Who's fastest, and what it costs to run them.
As of Oct 16, 2026, Opus 5.5 leads the Model Standings at 58, 2 points clear of Sonnet 5.5. Haiku 5.5 gives the most score per dollar, at $0.10 in and $0.50 out per million tokens. Most used this week: DeepSeek V4.1 Flash, with 21% of OpenRouter tokens and 26% on Vercel AI Gateway. 14 LLMs from 8 labs; 3 not yet scored.
- Busiest model on OpenRouter that is not on our grid: Step 5 Preview (StepFun), 9.8% of tokens.
- Biggest move on Vercel AI Gateway since the day before: GLM-5.3, 8.4% → 9.8% of tokens.
- Biggest move on OpenRouter since the day before: DeepSeek V4.1 Flash, 20.8% → 21.2% of tokens.
- Grok 4.7 more than doubled its Vercel AI Gateway share week on week: 0.3% → 1.9%.
- Gemini 4 Argon is new on OpenRouter this week, with 0.3% of tokens.
- Muse WonderMeta · OpenRouter · Vercel · under 0.1% of OpenRouter · listed Oct 12
- Muse Spark OpenMeta · OpenRouter · Vercel · listed Oct 15
- LFM2.5-24B-A2BLiquidAI · OpenRouter · Vercel · listed Oct 13
| Pos | Model | Score | Gap | $ in / out | Laps | Test lap |
|---|---|---|---|---|---|---|
| P1 | Opus 5.5Anthropic | Leader | $4 / $20 | 3.4%▲6.9%▲ | – | |
| P2 | Sonnet 5.5Anthropic | −2 | $2 / $10 | 1.6%▲2.5%▲ | – | |
| P3 | Gemini 4 ArgonGoogle | −5 | $2 / $10 | 0.3%▲– | – | |
| P4 | Fable 5.1Anthropic | −5 | $10 / $50 | 0.1%▲– | – | |
| P5 | GPT-6 AstraOpenAI | −5 | $10 / $50 | 0.2%▼– | – | |
| P6 | GPT-6.1 SolOpenAI | −6 | $2 / $10 | 1.2%▲2.4%▼ | – | |
| P7 | Grok 4.7xAI | −12 | $2 / $6 | 0.1%▼1.9%▲ | – | |
| P8 | GLM-5.3Zhipu | −13 | $1.40 / $4.40 | 3.0%▲9.8%▲ | – | |
| P9 | Kimi K3Moonshot | −14 | $3 / $15 | 1.4%▲– | – | |
| P10 | Haiku 5.5TrackAnthropic | −15 | $0.10 / $0.50* | 1.1%▲– | Track | |
| P11 | GPT-6 LunaOpenAI | −20 | $0.10 / $0.50 | 3.7%▼6.7%▼ | – | |
| NC | V4.1 FlashDeepSeek | $0.30 / $1.20* | 21%▼26%▼ | – | ||
| NC | Gemini 3.8 FlashGoogle | $0.75 / $3.75 | 1.1%▼– | – | ||
| NC | Muse Spark 1.3Meta | $1.25 / $4.25 | 0.1%– | – |
In the news this week.
Headlines from the labs' own news pages and the two gateways' launch notes, in the seven days to Oct 16, 2026. Their words, linked to the source; we only pick which ones.
- OpenAIHow Itaú is building generative AI for its entire customer base
- OpenAIStack Overflow for Agents triples signups with ChatGPT plugin
- OpenAIBenefit Street Partners answers questions faster with Codex
- VercelElorian AI's Glimpse now available on AI Gateway
- VercelPerplexity's Decider V1.2 models now available on AI Gateway
- VercelReflection AI's Beam now available on AI Gateway for free
- AnthropicAn update on independent evaluators
- Google DeepMindAccelerating the search for antimicrobial peptides with Google Antigravity
- AnthropicPreparing for the 2026 US midterms
- AnthropicAnthropic opens Singapore office and signs MOU with the Government of Singapore
- Google DeepMindGoogle DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection
- OpenRouterPrivate Endpoints: Bring Your Fine-Tuned Models and Capacity to OpenRouter
Smarter costs more. Not always.
Score against blended list price, on a log scale. The line joins the models nothing else beats on both: every model off the line has a rival at least as smart for less. Hover or tap a dot for its numbers. Unscored models sit this one out.
scroll the chart sideways
Score is the Artificial Analysis Intelligence Index, taken from their public results and compared only within the same index version. Price is list, in US dollars per million tokens, input / output; open-weight prices are the first-party API, and hosts on OpenRouter often go lower. Blended price weighs input three to one, the shape of a typical agent turn. Laps is each model's share of all tokens over the last seven days on OpenRouter and on Vercel AI Gateway, two different crowds of developers; the arrow compares it with the seven days before, and a dash means too small to call. Source: OpenRouter (openrouter.ai/rankings), as of Oct 17, 2026. Licensed under CC BY 4.0. Source: Vercel AI Gateway usage data, as of Oct 16, 2026. Licensed under CC BY 4.0.
A verdict from putting the model on real work for a while, not a benchmark. Track means we would send real work to it today. Traffic means fine, nothing special. Pit lane means not ready. Most models have not had a lap yet.
Timing by sleds, which routes work across every lab and favours none. The board updates every morning, and every day keeps its own page, so a link to a day always shows that day. Quote it freely with a link back. The numbers behind it are open too, as JSON.
Past days.
Every day's board is kept as it stood, with that day's notes and headlines. Pick a day to see who led and what moved.