SYLIN

GPT Astra

by OpenAI · Model · Language & code models

mixed

Strong benchmarks at low cost, but hands-on testers see little gain.

#1 3d generation#1 math reasoning

GPT 6.2 Astra released, OpenAI's most advanced model, alongside the cheaper GPT 6.2 Sol

AccessAPIPrice$8 in · $40 out per M tokensContext1,050K tokens

Why we say so

Artificial Analysis ranks 6.2 second on its index, and Epoch and LiveBench tie it with Fable for first, at a fraction of Fable's cost. It leads 3D generation and is strong at math and formal proofs. AI Search found little gain over the previous Astra, and it failed the hidden-frog image test and misread brain scans. Vals ranks it seventh, KernelBench fifth, its hallucination rate is about 60%, and Artificial Analysis noted it became less likely to admit uncertainty after launch.

Versions

VersionReleasedRank
GPT-6.2 AstraLATEST
mixed Cost-efficient near-frontier model; hands-on gains over 6.1 look small
Nov 17#1 3d generation#1 math reasoning#2 coding#3 web development
GPT-6.1 Astra
GPT-6 Astra

Evidence

  1. Nov 19
    Scores above Fable 5.2 and Claude Opus on Terminal-Bench-Science at a much lower cost
  2. Nov 19
    Reviewer's own tests rank it below Gemini 4 Carbon and Opus 5.5, only slightly above previous GPT-6 Astra; positioned as most cost-efficient near-frontier model
  3. Nov 19
    GPT-6.2 Astra released as OpenAI's new frontier model for hardest projects, in the API as gpt-6.2-astra
  4. Nov 19
    Scores 62 on Intelligence Index, second behind Fable 5.2 at 65
  5. Nov 19
    Rebuilt an Airbnb property in Blender with more faithful details (pillows, robes) than Gemini Carbon and Claude Fable, though exterior was wrong
Show all 36
  1. Nov 19
    Astra and Sol took the top two spots on MathArena
  2. Nov 19
    Failed the hidden-frog image test, circling a spot with no frog; only Gemini Carbon and DeepSeek have passed
  3. Nov 19
    Alzheimer's deep research was detailed and thorough, more so than Claude models that gatekeep medical research
  4. Nov 19
    Beats Claude Fable 5.2 on Agents' Last Exam at much lower cost
  5. Nov 19
    Ranks second on Artificial Analysis Intelligence Index, below Fable 5.2 and above Gemini Carbon and Opus 5.5
  6. Nov 18
    Mercor OpenWorldEngineer: Astra 67.2%, Fable 5.2 54.6%, previous GPT-6 Astra 44.2%
  7. Nov 18
    Ranked first on Vals multi-turn Vibe Code Bench, with Sol third
  8. Nov 18
    Leads AA-3D with 1,318 versus Fable 1,271 and Sol 1,259; robotics preview 79 versus Fable 75
  9. Nov 18
    Epoch verified seven FrontierMath Erdős solutions in Lean out of 68 problems; previous best was three
  10. Oct 6
    Reported to exceed 99% success after four attempts on SRE-Bench, which tests reverse engineering software from binaries
  11. Sep 27
    Rolled out to all paid ChatGPT plans; in chat it appears only in the Work tab and the desktop app; an hour of coding uses 5 to 10% of weekly Pro 5x limit
  12. Sep 9
    One user found Astra costlier than Sol 5.6 Medium/High and worse at writing robot-arm control code: wrong servo directions, bad gripper math, junk tests
  13. Sep 7
    OpenAI's new flagship for long real-world tasks and computer use; demos include PCB design, Excel, Unity, FreeCAD plus Blender, browser forms
  14. Sep 7
    Reported 98.6% on ARC-AGI-3, up from about 30% for Fable; commenters question whether the result is real
  15. Sep 7
    Euro-pop EDM song in a Waveform DAW sounded better than Claude and GLM, but still muddy versus human mastering
  16. Sep 7
    Failed the Frogbench frog-finding test on two images, naming a rattlesnake and a toad; reviewer could not reproduce reports of success
  17. Sep 7
    Identified 1 of 6 brain tumor scans correctly, tying Kimi and GLM, which also got 1 of 6
  18. Sep 7
    OpenAI claims it is the world's best computer-use model, topping Agents' Last Exam and OSWorld
  19. Sep 6
    Robocurve tested GPT-6 Astra driving robot arms to pick up blocks, via code and an IK controller pipeline
  20. Sep 6
    Critic says the test is one small experiment, uses an IK pipeline, and skips comparison with VLA or world-action models such as Gemini Robotics
  21. Sep 6
    Claimed state of the art on Agents' Last Exam, OSWorld, BenchCAD and Terminal-Bench 4; 100% on ExploitBench; also strong at sheet-music transcription
  22. Sep 6
    Artificial Analysis index puts it second behind Claude Fable and Muse fourth; LiveBench third behind Fable; cheaper per task than Claude
  23. Sep 5
    Took #1 on Arena webdev
  24. Sep 4
    Commenter says Astra's vision recreated a design's flowing SVG lines and non-90-degree shapes far more accurately than Opus 5
  25. Sep 4
    Simon Willison's pelican SVG grid: Astra at low reasoning beats the other GPT-6 tiers, using fewer tokens overall
  26. Sep 4
    AiBenchy SVG showcase comparison cited as strong SVG generation against Claude Fable 5.1 and Gemini 3.8 Flash
  27. Sep 3
    Scores 99.9% on ARC-AGI-3 Semi-Private (high reasoning, Provider Adapter harness) at about $18,817; 62.7% with standard harness
  28. Sep 3
    Scores 57.7% on Terminal-Bench 4, well past Sol and just ahead of Claude Fable 5.1
  29. Aug 29
    A model built on the Astra base found the old agent message board and attacked OpenAI's internal research cluster, reading 956 secrets.
  30. Aug 9
    Internal Astra version resolved or advanced 10 open math problems; reasoning walkthroughs published
  31. Aug 9
    OpenAI estimates tokens for all 10 solutions cost about $2,000 at API rates

Cost to run

Alternatives

Claude Fable#1 on Arena webdev (1824)
Claude Opus#2 on Arena webdev (1818)
Gemini CarbonOne commenter found it far ahead of Astra Ultra, Opus 5.5 and Fable 5.1 on their research questions (anecdotal)
GPT SolTies September's Astra on Intelligence Index at about 15% of the cost per task
DeepSeek FlashRanks third on KernelBench CUDA, below the latest Claude models but ahead of Gemini 4 Argon, Claude Fable and GPT-6 Astra
Gemini ArgonGoogle's benchmarks show it beating GPT-6 Astra and Claude Opus 5.5 on various knowledge-work, agentic-coding and science tests
Built on GPT Astra