Polarison

AI model benchmarks vs. API price

The most capable model we track is Claude Opus 5.5 (Epoch Capabilities Index 166.8). For the best capability per dollar, look at Claude Haiku 5.5, GPT-6.1 Sol, Claude Opus 5.5: no cheaper model beats any of them.

Benchmark data from Epoch AI, retrieved October 17, 2026.

Capability vs. price

Epoch Capabilities Index (higher is more capable) against blended API price per 1M tokens. Hover or tab to a dot for details.

Best value: no cheaper model scores higherOther modelsConfidence interval
135140145150155160165170175$0.10$0.30$1$3$10$30Blended price per 1M tokens (3 input : 1 output, log scale)Epoch Capabilities IndexClaude Haiku 5.5GPT-6.1 SolClaude Opus 5.5

Best model for each job

Most capable overall
Claude Opus 5.5
Epoch Capabilities Index 166.8
Best for science reasoning
Gemini 4 Argon
GPQA Diamond: 96.4%
Best for advanced math
GPT-6 Astra
FrontierMath (Tiers 1–3): 93.7%
Best for factual accuracy
Gemini 4 Argon
SimpleQA Verified: 76.0%
Best for abstract reasoning
GPT-6 Astra
ARC-AGI-2: 95.0%
Best for coding
GPT-6 Astra
DeepSWE: 74.1%
Best for agentic work
Gemini 4 Argon
APEX-Agents: 82.4%
Most capable under $1 per 1M tokens
Claude Haiku 5.5
ECI 159.0 at $0.20 blended
Most capable under $5 per 1M tokens
GPT-6.1 Sol
ECI 166.3 at $4.00 blended

Leaderboard

Sorted by Epoch Capabilities Index. Prices are blended USD per 1M tokens (3 input : 1 output).

#ModelECIPriceGPQA DiamondFrontierMath (Tiers 1–3)SimpleQA VerifiedARC-AGI-2DeepSWEAPEX-Agents
1Claude Opus 5.5AnthropicBest value166.8(164–171)$8.0090.6%91.2%72.2%93.3%—73.5%
2GPT-6 AstraOpenAI166.4(163–171)$20.0095.8%93.7%75.6%95.0%74.1%64.7%
3GPT-6.1 SolOpenAIBest value166.3(163–171)$4.0095.4%93.7%73.9%94.2%—60.0%
4Claude Fable 5.1Anthropic164.7(162–168)$20.00—90.2%70.8%90.0%—68.6%
5Gemini 4 ArgonGoogle164.6(162–169)$4.0096.4%86.0%76.0%——82.4%
6Claude Sonnet 5.5Anthropic164.4(161–168)$4.0095.6%88.8%46.5%——75.5%
7Claude Opus 5Anthropic162.6(160–166)$10.0093.9%85.6%59.9%90.4%73.6%65.8%
8GPT-6 SolOpenAI162.6(160–166)$4.0094.3%89.8%60.7%89.6%—54.3%
9GPT-5.6 SolOpenAI161.7(159–165)$8.0093.5%89.1%69.7%92.5%72.7%51.4%
10GPT-5.6 TerraOpenAI159.4(157–162)$4.5093.3%86.0%43.2%83.9%69.6%58.2%
11Claude Haiku 5.5AnthropicBest value159.0(157–162)$0.2089.6%75.1%23.8%——56.1%
12Gemini 3.8 FlashGoogle157.2(155–160)$1.5095.4%68.4%69.7%89.2%73.8%64.3%
13Claude Sonnet 5Anthropic156.4(154–159)$4.0090.5%65.6%33.7%—53.8%54.5%
14GPT-5.6 LunaOpenAI156.3(154–159)$0.4591.6%82.1%41.0%59.5%67.2%43.0%
15GPT-6 LunaOpenAI156.3(154–159)$0.2090.5%78.9%41.4%59.3%—44.3%
16Grok 4.6xAI156.3(155–159)$3.0094.0%66.0%49.3%67.1%67.5%65.3%
17DeepSeek V4.1 FlashDeepSeek155.6(154–158)$0.52589.8%67.4%42.9%72.9%—39.5%
18DeepSeek V4 ProDeepSeek155.3(154–158)$1.9891.7%64.6%52.9%61.3%—47.3%
19Gemini 3.1 ProGoogle154.7(152–157)$4.5094.4%59.6%73.5%77.1%11.7%35.3%
20Grok 4.7xAI154.4(153–156)$3.0092.7%53.0%56.0%61.4%—54.6%
21Gemini 3.6 FlashGoogle154.2(153–156)$1.5094.1%58.9%66.2%60.4%46.7%46.9%
22Gemini 2.5 ProGoogle145.3(144–147)$3.43885.3%24.6%—4.9%——
23Gemini 3.5 Flash-LiteGoogle145.2(143–147)$0.8583.3%26.0%—10.3%—29.3%
24Claude Haiku 4.5Anthropic142.4(139–144)$2.0071.2%—13.2%4.0%——
25Gemini 2.5 FlashGoogle140.8(139–142)$0.85——————

Not yet rated by Epoch AI: Grok 4.3.

How to read these scores

The Epoch Capabilities Index combines many benchmarks into one general-capability scale. The numbers in brackets are its confidence interval: when two models’ intervals overlap, treat them as roughly equal rather than declaring a winner.

Each benchmark score is the best result Epoch AI recorded for that model, usually at its highest reasoning setting. Higher reasoning settings produce more output tokens, so a model can cost more in practice than its per-token price suggests.

The benchmarks

  • GPQA Diamond (science reasoning): Graduate-level biology, chemistry and physics questions designed to be hard to look up.
  • FrontierMath (Tiers 1–3) (advanced math): Original, unpublished math problems ranging from advanced undergraduate to research level.
  • SimpleQA Verified (factual accuracy): Short fact-seeking questions that reward correct answers over confident guesses.
  • ARC-AGI-2 (abstract reasoning): Visual pattern puzzles that are easy for people but hard for AI systems.
  • DeepSWE (coding): Agentic software engineering tasks that require changing real codebases.
  • APEX-Agents (agentic work): Long, multi-step professional tasks completed by an AI agent.

A dash means Epoch AI has not published a score for that model on that benchmark. Benchmarks measure narrow skills; test candidates on your own tasks before committing. To turn these numbers into a monthly bill, use the cost calculator or compare two models.

Data: Epoch AI, “Capabilities & Benchmarking”, epoch.ai/benchmarks, licensed CC BY 4.0. Prices from each provider’s official pricing page.

Frequently asked questions

What is the most capable AI model right now?
Among the models we track, Claude Opus 5.5 from Anthropic has the highest Epoch Capabilities Index score (166.8), followed by GPT-6 Astra (166.4) and GPT-6.1 Sol (166.3). Scores this close can fall within each other’s confidence intervals.
Which AI model is best for coding?
On the DeepSWE agentic coding benchmark, GPT-6 Astra scores highest among tracked models at 74.1%.
What is the best value AI model?
These models offer the most capability for their price, because no cheaper model we track scores higher: Claude Haiku 5.5 ($0.20 per 1M tokens), GPT-6.1 Sol ($4.00 per 1M tokens), Claude Opus 5.5 ($8.00 per 1M tokens).
What is the Epoch Capabilities Index?
The Epoch Capabilities Index (ECI), published by Epoch AI, combines results from many AI benchmarks into a single general-capability scale, so models can be compared even after individual benchmarks saturate. Higher is better.