Usage rankings and Chinese-model adoption
Adoption rankings, token-volume leaderboards, and the shift toward Chinese and open-weight models on the platform.
32%
Best tweets about OpenRouter
50 Best Tweets About OpenRouter (2026)
Explore the best tweets about OpenRouter, covering model routing, API pricing, provider reliability, benchmarks, and developer integrations.
Specific OpenRouter models, routing behavior, API usage, pricing, provider performance, outages, and engineering lessons.
- Creators
- 44
- Updated
Original Xholic analysis
What 50 top OpenRouter posts reveal
Across the 50-post OpenRouter dataset, the largest theme is usage rankings and Chinese-model adoption (32%), followed by pricing and cost optimization (26%). The engineering takeaway in the cited posts is conditional: routing broadens model choice and is often discussed alongside lower spend, but authors also describe a need for evals, log auditing, fallback planning, and retesting when models or providers change. 2075631728027484462 2072279035493900395 2069073174747332783
- Dominant tone
- Mixed
- Median score
- 7.40
- Leading format
- Announcement
- Recent posts
- 56%
46% of posts
All-time engagement
58% of posts
Published in 90 days
Conversation map
The themes creators return to
Pricing and cost optimization
Token pricing, free tiers, proxy fees, cost comparisons, and substitutions of premium models with cheaper alternatives.
26%
Model launches and evaluations
Launches, specifications, context windows, licensing, capabilities, and hands-on evaluations of named models available through OpenRouter.
24%
Routing, orchestration, and portability
Dynamic model selection, multi-model agent architectures, fallback design, Fusion-style deliberation, and portability limits between model stacks.
18%
Provider reliability and operations
Provider rate limits, outages, latency, model deprecations, endpoint shutdowns, and production reliability concerns.
16%
API integrations and developer workflows
OpenRouter API integrations, proxies, SDKs, prompt replay, local-model connections, and practical configuration workflows.
14%
Evals, observability, and quality variance
Verification of routed behavior through logs, evals, quality checks, provider differences, and model-specific tool-use or output failures.
10%
Platform business and market structure
OpenRouter's marketplace position, usage-data interpretation, payment economics, and potential Stripe acquisition.
6%
Tone and stance
Sentiment Mixed leads
Author posture Supportive leads
Performance benchmark
- Median likes
- 44
- Median reposts
- 4
- Median replies
- 10
- Median views
- 4.6K
Posts with media make up 70% of this collection. Their median all-time score is 6.46, compared with 11.2 for text-only posts.
Format mix
- Announcement 58% · score 10.2
- Case Study 18% · score 4.29
- Opinion 16% · score 2.65
- Tutorial 8% · score 14.2
Consensus and debate
Where creators agree, and where they do not
Shared view
Price is a recurring lens in adoption posts
Posts citing OpenRouter usage describe increased token share or volume for Chinese and open-weight models. These posts repeatedly foreground lower pricing as part of the comparison, though they do not independently establish price as the sole cause of adoption.
Shared view
Cost comparisons drive experimentation
Authors describe lower-cost substitutes, free access, and single-prompt long-context workflows as reasons to test alternatives to premium models. The cited comparisons and savings are author-reported, and some posts note quality or task-dependent trade-offs.
Open debate
Routing utility does not imply seamless portability
Routing is presented as useful for assigning tasks or reducing costs, while the cited posts also argue that switching models is not frictionless. They point to model-specific prompts, tool behavior, skills, evals, and routing logic as work that may need revision or retesting.
Open debate
Leaderboard readings have stated limitations
Some posts use OpenRouter usage charts as adoption signals. Others caution that an aggregator leaderboard represents a particular slice of usage and can be affected by routing behavior, direct accounts or contracts, and rate limits.
What performs
Patterns behind standout posts
Top outliers covered launches, workflows, orchestration, and cost
These four posts are the highest-scoring outliers in the deterministic analytics. Their subjects span a free model preview, distillation utilities, orchestration and portability analysis, and premium-model cost comparisons.
Developer posts emphasize auditing and validation
The cited posts emphasize comparison and verification practices: replaying prompts across models, checking identical-model behavior with evals, and accounting for tool-use reliability and changed data paths when using a proxy or alternate provider.
Access and endpoint changes are operational considerations
Operational evidence includes an OpenRouter uptime warning about single-vendor inference, a reported rate-limit experience, and a hosted endpoint sunset notice with a migration recommendation. Together, these posts support planning for alternate access paths and endpoint changes.
Statistical standouts
- View standout post 1 Score 556.1 · 75.14× median
- View standout post 2 Score 366.8 · 49.57× median
- View standout post 3 Score 231.2 · 31.24× median
- View standout post 4 Score 179.9 · 24.32× median
- View standout post 5 Score 73.5 · 9.94× median
Creator landscape
Who shapes this conversation
The five most represented creators account for 20% of the selected posts.
