--- title: "Ranges · Undominated.ai" canonical: https://undominated.ai/ranges/ description: "The leaderboard gives you a position. The benchmark also publishes an interval around it. At rank 40, that interval covers 25 different models — and they cost between $0.28 and $50 per million tokens." --- # Ranges · Undominated.ai > The leaderboard gives you a position. The benchmark also publishes an interval around it. At rank 40, that interval covers 25 different models — and they cost between $0.28 and $50 per million tokens. Understand the numbers # Ranges The leaderboard gives you a position. The benchmark also publishes an interval around it. At rank 40, that interval covers **25** different models — and they cost between **$0.280** and **$50.00** per million tokens. [Stand in a position ↓](#slot-explorer) [Ranges ↓](#ranges-figure) A board that publishes a confidence interval for every model also publishes **419** distinct integer positions for 419 models, with **no ties**. Across the whole board, **416** of 418 neighbouring intervals overlap. That is what the published uncertainty looks like when you draw it instead of sorting by it. These bands are wider than the truth, on purpose. They are built from each model’s own published interval and nothing else, which ignores that the scores are jointly estimated from shared battles — a covariance the published intervals cannot recover. So a band shows what is *not ruled out*, never what is identified, and two models sharing a position are not thereby tied or equivalent. ## Stand in a position Pick a rank. These are the models whose published interval covers it. Rank Stand in a position 1 142 **25** models are not ruled out of rank 40. The cheapest bills **$0.280** per million output tokens and the dearest **$50.00** — a spread of **178.6×** at one position. Stand in a position | Model | Not ruled out of | $/M out | | --- | --- | --- | | [Qwen3.7 Plus](/models/qwen__qwen3.7-plus/) | 26–40 | $1.28 | | [GPT-5.4](/models/openai__gpt-5.4/) | 28–42 | $15.00 | | [MiMo-V2.6-Flash](/models/xiaomi__mimo-v2.6-flash/) | 28–46 | $0.280 | | [Claude Opus 4.5](/models/anthropic__claude-opus-4.5/) | 30–43 | $25.00 | | [GPT-5.6 Terra](/models/openai__gpt-5.6-terra/) | 31–47 | $12.00 | | [Grok 4.5](/models/x-ai__grok-4.5/) | 31–49 | $6.00 | | [DeepSeek V4 Pro 0423](/models/deepseek__deepseek-v4-pro/) | 31–56 | $3.14 | | [GLM 5](/models/z-ai__glm-5/) | 33–50 | $2.24 | | [GPT-6.1 Sol](/models/openai__gpt-6.1-sol/) | 31–56 | $10.00 | | [Kimi K2.5](/models/moonshotai__kimi-k2.5/) | 34–50 | $3.00 | | [Gemma 4 31B](/models/google__gemma-4-31b-it/) | 32–58 | $0.400 Price under confirmation since Nov 4, 2026 | | [Claude Sonnet 5](/models/anthropic__claude-sonnet-5/) | 35–56 | $10.00 | | [Step 5 Preview](/models/stepfun__step-5-preview/) | 33–56 | $2.70 | | [Inkling](/models/thinkingmachines__inkling/) | 35–56 | $4.05 | | [GPT-6 Astra](/models/openai__gpt-6-astra/) | 35–59 | $50.00 | | [GLM 4.6](/models/z-ai__glm-4.6/) | 37–57 | $1.75 | | [Hy3](/models/tencent__hy3/) | 36–63 | $0.528 | | [Claude Sonnet 4.5](/models/anthropic__claude-sonnet-4.5/) | 40–58 | $15.00 | | [Qwen3.8 27B](/models/qwen__qwen3.8-27b/) | 38–62 | $2.55 | | [GLM 5V Turbo](/models/z-ai__glm-5v-turbo/) | 37–65 | $4.00 | | [Qwen3.5 397B A17B](/models/qwen__qwen3.5-397b-a17b/) | 40–62 | $2.34 | | [Qwen3.6 Plus](/models/qwen__qwen3.6-plus/) | 40–64 | $1.95 | | [Gemini 3.5 Flash Lite](/models/google__gemini-3.5-flash-lite/) | 40–65 | $2.50 | | [GLM 4.7](/models/z-ai__glm-4.7/) | 40–65 | $1.75 | | [Gemma 4 26B A4B](/models/google__gemma-4-26b-a4b-it/) | 40–69 | $0.330 Price under confirmation since Nov 4, 2026 | The widest position on this board is rank 65: 25 models, spanning 417× in price. ## Every rated model, drawn as the positions its interval does not rule out Each bar is one model. Its length is the span of rank positions the published interval does not rule out; the faint diagonal behind it is the single position the board displays. Colour is price, not quality. This is not a claim that the ranking is wrong. A rank column is a deterministic display of a sort over point estimates, and the benchmark never claimed statistical separation. It is also the only board in this market that publishes an interval at all, which is the only reason this page can exist. 