GPT-6 Astra
OpenAI·Released Sep 3, 2026
Updated Oct 8, 11:55 PM ET
Overall score
74.1
Rank #3 overall
Evidence
34 results
97% of scored kinds covered
Context window
1.1M tokens
From OpenRouter
API price, USD per 1M tokens
$10 / $50
Input / output · cached input $1
The score is on the board's scale (reference average 50, about 15 points per standard deviation), not a percentage.
By domain
Each domain is scored on the same scale as the overall score. Coverage is the share of the domain's kinds of evaluation with a result; it says how much evidence there is, not how good the model is.
Every result behind the score
The result as the evaluator publishes it, and the reading it gives on the board's scale. A hard evaluation can give a high reading for a modest result; results near 95% or 5% only say "at least" or "at most".
Coding
10 results · 71.2- Terminal-Bench 4.0Vals AI · max effort59.6%Reading 83.8
- Terminal-Bench 4.0 (AA run)Artificial Analysis · high effort54.0%Reading 83.7
- Code MigrationVals AI · max effort67.7%Reading 82.3
- Vibe Code Bench 1–100Vals AI · max effort27.6%Reading 78.8
- Vibe Code BenchVals AI · max effort89.6%Reading 73.4
- ProgramBenchVals AI · max effort5.5%Reading 68.7
- APEX-SWEMercor · max effort50.0%Reading 66.8
- SciCodeArtificial Analysis · high effort55.4%Reading 64.3
- LiveBench CodingLiveBench · max effort68.8%Reading 64.2
- CyberBench PatchVals AI · max effort82.1%Reading 56.8
Research and reasoning
9 results · 91.5- Mystery Game PuzzlesEpoch AI · max effort84.0%Reading 113.1
- FrontierMath Tier 4Epoch AI · high effortNear the ceiling97.6%Reading at least 101.0
- Chess PuzzlesEpoch AI · max effort72.0%Reading 100.6
- Terminal-Bench ScienceVals AI · max effort62.9%Reading 94.3
- FrontierMath Tiers 1–3Epoch AI · max effort93.7%Reading 90.3
- ProofBenchVals AI · max effortNear the ceiling99.0%Reading at least 89.2
- MysteryMechanismVals AI · max effort53.2%Reading 88.8
- LiveBench ReasoningLiveBench · max effort94.7%Reading 84.6
- Humanity's Last Exam (AA run)Artificial Analysis · high effort53.1%Reading 73.9
Professional work
9 results · 63.1- LiveBench Data AnalysisLiveBench · max effort83.0%Reading 86.1
- APEX-AgentsMercor · max effort64.7%Reading 75.0
- EMBVals AI · max effort71.7%Reading 71.3
- τ-Bench Banking (AA run)Artificial Analysis · high effort40.0%Reading 62.9
- MedCodeVals AI · max effort48.5%Reading 62.8
- Legal Research BenchVals AI · max effort39.4%Reading 61.3
- Finance AgentVals AI · max effort53.5%Reading 60.6
- Tax Agent BenchVals AI · max effort20.7%Reading 48.2
- Harvey Legal Agent BenchmarkVals AI · max effortCapped5.4%Reading 29.1
Knowledge and accuracy
5 results · 79.1- SimpleQA VerifiedEpoch AI · max effort75.6%Reading 94.5
- LiveBench LanguageLiveBench · max effort89.4%Reading 92.0
- LiveBench Instruction FollowingLiveBench · max effort75.6%Reading 87.0
- BullshitBenchBullshitBench · max effort47.3%Reading 75.0
- AA-LCRArtificial Analysis · high effort80.0%Reading 49.1
Human preference
1 result · 57.2- Arena TextLMArena · max effort1,477.1Reading 48.5
Shown, not scored
Composite indices, saturated or older evaluations, evaluations still being checked, and vision, writing and multilingual results. They do not move the score.
- IOIVals AI · CodingReference only100.0%
- MirrorCodeEpoch AI · CodingWatching46.7%
- SRE BenchVals AI · CodingWatching56.9%
- BioMysteryBenchVals AI · Research and reasoningWatching79.3%
- OTIS Mock AIMEEpoch AI · Research and reasoningReference only100.0%
- APEX-AccountingMercor · Professional workWatching11.9%
- EBR-benchEpoch AI · Professional workWatching76.2%
- MedScribeVals AI · Professional workWatching87.9%
- GPQA DiamondEpoch AI · Knowledge and accuracyReference only95.8%
- AA Intelligence IndexArtificial Analysis · Composite indicesReference only50.9
- LiveBench averageLiveBench · Composite indicesReference only82.2%
- Vals IndexVals AI · Composite indicesReference only63.1%
- Arena VisionLMArena · VisionReference only1,284.3
- Arena WebDevLMArena · Writing and designReference only1,787.6
No result yet
Scored evaluations this model has not taken. A missing result neither adds nor subtracts.
- τ²-Bench Telecom (AA run) · Artificial Analysis
- IFBench (AA run) · Artificial Analysis