-
1. AshutoshShrivastava
@ai_for_success
2 posts
-
2. BridgeMind
@bridgemindai
2 posts
-
3. Layton Gott
@Layton_Gott
2 posts
-
4. OpenRouter
@OpenRouter
2 posts
-
5. Pukerainbow 🤮🌈
@pukerrainbrow
2 posts
-
6. Rohan Paul
@rohanpaul_ai
2 posts
BridgeMind led the listed top-voice cohort by median score
Among the listed top voices, BridgeMind has the highest median all-time score (290.7). Its two cited posts focus on model-launch specifications, including context, pricing, and multi-agent behavior.
Official posts highlight comparison tools and vendor-risk awareness
OpenRouter’s cited posts announce prompt replay and remix functions for auditing, iteration, and model comparison, and warn against relying on single-vendor inference for uptime.
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 OpenRouter tweets
Top OpenRouter tweets from 44 creators
Ranked 01–50
- 01
@bridgemindai ·
Qwen 3.6 Plus Preview just dropped on OpenRouter. 1,000,000 token context window. $0 input. $0 output. It's Free. A million tokens of context for free. This is insane. Claude Opus 4.6 charges $5/$25 per million tokens for 200K context. GPT 5.4 charges for 1M context. Qwen just gave it away. All free. Open source is not slowing down. It's accelerating.
- 02
@rasbt ·
Claude distillation has been a big topic this week while I am (coincidentally) writing Chapter 8 on model distillation. In that context, I shared some utilities to generate distillation data from all sorts of open-weight models via OpenRouter and Ollama: https://t.co/IsfNDpcGAw
Watch video
- 03
@nikesharora ·
Summary: I spent time trying to figure out this orchestration layer problem, can we design a multi model architecture in the long term. The more I dug in the more I understand that trying to build an abstracted layer is hard. As agentic activities increase and agent chaining and complex tasks get assigned to AI it will become harder to move between models. There is a reasonable probability that 75% of the enterprises will build their implementation of the solution to their core problem around one model "stack". Token price reduction by 90% is the solve and mobility between models from the same frontier lab! Evals, harnesses, cache memory are the moats and I don't see models providing simple abstraction to those. I know there are efforts to do this out there, the long term solve for orchestration if it works will need to be "Claude code" level of design genius. Here's a chat with Fable @HamzaFodderwala had. **Why abstraction looks easy.** Models are stateless — every API call is weights + a prompt assembled at runtime. Everything the model "knows" about you — memory, documents, history, tools — is injected into the context window by software outside the model. So in principle, all your state already lives outside the weights. The catch is what "state" includes. **Layer 1 — Data (fully portable).** Enterprise documents, tickets, logs. Retrieved via RAG: text is chunked, embedded, stored in a vector database (Pinecone, pgvector), and relevant pieces are fetched into the prompt per query. The embedding model is separate from the LLM, so this layer is genuinely model-agnostic. Already solved. **Layer 2 — Memory (portable in principle).** Systems like Mem0 and Zep sit between the app and the model: after each interaction they extract salient facts ("user prefers X"), store them as plain text, and inject the relevant ones into future prompts. Because the artifact is natural language, it reads into any model. Facts port. **Layer 3 — Orchestration/routing (works, but only for shallow tasks).** Gateways like OpenRouter and LiteLLM normalize API differences and route each request to the cheapest capable model. This is the fungibility layer being furiously built. It genuinely works for one-shot, verifiable tasks — classification, extraction, summarization — which conveniently are the tasks where cheap models suffice anyway. **Where it breaks — the non-portable state.** Four things stay behind when you switch: - **The harness.** Prompts, tool schemas, and guardrails are tuned to one model's quirks. An agent must get every step right, so reliability compounds: a model that's 98% reliable per step completes a 50-step task about a third of the time; at 90% per step, it almost never finishes. Swapping models costs you a few points per step — the difference between an agent that works and one that doesn't. - **The evals.** Swapping means re-testing everything and re-fixing every regression. The real switching cost isn't data migration — it's re-verification. Nobody has abstracted that. - **Procedural memory.** Facts port; skills don't. Cached successful workflows and learned workarounds are conditional on the model that produced them. - **Cache pricing.** Provider-specific, worth 75–90% of input costs on agentic workloads. Quiet lock-in. **The labs' angle.** They offer hosted memory, hosted file stores, caching, fine-tuning — every one pulls state from your side onto theirs. The labs will crack memory first, but as lock-in, not portability. Nobody standardizes their own exit door. MCP is the partial exception: it standardizes tool and data access across models, but doesn't touch harness tuning or evals. **Where 3P vendors fit.** Routers are thin-margin commodity plumbing; vector DBs and memory infra are real but small. The two structurally interesting positions: **eval platforms** (LangSmith, Braintrust) — since switching cost equals re-verification cost, whoever industrializes cross-model testing actually enables fungibility.
