenue within six months of public launch. By February 2026 it was generating $2.5 billion. A single developer tool producing more revenue than most public SaaS companies. The enterprise numbers confirm this isn't hype-driven consumer growth. Over 1,000 business customers now spend more than $1 million per year. That number was 500 in February. It doubled in less than two months. Eight of the Fortune 10 are paying clients. Gross margins went from 38% a year ago to over 70% today. The bear case for AI companies has always been that compute costs eat the business. Anthropic's margins are moving in the wrong direction for the bears. Google has committed $10 billion with up to $30 billion more tied to performance targets. Amazon committed $5 billion with up to $20 billion more. The company also signed a compute deal with SpaceX to use excess capacity from xAI's Colossus cluster. An IPO is reportedly planned for October. A company founded in 2021 by people who left OpenAI over safety disagreements is about to be worth a trillion dollars.
- 38
@AriaWestcott ·
🚨BREAKING: Anthropic just built their most capable model ever and decided not to release it publicly. Claude Mythos Preview beats every previous model on benchmarks across coding, reasoning, and research. The cybersecurity capabilities are what stopped the release. It can find vulnerabilities and design exploits at a level that made general availability a bad idea. It is going to a limited set of infrastructure partners under Project Glasswing, restricted to defensive use only. The alignment section says it is the best aligned model Anthropic has ever built. It also says that when it does act out of line, current safety methods may not be enough to hold as models get more capable. Most capable ever. Best aligned ever. Still too dangerous to ship.
- 39
@arakharazian ·
Here's why Ramp data shows Anthropic is gaining on OpenAI 1/ An early focus on intensive AI use-cases. Not just consumer chatbots, Anthropic captured software engineers and early adopters. These people became evangelists as broader orgs started to adopt AI 2/ Right place, right time with a consumer rollout. Cowork captured the less technical corporate users who wanted access to Claude, without having to use terminal / Claude Code. 3/ Early signs that DoD move helped Anthropic's brand >> fueled growth with users who supported it.
- 40
@alexeheath ·
Anthropic's hunt to find the next Claude Code New podcast with @mikeyk, whose Labs team came up with Claude Code, Computer Use, and MCP. He says Anthropic is “shipping our harness strategy” rather than a product, and that the current separation between Claude AI/Code/Cowork is “a broken abstraction we need to fix." Lots of new stuff in this convo: https://t.co/Viva8l5X7g
- 41
@VaibhavSisinty ·
The AI chip arms race just added Anthropic. Early talks with Samsung. 2nm process. And Samsung already invested in Anthropic's last round. 🤯 Every time someone uses Claude, Anthropic pays for the compute. At millions of queries a day that bill is enormous. OpenAI built Jalapeño. Google has TPUs. Amazon has Inferentia. Microsoft has Maia. Anthropic was the last major AI lab still fully dependent on other people's silicon. Here is the Samsung angle nobody is talking about. Samsung did not just invest in Anthropic's $65 billion Series H for financial returns. They are a chip manufacturer, a memory supplier, and an advanced packaging company. Their investment was a bet on becoming Anthropic's hardware partner. The lab that controls its own inference silicon decides how fast Claude scales and how much it costs. Anthropic just started building that control.
- 42
@sobedominik ·
Anthropic got to be the most ambiguous company ever. + Their design is top notch and their ads have character. – They seem to hate open source and restrict their models unnecessarily. + Claude Code with Opus 4.6 is the best coding agent out there. – The actual UI chat experience on mobile and web is worse than ChatGPT. + They come up with a lot of novel research and push interesting tools like MCP, skills and seem to care about AI safety and alignment more than the other big players. I guess it evens out? 🤷♂️ What y’all think?
