Updated 1 Nov 2026 · 34 models · 3 benchmarks
Best AI models right now
On Humanity's Last Exam, GPT-6 Astra at 54.8% (±1.94) leads. For coding, GPT-6 Astra at 74.1% (±2.9) leads DeepSWE v1.1 (full coding ranking). Every score comes from an official leaderboard or the model maker, with its source.
Ask an AI about this pageGChatGPTCClaudePPerplexityGeGeminiXGrok
Cite
Cite this page
ModelsAtlas. “Best AI models right now”. Retrieved 1 Nov 2026. https://modelsatlas.com/best
Ranked by Humanity's Last Exam
accuracy, % of 2,500 expert questions answered correctly
| # | Model | Humanity's Last Exam | Setting | Price / 1M (in / out) | Points per $ |
|---|---|---|---|---|---|
| 1 | GPT-6 AstraOpenAI | 54.8%asserted Official leaderboard states
| leaderboard entry 'GPT 6 Astra' | $10.00 / $50.00observed We measured
| 2.7 |
| 2 | Claude Fable 5.1Anthropic | 46.5%asserted Official leaderboard states
| leaderboard entry 'Fable 5.1 (xhigh)' | $10.00 / $50.00observed We measured
| 2.3 |
| 3 | Gemini 3.1 Pro PreviewGoogle | 46.44%asserted Official leaderboard states
| leaderboard entry 'gemini-3.1-pro-preview (thinking high)' | $2.00 / $12.00observed We measured
| 10.3 |
| 4 | Claude Sonnet 5.5Anthropic | 46.12%asserted Official leaderboard states
| leaderboard entry 'Sonnet 5.5 (xhigh)' | $2.00 / $10.00observed We measured
| 11.5 |
| 5 | Gemini 3.8 FlashGoogle | 44.52%asserted Official leaderboard states
| leaderboard entry 'Gemini 3.8 Flash' | $0.83 / $4.13observed We measured
| 27.0 |
| 6 | GPT-5.4 ProOpenAI | 44.32%asserted Official leaderboard states
| leaderboard entry 'gpt-5.4-pro-2026-03-05' | $30.00 / $180.00observed We measured
| 0.7 |
| 7 | GPT-5.4OpenAI | 36.24%asserted Official leaderboard states
| leaderboard entry 'gpt-5.4-2026-03-05 (xhigh thinking)' | $2.50 / $15.00observed We measured
| 6.4 |
| 8 | Claude Opus 4.7Anthropic | 36.2%asserted Official leaderboard states
| leaderboard entry 'claude-opus-4-7' | $5.00 / $25.00 | 3.6 |
| 9 | Claude Opus 4.6Anthropic | 34.44%asserted Official leaderboard states
| leaderboard entry 'claude-opus-4-6-thinking-max' | $5.00 / $25.00 | 3.4 |
| 10 | Claude Haiku 5.5Anthropic | 30.12%asserted Official leaderboard states
| leaderboard entry 'Claude Haiku 5.5' | $0.10 / $0.50observed We measured
| 150.6 |
Best for coding: DeepSWE v1.1
All 21 →pass@1, % of 113 long-horizon software tasks solved (mean of 4 runs)
| # | Model | DeepSWE v1.1 | Setting | Price / 1M (in / out) | Points per $ |
|---|---|---|---|---|---|
| 1 | GPT-6 AstraOpenAI | 74.1%asserted Official leaderboard states
| xhigh reasoning effort | $10.00 / $50.00observed We measured
| 3.7 |
| 2 | Gemini 3.8 FlashGoogle | 73.8%asserted Official leaderboard states
| high reasoning effort | $0.83 / $4.13observed We measured
| 44.7 |
| 3 | Claude Opus 5Anthropic | 73.6%asserted Official leaderboard states
| max reasoning effort | $5.00 / $25.00 | 7.4 |
| 4 | GPT-5.6 SolOpenAI | 72.7%asserted Official leaderboard states
| max reasoning effort | $4.00 / $20.00observed We measured
| 9.1 |
| 5 | Claude Fable 5Anthropic | 69.9%asserted Official leaderboard states
| xhigh reasoning effort | $10.00 / $50.00 | 3.5 |
| 6 | GPT-5.6 TerraOpenAI | 69.6%asserted Official leaderboard states
| max reasoning effort | $2.00 / $12.00observed We measured
| 15.5 |
| 7 | Grok 4.6xAI | 67.5%asserted Official leaderboard states
| medium reasoning effort | $2.00 / $6.00observed We measured
| 22.5 |
| 8 | GPT-5.6 LunaOpenAI | 67.2%asserted Official leaderboard states
| max reasoning effort | $0.20 / $1.20observed We measured
| 149.3 |
| 9 | GPT-5.5OpenAI | 67%asserted Official leaderboard states
| xhigh reasoning effort | $5.00 / $30.00observed We measured
| 6.0 |
| 10 | Gemini 3.7 FlashGoogle | 65.5%asserted Official leaderboard states
| medium reasoning effort | $0.75 / $3.75observed We measured
| 43.7 |
| # | Model | Terminal-Bench 4.0 | Setting | Price / 1M (in / out) | Points per $ |
|---|---|---|---|---|---|
| 1 | Claude Opus 5.5Anthropic | 64.85%asserted Official leaderboard states
| max reasoning effort, Claude Code agent | $4.00 / $20.00observed We measured
| 8.1 |
| 2 | Claude Sonnet 5.5Anthropic | 61.82%asserted Official leaderboard states
| max reasoning effort, Claude Code agent | $2.00 / $10.00observed We measured