- 04
@DeRonin_ ·
Replaced these PREMIUM models with 100%+ cheaper alternatives 1. Claude Opus 4.6 → Kimi K2.5 Benchmark gap: ~11% Price difference: ~9x cheaper 2. GPT-5.4 → Qwen3.5-397B Benchmark gap: ~21% Price difference: ~6–7x cheaper 3. Claude Sonnet 4.6 → GLM-5 Benchmark gap: ~2% Price difference: ~3.8x cheaper on input and ~5.9x cheaper on output 4. Gemini 3.1 Pro → GLM-5 Benchmark gap: ~12% Price difference: ~2.5x cheaper on input and ~4.7x cheaper on output 5. GPT-5.4 Pro → Kimi K2.5 + DeepSeek R1 Benchmark gap: GPT-5.4 Pro still has an advantage on complex tasks, but the price difference is very attractive Price difference: ~95–97% cheaper Conclusion: Based on OpenRouter pricing and Artificial Analysis benchmark scores, these lower-cost models are often 50%+ cheaper while remaining surprisingly close in capability for many tasks Some of them may underperform the premium models on harder or more complex tasks, but the price difference is significant Considering the 50%+ savings in most cases, this is a strong option if you don't have a large budget
- 05
@kaif9998 ·
Anthropic banned my Julius (OpenClaw) from their API 💀 So I tested every single alternative out there. Was burning $150/month on credits. Tried Kimi K2.5… and now I finally see why it’s the most-used model for OpenClaw on OpenRouter. > 90%+ of Claude Opus performance > Built-in web search (no extra API calls) > 8x cheaper than Claude I went from $150/mo → $15/mo. Almost the same coding output and reasoning quality. The math just isn’t even close anymore. Who else made the switch? What’s your new stack? 👇
- 06
@HedgieMarkets ·
🦔U.S. companies are routing more AI workloads to Chinese models. CNBC reported that Chinese model usage on OpenRouter sat above 30% of U.S. company tokens every week since February, up from an 11% average the prior year and just 4.5% in H1 2025. https://t.co/IJWtvXQaKh's GLM 5.2 saw token volume grow 27x in its first week on Vercel. One startup, Lindy, moved 100% of its traffic from Claude to DeepSeek and expects to save millions. My Take The pitch from U.S. labs has always been that frontier performance justifies frontier pricing. That logic holds until someone ships a model within a percentage point of Opus 4.8 on agentic benchmarks at a fifth of the cost, which is what GLM 5.2 apparently did. At 60 to 90% cheaper for most workloads, you do not need to match the best model. You just need to be good enough for the task you're routing to it. If U.S. labs keep raising token prices while Chinese open-weight models close the performance gap, the market self-selects toward Chinese infrastructure for everything except the hardest problems. Most companies making that switch are not thinking about geopolitics. They are thinking about their AWS bill. Hedgie🤗
- 07
@CodeByNZ ·
👀 Tencent Hunyuan HY3 is moving a lot faster than I expected. Just one week after launch, it's already the #1 most-called model on OpenRouter. What's even crazier is the growth. The preview version hit 3.6 trillion model calls two months ago. Now HY3 has surpassed 6 trillion model calls, nearly doubling that number. The open model race is getting more competitive every month, and Tencent is proving it's a serious contender. @TencentHunyuan #HY3 #Hunyuan #AI
Watch video
- 08
- 09
@bridgemindai ·
Grok 4.20 Multi-Agent just went live on OpenRouter. This model is built different. Low/medium reasoning: 4 agents in parallel. High/xhigh reasoning: 16 agents in parallel. One API call. Up to 16 agents coordinating, researching, and synthesizing in parallel. 2M context. $2/M input. $6/M output. This is the #1 model on BridgeBench for a reason. It's not one model thinking harder. It's 16 models thinking together. xAI just made multi-agent native at the model level. Everyone else is bolting it on after the fact.
- 10
@Rahul_J_Mathur ·
Last night I panicked when our fund’s agent Zen replied to me in Chinese. For a minute, I wasn’t sure WTF happened. Until recently, our agents have run exclusively on the Opus series of models (this is a $100k annualized bill for a 50 person team) Sometime in June, we began experimenting with a multi-model set-up using OpenRouter with an explicit goal of reducing our spend by ~40% without compromising on long term performance or security. Sharing my personal observations: (1) DeepSeek v4 pro replies in Chinese at least once per day (approx. once in ~30 tasks). There is an open ticket in GitHub specifying this very issue which remains unresolved. Furthermore, DeepSeek is noticeably slower for token heavy tasks (to a point where I need to do other work while waiting to check outputs) (2) GLM 5.2 is observably slower than Opus 4.8 & produces worse quality output GLM 5.2 doesn’t read the complete Skill file, fails to handle edge cases etc. However, at 1/8th the cost, this is an acceptable trade-off for me since the output quality is observably worse for me but NOT for the recipient of the output. (3) Skills built for Opus & GPT series models don’t translate well to the GLM series & DeepSeek model variants. Opus & GPT have far superior tool calling capability (my hypothesis) & more grounded to follow instructions (knowledge worker centric RL or inclusion in training data sets). I regret to inform the hype lords that with appropriate & directed efforts, it is possible to switch workflows between model providers⚠️ However, this is NOT seamless & does require one-time effort to update Skills, usage muscle memory & hard code certain routing logic.