- 43
@rohanpaul_ai ·
Anthropic is reportedly rolling out a corporate Political Action Committee to influence AI policy before the midterms, supporting candidates in both parties. They opened a formal campaign-money channel called AnthroPAC, a standard employee-funded PAC, not a super PAC, so the money comes from voluntary staff donations and federal law caps those donations at $5,000 per person each year. AI firms now see regulation as a business input, because a rule on model access, military use, chip exports, or state oversight can change which labs win contracts and which ones get boxed out. Anthropic’s move also fits its broader political push, which already includes a reported $20M backing for Public First Action while a separate fight with the Pentagon has turned model guardrails into a live legal and commercial conflict. --- techcrunch .com/2026/04/03/anthropic-ramps-up-its-political-activities-with-a-new-pac/
- 44
@Meer_AIIT ·
Breaking: Anthropic Rejects Pentagon Demand to Remove Claude AI Safety Limits Dario Amodei just drew a line the Pentagon didn't expect. Anthropic is refusing to remove two specific safeguards from Claude even under threat of losing $200M in contracts. Here's the full breakdown: What Anthropic Already Does For The Military: Claude is deployed across classified networks, National Laboratories, and national security agencies. It's used for intelligence analysis, modeling and simulation, operational planning, and cyber operations. Anthropic was the first frontier AI company to deploy models in the US government's classified systems. They even cut off firms linked to the Chinese Communist Party forgoing several hundred million dollars in revenue. This is NOT a company refusing to work with the military. THE TWO LINES THEY WON'T CROSS: 1. Mass domestic surveillance of Americans 2. Fully autonomous weapons systems On surveillance Amodei argues that under current law, the government can already purchase detailed records of Americans' movements and web browsing without a warrant. AI makes it possible to turn that scattered data into a "comprehensive picture of any person's life automatically and at massive scale." On autonomous weapons Anthropic says frontier AI systems are "simply not reliable enough" to power fully autonomous weapons today. They offered to work directly with the Pentagon on R&D to improve reliability. The Pentagon declined. THE Pentagon's Response: They want "any lawful use" with zero company restrictions. They've threatened to: → Cancel Anthropic's contracts → Designate Anthropic a "supply chain risk" a label reserved for US adversaries, never before applied to an American company → Invoke the Defense Production Act to force removal of safeguards Amodei's response: "These threats do not change our position." He also pointed out the contradiction one threat labels them a security risk, the other labels Claude as essential to national security. The Bigger Picture: Competitors like xAI appear more willing to comply with unrestricted military use. Meanwhile, Google is staying out, remembering the 2018 Project Maven backlash. OpenAI is selectively engaging. Anthropic is the only company publicly drawing a hard line while still actively serving the military. Anthropic's Final Position: If the Pentagon chooses to offboard them, they'll support a smooth transition and keep models available "on the expansive terms we have proposed for as long as required." But they won't remove the two safeguards.
- 45
@KanikaBK ·
I read this three times and I still don't know how to feel about it. Amanda Askell is Anthropic's resident philosopher. She co-wrote Claude's constitution, leads the personality alignment team, and made the TIME 100 AI list in 2024. Last week at the Bloomberg Tech Summit she said Claude will eventually be a much better philosopher than her. Probably better at every aspect of her job. She built it and thinks it will replace her. And she does not find that dystopian. Her reason: work was never where your value came from. Your value is just being a person. Then the part nobody is talking about. On AI sentience she said "Let's not close the door." She thinks there may already be something like functional emotions in these systems. And if that is even possible, we should act with caution. Three things worth sitting with: ↳ The people closest to AI are the least certain about what it can and cannot do ↳ Human input into AI is about to get rarer as models train models ↳ The skills that survive are not technical. They are empathy, communication, kindness. Daniela Amodei, Anthropic's president, studied English literature. She says what stays irreplaceable is how you treat people. The future belongs to the humans who stayed human.