| 15.5 |
| 3 | GPT-6 AstraOpenAI | 58.18%asserted Official leaderboard states
| max reasoning effort, Codex agent | $10.00 / $50.00observed We measured
| 2.9 |
| 4 | GPT-6.1 SolOpenAI | 58.18%asserted Official leaderboard states
| max reasoning effort, Codex agent | $2.00 / $10.00observed We measured
| 14.5 |
| 5 | Claude Fable 5.1Anthropic | 57.88%asserted Official leaderboard states
| max reasoning effort, Claude Code agent | $10.00 / $50.00observed We measured
| 2.9 |
| 6 | Claude Opus 5Anthropic | 53.94%asserted Official leaderboard states
| xhigh reasoning effort, Claude Code agent | $5.00 / $25.00 | 5.4 |
| 7 | GPT-6 SolOpenAI | 49.39%asserted Official leaderboard states
| max reasoning effort, Codex agent | $2.00 / $10.00observed We measured
| 12.3 |
| 8 | Claude Fable 5Anthropic | 44.55%asserted Official leaderboard states
| max reasoning effort, Claude Code agent | $10.00 / $50.00 | 2.2 |
| 9 | Grok 4.7xAI | 37.58%asserted Official leaderboard states
| xhigh reasoning effort, Grok Build agent | $2.00 / $6.00observed We measured
| 12.5 |
| 10 | GPT-5.6 SolOpenAI | 37.27%asserted Official leaderboard states
| max reasoning effort, Codex agent | $4.00 / $20.00observed We measured
| 4.7 |
Best value
Humanity's Last Exam points per dollar, for models at or above the median score.
| # | Model | Humanity's Last Exam | Setting | Price / 1M (in / out) | Points per $ ↓ |
|---|---|---|---|---|---|
| 1 | Gemini 3.8 FlashGoogle | 44.52%asserted Official leaderboard states
| leaderboard entry 'Gemini 3.8 Flash' | $0.83 / $4.13observed We measured
| 27.0 |
| 2 | Claude Sonnet 5.5Anthropic | 46.12%asserted Official leaderboard states
| leaderboard entry 'Sonnet 5.5 (xhigh)' | $2.00 / $10.00observed We measured
| 11.5 |
| 3 | Gemini 3.1 Pro PreviewGoogle | 46.44%asserted Official leaderboard states
| leaderboard entry 'gemini-3.1-pro-preview (thinking high)' | $2.00 / $12.00observed We measured
| 10.3 |
| 4 | GPT-6 AstraOpenAI | 54.8%asserted Official leaderboard states
| leaderboard entry 'GPT 6 Astra' | $10.00 / $50.00observed We measured
| 2.7 |
| 5 | Claude Fable 5.1Anthropic | 46.5%asserted Official leaderboard states
| leaderboard entry 'Fable 5.1 (xhigh)' | $10.00 / $50.00observed We measured
| 2.3 |
How we rank
- We do not run benchmarks. Every score is copied from the benchmark's official leaderboard or the model maker's own model card or launch post. Open any score to see the page we read, the date and the eval setting.
- Each benchmark is ranked on its own. We never add or average scores across benchmarks, because they measure different things.
- One row per model: the best configuration the source lists. The Setting column shows the reasoning effort and agent harness.
- Highlighted scores are ties. They fall within the leader's 95% margin of error, so the order between them is not meaningful.
- Points per $ = score ÷ blended price per 1M tokens (3 input tokens to 1 output, cheapest standard offer we list). It does not count how many tokens a model spends per task, so use it as a rough guide. The best-value tables only include models at or above the median score.
- No score, no rank. A model missing from a table has no published score on that benchmark. That is not a verdict on the model.
Questions
What is the best AI model right now?
On Humanity's Last Exam, a general reasoning test, GPT-6 Astra at 54.8% (±1.94) leads. For coding, GPT-6 Astra at 74.1% (±2.9) leads DeepSWE v1.1. No single benchmark settles it, so we rank each one separately.
Which AI model is the best value?
Among models at or above the median Humanity's Last Exam score, Gemini 3.8 Flash gives the most points per dollar: 44.52% at $0.83 input and $4.13 output per 1M tokens.
How are AI models ranked here?
By scores published on the official leaderboards of DeepSWE v1.1, Terminal-Bench 4.0 and Humanity's Last Exam, or by the model maker. We do not run benchmarks, and we never average scores across benchmarks. Every score links to the page it came from.
Keep going in your AI
Opens with this page as the source. We add nothing else to the prompt.
- What is the best AI model right now?
- Which AI model is the best value?
- How are AI models ranked here?