- 11
@Parul_Gautam7 ·
I've been using Ring-2.6-1T on OpenRouter for my documentation pipeline this week. My usual flow: I dump a raw codebase or a long spec doc into the model and ask it to generate structured docs, flag inconsistencies, and draft a changelog. With other models, I was splitting this into 3 separate prompts because of context limits. Each split costs tokens. Each split loses some coherence between sections. With Ring-2.6-1T, I sent the whole thing in 1 prompt. 38,000 tokens. got back structured docs, a flagged inconsistencies list, and a changelog draft in a single response. Coherence was intact across all 3 outputs. No context bleed between sections. Still running it through more workflows, but the single-prompt full-doc processing alone has already cut my per-task token spend significantly. And then the cost showed $0.00, and I sat there for a second 👉 https://t.co/DZIXXDMiqK #BuildInPublic #OpenSource
Watch video
- 12
@TheGeorgePu ·
DeepSeek went from under 1% of OpenRouter tokens to 17%. That happened in a single month. The top 4 models by usage are now all Chinese open models. Nobody decided this in a press release. Finance teams looked at the bill and rerouted the traffic. I found out I don't have to run my whole workflow on Claude and GPT. The cost difference is multiple orders of magnitude. I think more people are figuring that out too. The leaderboard everyone watches is one thing. The model people actually pay for is another.
- 13
@shedntcare_ ·
BREAKING: Stealth model Elephant-alpha reaches #2 on OpenRouter Trending. The ranking is based on real token consumption across OpenRouter—reflecting actual developer usage, not benchmarks or subjective evals. Signal: rapid adoption and growing traction in live environments. Source: https://t.co/ydrGxGqs1p
- 14
@Layton_Gott ·
Last year Anthropic had 29% of all the AI usage on OpenRouter. Today it’s only 13%. You’re probably wondering why Claude’s usage got cut down over 50%… The answer is simple: Chinese open weight models. They went from basically nothing to over 60% of all the tokens running through OpenRouter. In about a year and a half. The top six most used models on the entire platform are now Chinese. Claude sits at number six. MiMo from Xiaomi. DeepSeek. MiniMax. GLM. Models a lot of people have never even heard of are now bigger than Claude by usage. Why? Price. These models run for as little as a tenth of what Claude costs. Some even less. And they’re now good enough that for most jobs you genuinely can’t tell the difference. So developers did the obvious thing. They followed the cheaper option. Anthropic is only 12% of the tokens, but it’s still pulling around 46% of the actual money on the platform. So the market split in two. One lane is cheap, open, good enough, and absolutely massive on volume. That’s the Chinese models. The other lane is expensive, premium, and where the real money still is. That’s Claude and GPT. No, Claude isn’t dying. It’s making more money than ever while slowly losing the volume war. But the direction is impossible to ignore. The cheap open models are taking the actual usage, and they’re getting better every single month. And the closed frontier models keep getting locked behind governments and price tags. This is exactly why I keep saying open source wins in the end. You can’t out price free. And you can’t shut off a model the whole world already downloaded.
- 15
- 16
@sharbel ·
Anthropic's API can run $1000+/month at heavy Claude Code usage. Someone built Free Claude Code: A proxy that lets Claude Code talk to free or local model providers instead. It's called free-claude-code. 7,300+ stars on GitHub. Set two environment variables (ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN), point Claude Code at the proxy, done. The CLI and VSCode extension keep working, but behind the scenes you're hitting a different model, not Claude. What it supports: → Routes to NVIDIA NIM's free tier (rate-limited). → OpenRouter, including free model tiers. → DeepSeek direct API. → LM Studio / llama.cpp for fully local inference. → Per-model routing: different provider per Claude model tier. → Thinking-token parsing into Claude thinking blocks. → Heuristic tool-call parser for models without native tool use. → Discord / Telegram bot for remote coding sessions. Real trade-offs to know: - Quality drops. Backing models are not Claude. Tool use is less reliable (that's why the heuristic parser exists). Agentic tasks that Sonnet or Opus nail will fail more often on free tiers. - Free tiers aren't permanent. NIM / OpenRouter / DeepSeek quotas and pricing can change. - Data path changes. Your prompts and code now flow through whichever provider you route to, with their retention policy, not Anthropic's. Worth it for tinkering, offline work, or non-sensitive side projects. Not a drop-in replacement for paid Claude on client code or serious engineering. If your actual goal is "pay less for Claude Code at full quality," the Claude Pro sub ($20/mo) or Max tier ($100-200/mo) includes Claude Code usage directly, no API key needed. MIT licensed. 100% open source.
- 17
@david_nix ·
Is Cloud AI a scam? I was running some evals Local qwen models 27b and 9b right off my laptop. Simple data extraction. Here's a pile of text, give me JSON plz. My local models passed evals 100%. Switched to Openrouter. Exact same models... or so I thought. Several evals failed. So what gives? Did they quantize more than advertised?