207 models in the catalogue carry no entry on this board. They are absent from the figure and named as absent — never drawn at rank zero, and never counted as scoring badly. Absence of a measurement is not a measurement. ## Try it yourself The blind rank test shows two models with their names hidden and asks you to pick the higher-scoring one — then tells you whether the published intervals can tell them apart at all. [Blind rank test](/blind-rank/) Scores and intervals are LMArena Elo, used under CC BY 4.0 from the official dataset. Prices are list output price per million tokens, as published by each provider. ## Questions this page answers Does this mean the leaderboard is wrong? No. A rank column is a deterministic display of a sort over point estimates, and the benchmark never claimed statistical separation. It is also the only board in this market that publishes an interval for every model, which is the only reason this page can be drawn. Are two models sharing a position tied? No, and this page will not say so. Failure to rule a position out is not evidence of equivalence. The bands are built from each model's own published interval, which ignores that scores are jointly estimated from shared battles, so they are wider than the truth. Why are some models missing from the figure? A model is drawn only if it carries an entry on this board and a paid price. Models with no entry are named as absent rather than drawn at rank zero. Absence of a measurement is not a measurement, and it is never a verdict. Where do the numbers come from? Scores and intervals are LMArena Elo under CC BY 4.0 from the official dataset. Prices are list output price per million tokens as published by each provider. Everything on this page is arithmetic on those two. Copy evidence Download CSV · 142 rows [Input hashes & method version](/data/publications/manifests/5d4c4fe0b25c71badf22f8e487146fa4dec33728cb96eb590d4c545fe54021b7.json) Read citation text and BibTeX 416 of 418 adjacent published positions have overlapping score intervals. The figure covers 142 models with measured rank bands and a paid output-token price. Intervals overlapping a position do not establish equivalence or task success. Prices are USD per million output tokens. As of 2026-11-05. Lens: LMArena Elo (CC BY 4.0). https://undominated.ai/ranges/ Method: https://undominated.ai/methodology/ Arena data: https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset (CC BY 4.0). Derived calculations by Undominated.ai. Publication: sha256:5d4c4fe0b25c71badf22f8e487146fa4dec33728cb96eb590d4c545fe54021b7; calculation SHA-256: fe6e66bba404c2382400785273cdd10c512261febefcd31122d0a3785c1a83b7; permitted inputs: https://undominated.ai/data/publications/manifests/5d4c4fe0b25c71badf22f8e487146fa4dec33728cb96eb590d4c545fe54021b7.json Copy BibTeX Download citation @misc{undominated-ranges-2026-11-05, author = {{Undominated.ai}}, title = {ranges evidence — Undominated.ai}, year = {2026}, url = {https://undominated.ai/ranges/}, note = {as of 2026-11-05; LMArena Elo (CC BY 4.0)}, annote = {Publication: sha256:5d4c4fe0b25c71badf22f8e487146fa4dec33728cb96eb590d4c545fe54021b7; calculation SHA-256: fe6e66bba404c2382400785273cdd10c512261febefcd31122d0a3785c1a83b7; permitted inputs: https://undominated.ai/data/publications/manifests/5d4c4fe0b25c71badf22f8e487146fa4dec33728cb96eb590d4c545fe54021b7.json} } ## Continue your investigation - [Inspect interval overlap](/significance/) - [Compare exact models](/compare/) - [Compare independent measurements](/disagreement/) - [Read ranking rules](/methodology/)