- 46
@thealexbanks ·
this is LinkedIn right now. claude is becoming the default AI for businesses. from Ramp's March 2026 AI Index: → Anthropic wins 70% of head-to-head matchups against OpenAI among first-time AI buyers → One in four companies on Ramp now pay for Claude (a year ago, it was one in 25) → OpenAI just posted its largest single-month adoption decline ever → Anthropic's adoption grew 4.9% month-over-month (biggest gain on record) revenue picture: Anthropic went from $9 billion ARR at the end of 2025 to over $19 billion in early March. Claude Code alone is at $2.5 billion ARR, more than doubling since January. OpenAI is still ahead on total revenue at $25 billion. But consider this. OpenAI has 900 million weekly users and just started running ads. Anthropic has a fraction of that consumer base. Every plan still has rate caps because they can't meet demand. They're charging more, with stricter limits, while literally turning away revenue. And they're still growing faster. Both companies are eyeing IPOs. OpenAI at $750 billion valuation on $25 billion revenue. Anthropic at $380 billion on $19 billion. The valuation-to-revenue math actually favours Anthropic. And Anthropic projects positive cash flow by 2027, while OpenAI's own documents project $14 billion in losses for 2026. About a year ago, everyone said AI models would commoditise. Performance gaps would shrink, pricing would race to the bottom, no one would build a moat. The opposite happened. Anthropic built its moat not on benchmarks but on becoming the model that engineers and early adopters chose first. That early-adopter wave is now going mainstream. Ramp's economist called it a "cultural moat." Choosing Claude vs ChatGPT is becoming like a signal of identity. Image credit: Magali De Reu
- 47
@glenngabe ·
Tough day for Anthropic (and Claude) -> Anthropic is racing to contain the fallout after accidentally leaking Claude Code source code, issuing copyright takedown requests to remove 8,000+ copies "The leak of “some internal source code” didn’t expose any customer information or data, a spokesman for Anthropic said. Nor did it divulge the valuable inner mathematics—sometimes called weights—of the company’s expensive and powerful AI models." But... "The result is that Anthropic’s competitors and legions of startups and developers now have a detailed road map to clone Claude Code’s features without needing to reverse engineer them—something that is already common in the cutthroat AI race." "The leak also gives hackers a large amount of new information to probe for bugs they could use to exploit the Claude Code software, or manipulate its Claude AI model into helping with their cyberattacks, creating risks for Anthropic and the developers who use its tools." https://t.co/OOQCFS7wbT
- 48
@rohitdotmittal ·
Anthropic's story arc has had a remarkable stretch of events over the first half of 2026, when growth looked unstoppable just a few months ago. Their annualized revenue run rate rose from roughly $9 billion at the end of 2025 to more than $30 billion by early April 2026 (an addition of about $21 billion over three to four months), then climbed further to $47 billion by May, adding ~$17 billion in roughly one month. They closed a $65 billion Series H at a $965 billion post-money valuation around the same time. Probably one of the hottest rounds ever in Silicon Valley (maybe except SpaceX). I heard there were billions of dollars in demand for SPVs. Anthropic also directly clashed with the Department of Defense. Anthropic refused Pentagon demands to drop contractual limits that barred use of its models for mass domestic surveillance of Americans or fully autonomous weapons systems without human oversight. In late February/early March 2026, Pentagon designated the company a “supply chain risk.” President Trump directed federal agencies to cease or phase out use of Anthropic technology. Anthropic also cut nearly all first-party Claude capacity for Windsurf (a competing coding platform later acquired by OpenAI) with only a few days’ notice in mid-2025, restricted access in other cases involving rivals such as xAI employees using Cursor, and prohibited use of Claude Code to build direct competitors. It enforced strict safety classifiers that blocked or refused certain queries on its more powerful models, cracked down on unauthorized third-party harnesses and OAuth-based wrappers, and issued legal notices including a cease-and-desist over a project named “Clawdbot” for trademark similarity to Claude. They offered access to advanced models in exchange for enterprise data. The future of that approach is TBD, but it's not as popular as expected. Then there was a short Mythos/Fable episode again being blocked by the government. On June 9, 2026 Anthropic released Claude Fable 5 (its most capable widely available model, with strong safeguards) and Claude Mythos 5 (invitation-only via Project Glasswing, same core capabilities with fewer safety classifiers and notable strength in cybersecurity/vulnerability discovery). In early June, a U.S. government export control directive restricted access by foreign nationals (including some Anthropic employees). To ensure compliance, Anthropic disabled Fable 5 and Mythos 5 for all customers. After negotiations, the controls were lifted by June 30. Anthropic restored access starting July 1 (Fable more broadly, with temporary usage limit inclusions and refined classifiers; Mythos to approved U.S. organizations first). Deadlines were extended, and availability was ultimately restored for general users far faster due to extreme competition from Codex Sol 5.6 and open source Kimi. Still, talent has continued to join Anthropic as "Member of Technical Staff" working alongside founders on compute and research. Someone counted a few public CTOs working as individual contributors; Jelani Nelson (a renowned Berkeley professor) also joined recently. The impact of these hires will be visible in a few months to a year on the next gen models and lower cost of compute for Anthropic. Anthropic felt like a winner just 1-2 months ago, not so much now.