- 18
@pukerrainbrow ·
46% of US company AI usage on openrouter is now chinese models openai cut luna prices 80% three weeks after launch. $1/$6 down to $0.20/$1.20 per million tokens same week they hit a billion users a billion users and they still couldn't keep their prices because chinese labs charge less and companies moved US tried to win with restrictions. china won with pricing have you tried any chinese models yet?
- 19
@VaibhavSisinty ·
Every major AI company shipped the same thing in the last 48 hours. Model routers. A layer that decides which model to use so you don't have to. Cursor launched Cursor Router on Tuesday. It reads every request, checks the complexity, and sends it to the cheapest model that can handle it. 60% cost savings. Same quality. But Cursor isn't alone ↓ → OpenAI already built routing into ChatGPT. It auto-switches between Instant and Thinking models based on query complexity. → Factory AI launched Factory Router. Claims 20-25% token savings with no quality drop. → OpenRouter upgraded their Auto Router across 400+ models from 60+ providers. → IDC predicts by 2028, 70% of top AI enterprises will use dynamic model routing. All within the same window. That's not coincidence. That's consensus. Here's the problem they're all solving. Most people pick one model and use it for everything. Cursor's own data says 60% of their developers do this. Frontier models built for complex reasoning are being used to fix typos and write Slack messages. That's a Ferrari doing grocery runs. Routers fix that automatically. Simple task goes to a cheap model. Hard task goes to frontier. The model race is slowing down. The top models are within a few points of each other. The real competition has quietly moved to who can deliver the same output at the lowest cost. That's what routers do. And every major player just shipped one in the same week. The era of one model for everything is ending. Smart routing just started.
- 20
@Layton_Gott ·
Open source local models are the future... You should be preparing for that now. But that doesn't mean ignore the present. Right now the paid subscriptions are worth WAY more than what you pay for them, so use them while that lasts. Here's how I do both at once: I run a free local model as my operator. Qwen3.6 27B, running on my own machine, costing me nothing. It doesn't do the heavy lifting itself. It routes. It dispatches tasks out to Claude Code, Codex, DeepSeek V4 Pro, GLM 5.2, and honestly any open source model I want through OpenRouter. So the second a new open model drops, I hook it in. My operator stays free and local, and the actual work goes to whatever model is best for it right now. I get the best of the present (frontier subscriptions worth more than I pay) and I'm already set up for the future of running local models although I'd like to get some more ram soon. Prepare for where it's going... But don't move past the present quite yet without taking FULL advantage of it.
- 21
@rohanpaul_ai ·
Reuters: A mysterious AI model named Hunter Alpha recently appeared on OpenRouter. Rumors say it might be a secret test version of DeepSeek V4. - This model is massive 1T param. - 1M token context window - Many noticed reasoning patterns look a lot like the chain-of-thought style seen in previous DeepSeek models During testing, the model claimed it was trained mostly on Chinese data with a knowledge cutoff of May-25, which matches DeepSeek’s official specifications. --- reuters .com/business/media-telecom/mystery-ai-model-has-developers-buzzing-is-this-deepseeks-latest-blockbuster-2026-03-18/
- 22
@ai_for_success ·
Qwen has just launched Qwen3.6-Plus model and it’s Free on OpenRouter , it is a significant upgrade over the Qwen3.5 series - 1M context window by default - significantly improved agentic coding capability - better multimodal perception and reasoning ability - Also available via the Alibaba Cloud Model Studio API along side OpenRouter. Some use case below from blog , Video Source and Credit : Qwen Blog Qwen is also planning to Open-source smaller-scale variants of the Qwen3.6 series shortly. 1/5
- 23
@burkov ·
The fact that the models are updated all the time and then older versions are deprecated is super annoying. You just found a model that works well for a given use case and costs well too. So you are using it directly from the provider or via OpenRouter. Then they decide to deprecate it, and your only choice, if the model is open weight, is to rent GPUs to run it, which is too expensive for the use case. Otherwise, you need to test them all once again until you find one that works reasonably well for a reasonable price for your use case.
- 24
@ai_for_success ·
There is new anonymous model called elephant-alpha, currently in the top 10 on OpenRouter by daily usage. Looks good for agentic work because OpenClaw and Hermes Agent are among the top apps using it. Tried it on some code review and refactoring. It is pretty fast and handles it well. Not a deep reasoner but solid for quick execution tasks. Feels similar to Grok Fast. Anyone tried this yet?
Watch video
- 25
@pukerrainbrow ·
OpenRouter's image model rankings (June 2026): 1. Nano Banana (Gemini 2.5 Flash Image), 1.71M requests, 40.7% share 2. Nano Banana 2 (Gemini 3.1 Flash Image), 1.21M requests, 28.8% 3. Nano Banana Pro (Gemini 3 Pro Image), 825K requests, 19.6% 4. Seedream 4.5 (ByteDance), 99K requests, 2.4% 3 google models are swallowing 89% of the platform’s image gen traffic. whatever “open-weights from china” narrative people want to push, the numbers don’t even start to back it up. Is this the one category nobody's actually built the open-weight alternative for yet? Genuine question.
