GPT-6 Astra
GPT-6 Astra is a closed AI model from OpenAI with an index score of 193 (GPT-4 = 100), rank 6 of 217.
193178–209Index score#6 of 217 modelsPro mode: 196 · $10 / $50
$10/ $50Price per million tokensinput / output · via OpenRouter
1.05MContext windowtokens it reads at once
4 Sep 2026ReleasedReplaced by GPT-6.2
Where it sitson the frontier
Among the first models to reach this score.
Test resultsby area · tick = best by any model
Reasoning & knowledgeevidence
Mathsevidence
Codingevidence
Agentsevidence
Professional workevidence
Novel problemsevidence
Long tasksno evidence yet
Third-party results only, at the model’s best standard setting. Hover a bar for who ran it.
What the lab says8 of 18 also run by third parties
| Test | OpenAI says | Third parties |
|---|---|---|
| Terminal-Bench 4.0 | 57.9% | 59.6%Vals AI |
| DeepSWE v1.1 | 74.1% | 74.1%Datacurve |
| Terminal-Bench Science 0.1 | 64.6% | 62.9%Vals AI |
| FrontierMath Tier 4 (v2) | 97.6% | 97.6%Epoch AI |
| GPQA Diamond | 96.0% | 95.8%Epoch AI |
| ARC-AGI-3OpenAI's own harness. OpenAI ran it in its own API harness; ARC Prize's verified result uses the standard harness. | 99.9% | 62.7%ARC Prize |
| ARC-AGI-2 | 95.0% | 95.0%ARC Prize |
| ARC-AGI-1 | 98.5% | 98.5%ARC Prize |
| FrontierCode 1.1 (Main) | 53.3% | not run yet |
| Humanity's Last Examwith tools | 57.2% | not run yet |
| OSWorld 2.0offline set, partial credit | 72.6% | not run yet |
| Agents' Last Exam | 59.3% | not run yet |
| ScreenSpot-Prono tools | 92.7% | not run yet |
| AutomationBench | 41.4% | not run yet |
| BrowseComp | 91.5% | not run yet |
| BenchCAD | 95.9% | not run yet |
| HealthBench Professionallength-adjusted | 63.4% | not run yet |
| OpenAI MRCR v2, 8 needles, 512K–1M tokens | 96.3% | not run yet |
From OpenAI’s launch post, GPT-6 Astra: A new generation of intelligence, copied from the table in the post. Shown, never scored; settings can differ from the third-party run.
What others sayrank on 13 leaderboards
The Model Indexthird-party tests, our method#6of 217 · 193
LMArena · Codingcoding questions in chat#5of 357 · 1548Epoch Capabilities Indexoverall capability from many benchmarks#4of 231 · 169.2LMArena · WebDevbuilding web apps and front ends#5of 133 · 1763LMArena · Instruction followingfollowing detailed instructions#28of 362 · 1474LiveBench · Overalloverall, on regularly refreshed questions#3of 38 · 84.5LMArena · Long promptslong, complex requests#32of 340 · 1485LMArena · Creative writingcreative writing#34of 360 · 1451LMArena · Mathsmaths questions in chat#33of 346 · 1481LMArena · Textoverall chat quality, judged by people#36of 362 · 1474LMArena · Visionunderstanding images#18of 145 · 1283LMArena · Searchanswering with web search#7of 41 · 1231LMArena · Documentsworking with long documents#15of 40 · 1467Best attop three among current models
1stARC-AGI-2ARC Prize · third-party1stHumanity's Last ExamScale AI · third-party2ndanswering with web searchLMArena · Search2ndreasoning questionsLiveBench · Reasoning2ndagentic questionsLiveBench · Agentic2ndFrontierMath (tiers 1-3)Epoch AI · third-party3rdcoding questions in chatLMArena · Coding3rdworking with long documentsLMArena · Documents
Where to use itprices checked 18 Nov 2026
API providers 3 · per 1M tokens, in / out
Host prices from OpenRouter; going direct can differ.
Facts and sources
Released4 Sep 2026
Announcementopenai.com ↗
Release entryGPT-6 Astra (preview 09-03)
Scored asGPT-6 Astra
Pro modeindex 196 · $10 / $50 per M
Output speed–
Hugging Faceclosed weights