Claude Sonnet 5.5 vs GPT-6 Astra
| Benchmark | Sonnet 5.5 | GPT-6 Astra |
|---|---|---|
| Humanity's Last Exam | 45.8 | 52.7 |
| Arena Text · overall | 1,475 · xhigh | 1,477 · max |
| Arena Text · Chinese | 1,534.6 · xhigh | 1,489.3 · max |
| Arena Text · English | 1,485.5 · xhigh | 1,447.6 · max |
| Arena Text · French | 1,494.6 · xhigh | 1,424.7 · max |
| Arena Text · instruction following | 1,490.5 · xhigh | 1,455.3 · max |
| Arena Text · Japanese | 1,459.9 · xhigh | 1,403.7 · max |
| Arena Text · non-English | 1,455.5 · xhigh | 1,432.2 · max |
| Arena Text · Russian | 1,458.1 · xhigh | 1,441.5 · max |
| Arena Text · Spanish | 1,446 · xhigh | 1,422.5 · max |
| AA Omniscience Index | 21 | 42.2 |
| SimpleQA Verified v1 | 46.5 · max | 75.8 |
| CritPt (physics) | 24.6 | 29.1 |
| AA Intelligence Index 4.3.2 | 46.8 | 49.6 |
| AA Omniscience · Accuracy | 52 | 60.6 |
| AA Omniscience · Hallucination Rate | 64.6 | 46.5 |
| AA-LCR (long context) 1.1 | 78 | 79.7 |
| GDP.pdf · AA all-pass | 25.2 | 30.4 |
| MLCR-AA | 18.9 | 38.3 |
| ARC-AGI-1 · public evaluation | 92.4 | 98.8 |
| ARC-AGI-1 · semi-private | 92 | 97.5 |
| ARC-AGI-2 · public evaluation | 70.7 | 96.7 |
| ARC-AGI-2 · semi-private | 67.4 | 92.1 |
| ARC-AGI-3 · semi-private | 1.2 | 38.6 |
| Chess Puzzles | 59 · max | 74 · max |
| FrontierMath Erdos | 2.9 · max | 2.9 · max |
| FrontierMath Tier 4 (v2) | 80.5 · max | 97.6 |
| FrontierMath Tiers 1–3 (v2) | 88.8 · max | 93.7 · max |
| Mystery Game Puzzles | 65 · max | 84 · max |
| Vals MedCode | 52.9 | 48.5 |
| Vals MedScribe | 91.1 | 87.9 |
| Vals Legal Research Bench | 48.1 | 39.4 |
| Vals Science Index | 59.1 | 65.1 |
| Arena Text · German | — | 1,454.7 · max |
| Arena Text · Korean | — | 1,427.2 · max |
| Arena Text · Polish | — | 1,427.4 · max |
| MathArena · ArXivMath · April 2026 | — | 90.8 · max |
| MathArena · ArXivMath · August 2026 | — | 88.6 · max |
| MathArena · ArXivMath · June 2026 | — | 94.4 · max |
| MathArena · ArXivMath · May 2026 | — | 95 · max |
| MathArena · BrokenArXiv · April 2026 | — | 97.5 · max |
| MathArena · BrokenArXiv · August 2026 | — | 81.9 · max |
| MathArena · BrokenArXiv · June 2026 | — | 99.1 · max |
| MathArena · BrokenArXiv · May 2026 | — | 91 · max |
| MathArena · ArXivLean · June 2026 | — | 65.2 · max |
| MathArena · ArXivLean · September 2026 | — | 37.7 · max |
Specs
| Spec | Sonnet 5.5 | GPT-6 Astra |
|---|---|---|
| Input, per 1M | $2 | $10 |
| Output, per 1M | $10 | $50 |
| Cost per task? | $0.88 | $1.54 |
| Speed, tok/s | 106 | 59 |
| Context | 1M | 1.05M |
| Weights | Closed | Closed |
| Released | Sep 2026 | Sep 2026 |
| See images | ✓ | ✓ |
| Read PDFs | ✓ | Partial |
| Hear audio | — | — |
| Watch video | — | — |
| Call tools | ✓ | ✓ |
| Force a tool call | — | Partial |
| Return strict JSON | ✓ | ✓ |
| Search the web | ✓ | ✓ |
| Run code | ✓ | ✓ |
| Use a computer | ✓ | ✓ |
What breaks when you switch
Monitoring can return HTTP 403›
Misalignment monitoring can return HTTP 403 before streaming or errors after output; match misalignment_policy_violation and handle stream errors.
No none reasoning effort›
The none reasoning effort is unsupported on GPT-6 Astra and GPT-6.1 Sol; use low instead.
Remove unsupported sampling parameters›
When reasoning effort is not none, remove temperature, top_p and top_logprobs, plus Chat Completions logprobs or Responses output logprobs inclusion.
Context: 1.05M → 1M
No forced tool use›
Forced tool_choice any or tool returns HTTP 400; remove the forced choice.
Replace disabled thinking›
Disabled thinking returns HTTP 400; use between_tools for the lowest thinking setting.
No manual thinking budgets›
Manual thinking budgets and extra between_tools fields return HTTP 400; remove manual budgets and send between_tools without display, budget_tokens or block_binding.
Remove sampling parameters›
Non-default temperature, top_p or top_k returns HTTP 400; remove these parameters.
No assistant prefill›
A prefilled final assistant turn returns HTTP 400; remove that prefill.
Old computer tool rejected · Claude API / Google Cloud›
The computer_20251124 tool returns HTTP 400 on the Claude API and Google Cloud; use computer_toolset_20260801 there.