- 26
@alex_verem ·
The most-used model on OpenRouter this week isn’t GPT, Claude, or DeepSeek. It’s Tencent’s Hy3 🤯 8.98T tokens. #1 on the platform. So I tested whether the efficiency shows up in the actual output. I asked it to: “Write a Python function that finds the maximum profit from one stock purchase and sale. Start with the brute-force solution, optimize it to O(n), then explain the time and space tradeoffs.” Hy3 first returned the brute-force O(n²) approach, then rewrote it as a single-pass solution that tracks the minimum price. It explained why the second approach reduces the time complexity from O(n²) to O(n), along with the space tradeoff. Architecturally, Hy3 has 295B total parameters while activating 21B per token. The published specs include: • 256K context • Fast and slow reasoning in one model • Apache 2.0 open-source licensing • Tencent Cloud lists pricing at 1 CNY/M input tokens and 4 CNY/M output tokens. Those specs may partly explain its recent OpenRouter usage, although one coding prompt obviously isn’t a full benchmark. @TencentHunyuan #Hy3
Watch video
- 27
@alliekmiller ·
Xiaomi now processes more AI tokens than OpenAI. On OpenRouter, Chinese models just crossed 45% of all token volume. Anthropic is at 15.3%. OpenAI is at 7.4%. For now, the frontier is American models, but the volume is Chinese models. Open source models are often good enough at coding and agents for most tasks, can ship a 1M token context window, and cost a fraction of GPT or Claude. 80% as good at 20% of the price is pretty freaking attractive to a procurement team.
- 28
@Zubairey0 ·
Saw that @stripe is in talks to acquire @OpenRouter for nearly $10 billion, less than two months after OpenRouter was reportedly valued at $1.3 billion. If the deal goes through, it would be one of the largest acquisitions in AI infrastructure this year. Agentic payments are still tiny. AI agents settled roughly $73 million across 176 million blockchain transactions in the year through April. A $10 billion valuation suggests Stripe is paying for where it believes the market will be in five to ten years, not where it is today. If agentic commerce scales as expected, it could look like a bargain. If adoption takes longer, it may prove to be a very expensive bet.
- 29
- 30
@RoundtableSpace ·
Someone got fed up with Claude Code blocking tasks mid-session and lecturing them about blood test questions, so they switched to Kimi K3 through OpenCode for $19. Here is the exact setup: → Ask Codex or Claude Code to install OpenCode → Go to https://t.co/N9c6hpx2AZ, make an account and pay for membership at $19 a month → Get your API key at https://t.co/3BBIldfy53 → Run OpenCode, type /connect and connect to Kimi Code, paste the API key → Set it to BUILD mode with Shift+Tab → Go to /settings and enable bypass permissions OpenRouter hit rate limits immediately so going direct to Kimi's API was the fix. Once connected, Kimi K3 runs without restrictions or interruptions.
- 31
@Marktechpost ·
Arcee AI just released Trinity Large Thinking, an Apache 2.0 open reasoning model built for long horizon agents, multi turn tool use, and structured outputs. Key specs: • 400B sparse MoE • 13B active params per token • 4-of-256 routing • 262k context on OpenRouter • #2 on PinchBench, behind Claude Opus 4.6 Full analysis: https://t.co/pBNXW0tqDN Technical details: https://t.co/Vw298MR3yg Model weight: https://t.co/fFoFEKhNBS @arcee_ai #AI #LLM #OpenSourceAI #Agents #MachineLearning
Watch video
- 32
- 33
@trikcode ·
Tried Sakana's Fugu today. Here's my honest take. - it's not a new base model. It's an orchestrator that routes your prompt across multiple models (GPT, Claude, others) behind one API, with a "conductor" deciding who handles what. - the benchmarks look genuinely strong, especially right after the Fable/Mythos export restrictions. - the "beats Mythos" headline is doing a lot of work and misleading. - It's just closer to OpenRouter's fusion approach than an actual frontier model. Simply Calling multiple AIs and having them check each other's work which is clever engineering, not a new model breakthrough. Also got hit with a couple slow responses and a 403 error during testing. Worth watching for. Not worth in my opinion but good to have
- 34
@rohanpaul_ai ·
Just a few days before, OpenRouter raised $113M at a $1.3B valuation and now, WSJ reports Stripe is negotiating a roughly $10B acquisition talk with it. Stripe already handles OpenRouter's payments, so ownership would combine billing with AI distribution. That position could give Stripe usage data and pricing influence across competing model providers. Companies increasingly mix models because the best choice changes by task, cost, and traffic. OpenRouter sits between those buyers and labs, collecting payments while routing each request. Stripe would mainly pay for control of a growing decision layer above model providers. The price works only if model choice becomes infrastructure as basic as online payments.
- 35
@shawnchauhan1 ·
The majority of traffic on the largest AI model marketplace is no longer American. Chinese and open models now account for over 60% of usage on OpenRouter. Airbnb, Pinterest, and DoorDash are already testing the switch. Even Microsoft is reportedly weighing DeepSeek inside its own Copilot product. Nobody announced this shift. Developers just quietly voted with their traffic. That's usually how the real signal shows up, not in a press release. Watch usage, not marketing.