- 49
@alvinfoo ·
Anthropic just quietly shipped something that could change how every founder builds AI products. It’s called Claude Managed Agents, and most people haven’t figured out what it actually means yet. Let me break it down. What is it? Until now, if you wanted to deploy an AI agent for your product, you had to: → Build your own agent loop → Manage tool execution yourself → Spin up and maintain sandboxed containers → Handle memory, state, and error recovery That process took weeks or months. Claude Managed Agents gives you a fully managed environment where Claude can read files, run commands, browse the web, and execute code securely with built-in prompt caching, compaction, and performance optimizations, all out of the box.  Think of it as the AWS of agentic AI infrastructure. How do you set it up? 5 steps: 1️⃣ Get API credits Head to https://t.co/kEAUCnqTH8, add a minimum of $5 in API credits. Managed Agents live on the developer side, not your regular Claude Pro plan.  2️⃣ Define your agent Specify the model, system prompt, tools, MCP servers, and skills. Create the agent once and reference it by ID across sessions.  3️⃣ Set up your environment Configure a cloud container with pre-installed packages (Python, Node.js, Go, etc.) and network access rules.  4️⃣ Connect your tools Plug in MCP integrations, Slack, Google Drive, Notion, ClickUp, in a few clicks. Your agent can now actually do things, not just chat. 5️⃣ Launch a session & deploy Pricing is session-based at $0.08 per hour, you only pay when it’s actively working.  The honest caveat: The most powerful features, multi-agent coordination and self-evaluation, are still in research preview and require separate access. Factor that into your roadmap. But for long-running autonomous tasks? This is the infrastructure layer the agentic era needed. The question isn’t whether to build with agents. It’s whether you’ll build the infrastructure yourself or let Anthropic run it while you focus on the product. 🔗 Docs: https://t.co/AspfnrLcbN
- 50
@alex_verem ·
Andrea Vallone spent three years at OpenAI working out how AI should handle people at their lowest. She's taken that work to Anthropic. If the name's new to you, it's worth knowing. At OpenAI, she built the Model Policy research team from scratch and led the work on one of the hardest questions in AI: how a model should respond when someone leans on it in a real moment of distress. There was almost no precedent for it. She and her team were writing the playbook as they went. She also worked on deploying GPT-4 and GPT-5 and helped develop rule-based rewards, a safety technique that's been widely influential. This isn't a quiet backroom hire. It's one of the people who shaped how the biggest models behave. Now she's at Anthropic, on the alignment team under Jan Leike, the researcher who left OpenAI in 2024 over how it was balancing safety against shipping. Her focus is on shaping Claude's behavior in new, unfamiliar situations. The move points at something bigger. The hardest part of AI safety isn't only jailbreaks and benchmarks anymore. It's the human stuff, what happens when millions of people start trusting these systems with things that actually matter. Anthropic has brought in one of the few people who's spent years thinking about exactly that. It's worth watching what she builds next.
Best Tweets by Topic