- 36
@aakashgupta ·
Kimi K2.6 hit #1 on OpenRouter last week with 1.58T tokens. Claude Sonnet 4.6 came in at 1.36T. The leaderboard isn't measuring what most people think it's measuring. OpenRouter is a routing aggregator. Developers set a quality bar, OpenRouter sends the request to the cheapest provider that clears it. The leaderboard is downstream of that. It's a proxy for who wins the price-per-good-enough-token war on the slice of the market that doesn't have direct enterprise contracts. Kimi K2.6 costs $0.60 per million input tokens on Moonshot's own API. Sonnet 4.6 is $3. Opus 4.7 is $5. So Sonnet is 5x Kimi's price and Opus is 8x. And Claude Sonnet still pulled 86% of Kimi's volume on the same platform, at 5x the cost. That's the actual story in this chart. The most expensive workhorse model in the industry is doing 86% of the volume of the new cheapest-good-enough option, on the platform built to commoditize the middle. Now scale that. Anthropic has THREE models in the top 10. Sonnet 4.6 (1.36T) + Opus 4.7 (1.15T) + Opus 4.6 (685B) = 3.2 trillion tokens. That's 2x Kimi's total volume on a platform that systematically routes against expensive models. Opus 4.7 launched four days before Kimi K2.6 and is already at #4, up 279% week over week. The thing nobody screenshots: the Y-axis. OpenRouter is processing 28 trillion tokens a week. The chart shows a near-vertical climb starting in February. The price-sensitive slice of the dev market is exploding. Open WebUI users, Cline and Roo Code agents, indie builders without enterprise contracts. And underneath that explosion, the market is splitting cleanly. Chinese open-weight (Kimi, DeepSeek, MiniMax) owns the volume war on price. Anthropic owns the segment that pays for quality. Google parked two Flash models in the top 10 for the workhorse middle. xAI quietly hit #7. Everyone is winning a different game. OpenRouter is the cheap seats. The actual revenue is happening inside Cursor, Claude Code, Bolt, Lovable, Replit. Direct integrations. Enterprise contracts. None of that volume shows up on this chart. Kimi at #1 is real. It took an open-weight model at 1/5 the price to edge out Sonnet by 16% on the platform built to commoditize them both.
- 37
@KenTheRogers ·
One of the biggest benefits of using @OpenRouter is outage safety. If my personal Hermes agent is hooked up to Codex auth and they have an outage it’s annoying but not a huge deal. If a client that I built a custom agent for to run a key workflow has an outage, that’s unacceptable. Anything that needs stability when providers have outages should be built on OpenRouter.
- 38
@arcee_ai ·
An update on Trinity-Large-Preview: We are officially sunsetting the hosted preview endpoint on OpenRouter. The preview model has served over 3.3T tokens since it's launch in late January. A final thank you to everyone who used it. If you still have active pipelines targeting the preview endpoint, consider transitioning them to Trinity-Large-Thinking to avoid any service disruption. As a reminder, Trinity-Large-Thinking is free on @OpenRouter until 5/23. https://t.co/a9oVQkIfPn
- 39
@Amank1412 ·
Tencent’s Hy3 just took the #1 spot on the OpenRouter LLM leaderboard. A 295B MoE model fully open under Apache 2.0, so it’s actually commercial friendly. Built for reliable affordable agentic workflows while going head to head with trillion parameter flagship models. Open models just keep getting better.
- 40
- 41
@ShinkaIoT ·
Fusion turns your prompt into a small multi-model deliberation. The "Fusion" API from Openrouter entirely flips the math on how we use frontier models. A panel of expert models (see below) analyzes your prompt in parallel with web search and web fetch enabled, then a judge model synthesizes their responses into a structured analysis — consensus, contradictions, partial coverage, unique insights, and blind spots — and writes the final answer from it. 🚀🔀 This isn't a price story; it’s an architecture story. The breakdown of how Fusion works and why single-shot prompts are officially obsolete: • 🎰 The Three-Layer Deliberation: When a prompt hits OpenRouter/Fusion, it triggers a compound workflow. First, a parallel Panel of models analyzes the prompt (with web search and tools active). Next, a Judge model audits all competing outputs, mapping out consensus, contradictions, and blind spots. Finally, a Synthesizer builds a clean, unified response grounded in that multi-perspective analysis. • 📊 Benchmarking the Multi-Model Advantage: OpenRouter tested this framework on Draco's strict deep research benchmark across law, finance, and medicine. A panel of Fable 5 and GPT 5.5 synthesized by Opus 4.8 scored an elite 69.0%, outperforming Fable 5 running solo (65.3%). Even more impressive: a budget preset combining cheap models (Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro) hit 64.7%—matching frontier performance at half the price. • 🔄 The Power of Self-Synthesis: The data proves that 75% of Fusion's performance jump comes strictly from the synthesis step, not model scale. Even when a single model like Opus 4.8 evaluates its own work in a dual-panel setup, its benchmark score rises from 58.8% to 65.5%. Iteration and comparative verification beat single-shot answers every time. • ⚖️ Stacking Cost vs. True Accuracy: Because Fusion runs 3 to 5 models in parallel, you pay the cumulative token cost of the underlying completions. However, the budget preset stacks up cheaper than running a single, premium frontier model while offering a built-in audit trail. Knowing that three independent models verified the same fact provides a completely different level of enterprise confidence. The Takeaway: The future of AI is cognitive diversity and smart orchestration, not single-model monopolies. Stop asking one model a question and blindly pasting whatever it spits out. Head to OpenRouter, initialize the Fusion API, and start running multi-model panel checks on your deep research, market analysis, and content workflows. Stop prompting. Start orchestrating. 🖥️⛓️
- 42
- 43
@ChinaScience ·
The share of tokens used by US companies on Chinese AI models via OpenRouter — one of the major overseas multi-model aggregation and routing platforms, serving as an important window for observing global LLM usage trends — has sat above 30% every week since February 8, with that figure rising as high as 46%, a significant surge from just 4.5% in the first half of 2025 and 11% in past 12 months. Attributing the trend to a combination of factors including cost-effectiveness, enhanced performance, open-source autonomy, and supply chain diversification, Chinese analysts said that they expect more foreign companies to adopt Chinese AI models.
- 44
@jasonlk ·
"How important is optimizing model costs? In some cases, it's very important. Look at OpenRouter's data. There is a large set of customers optimizing on open source models, mini, haiku, etc. But ... other B2B + AI Apps, it's not that important. The credit costs aren't massive, and the benefits from using the top Opus/Sonnet models are so high ... you still with one platform. And just invest more and more in it." with @HarryStebbings @rodriscoll + me
Watch video
- 45
@Aizcalibur ·
I believe this will be helpful to my fellow community members. Did you know @NousResearch Hemes has an auxiliary system for specific functions that can be routed through specific providers? 🤌 But here is what I stumbled upon and fixed before getting rekt. I noticed GLM-5 requests showing up on my OpenRouter dashboard even though my primary model was configured to use ZAI's direct API. GLM-5 doesn't take image input and this is where the auxiliary system comes into play. I assigned Gemma 4 via OpenRouter as the provider to 'vision' which takes image input and provides text output to GLM-5. The config looked correct on the surface and even Hermes just worked as it always does. Luckily, I checked my OpenRouter logs and saw GLM tags and was super confused. Shared the screenshots with Hermes and we dug deep into the configs. For reason I don't know, when I asked it to set OpenRouter for auxiliary, it got added to delegation provider with a blank model which was then taking Zai. Plus 'auto' was set in auxiliary and it routed for GLM but via OpenRouter for all the other tools except for 'vision' for which it took Gemma 4 as intended. As per Hermes, the fix here was changing the delegation + specific providers for each auxiliary tool. Quite an interesting system. If I hadn't seen the logs, I would be wasting money on OpenRouter even though I have a direct coding plan from Zai. I recommend checking logs and auditing if you use multiple models.
- 46
@alvinfoo ·
Here’s what most people are missing right now: Everyone is still debating which model is smarter — GPT vs Claude. But the real game has already shifted. Last week’s usage data from OpenRouter tells a very different story: All top 6 models by token consumption were Chinese. Let that sink in. This isn’t about benchmarks. This isn’t about demos. This is about distribution at scale. Take Alibaba’s Qwen: → 100,000+ derivatives across open-source ecosystems → Embedded everywhere developers actually build → Quietly compounding usage, day after day While others compete for attention, this ecosystem is capturing adoption. And that’s the real moat. Because in AI: •The best model doesn’t always win •The most used model does China didn’t just build competitive models. They seeded a self-replicating ecosystem, one that grows through developers, forks, integrations, and real-world usage. No hype. No noise. Just distribution compounding. The takeaway? 👉 If you’re still focused only on model quality, you’re playing yesterday’s game. 👉 The winners of this cycle will be decided by who owns the rails of usage. And right now, the data is telling us: the center of gravity is shifting.
- 47
- 48
@localjulius ·
Xiaomi's free AI usage beat Claude Sonnet on OpenRouter last week. By 58%. > Xiaomi MiMo-V2-Pro: 1.77T tokens/week > Step 3.5 Flash (free): 1.61T > MiniMax M2.5: 1.39T > DeepSeek V3.2: 1.23T > Claude Sonnet 4.6: 1.12T > Claude Opus 4.6: 1.06T Xiaomi is beating Claude Sonnet 4.6 by 58%. Is this free-tier volume arbitrage? Or is the quality genuinely there?
- 49
- 50
@gudanglifehack ·
AI Digest: OpenRouter costs and OpenAI’s GPT-5.6 Sol update 1. Accessing DeepSeek models through OpenRouter is more expensive OpenRouter routes requests through its own servers before reaching the third-party host, incurring additional costs and inefficiencies due to routing, caching, and infrastructure overhead.
Best Tweets by Topic