OpenAI Releases GPT 6.2 Astra & GPT 6.2 Sol, Beat Fable 5.2, Gemini 4 Carbon On Some Benchmarks

OpenAI has hit back barely a week after Anthropic and Google reshuffled the top of the AI leaderboards.

The company has released GPT-6.2 Astra, its new frontier model for the hardest professional and technical work, and GPT-6.2 Sol, a cheaper, faster model it is positioning as the default for everyday tasks. OpenAI’s headline claim is efficiency: both models are said to deliver better results with fewer output tokens, and the benchmark charts it published show GPT-6.2 Astra matching or beating Claude Fable 5.2 and Gemini 4 Carbon on several evaluations at a cheaper cost per task.

The release skips a version number in the Astra line. OpenAI had scrapped the planned October release of GPT-6.1 Astra after internal testing showed an alignment regression, so GPT-6.2 Astra is the first upgrade to GPT-6 Astra since it launched in September. GPT-6.2 Sol, meanwhile, succeeds GPT-6.1 Sol.

GPT-6.2 Astra vs GPT-6.2 Sol: What’s New

OpenAI describes GPT-6.2 Astra as setting “a new quality bar for AI in complex professional work.” The company says its slides, spreadsheets and documents read less like AI output and more closely follow user-supplied templates and the conventions of fields like law, finance, sales and accounting. It also claims cleaner code, better-tested changes in existing codebases, and stronger frontend and 3D work for websites and games.

GPT-6.2 Sol is pitched as the workhorse. OpenAI says it is smarter and more efficient than GPT-6.1 Sol, handles multistep agentic workflows well, and lets paid users complete more tasks within the same usage limits. According to the company, many OpenAI employees now use it as their default model for both technical and non-technical work.

GPT-6.2 Astra Benchmarks

OpenAI published most of its results as score-versus-cost charts rather than a single table, plotting each model at several reasoning-effort settings. That framing is interesting: in several cases GPT-6.2 Astra doesn’t post the highest raw score, but gets close to the top at a much lower cost per task. Figures below marked “~” are read from OpenAI’s charts and are approximate.

Agents’ Last Exam v1.1

Agents’ Last Exam tests long-horizon, multistep professional workflows. GPT-6.2 Astra sets a new high here, scoring about 78% at roughly $3 per task.

OpenAI says that is 5.2 points higher than GPT-6 Astra at less than a third of the cost, and 5.5 points ahead of Claude Fable 5.2 at just 5% of the cost. Fable 5.2 sits at about 73% at around $60 per task, while Claude Opus 5.5 lands at about 72% at roughly $17.

The more interesting result is Gemini 4 Carbon. Google’s model scores about 77.5% at roughly $4.50 per task, almost exactly level with GPT-6.2 Astra’s highest-effort setting. Astra’s lead over Carbon on this benchmark is under a point.

GPT-6.2 Sol tops out at about 73%, roughly matching both GPT-6 Astra (~73% at ~$9) and Fable 5.2 at a little over a dollar per task.

OpenAI notes that leaderboard results for Fable 5.2 exclude nine tasks that used a fallback model.

Terminal-Bench-Science 0.1

Terminal-Bench-Science tests how well models carry out scientific research workflows in a terminal, and is one of the nine components of the Artificial Analysis Intelligence Index v5. OpenAI calls GPT-6.2 Astra “the best model in the world for scientists” on the strength of this result.

GPT-6.2 Astra scores about 77% at roughly $7 per task, 9 percentage points above GPT-6 Astra (~68% at ~$24). Claude Fable 5.2, run on the mini-swe-agent harness, peaks at about 73% but needs around $45 per task to get there. Claude Opus 5.5 scores about 63%.

GPT-6.2 Sol is the standout value play: about 68% at roughly $2, matching GPT-6 Astra’s best score at under a tenth of the cost and comfortably ahead of GPT-6.1 Sol (~57%).

One caveat: OpenAI ran Fable 5.2 on a third-party harness, and Anthropic’s own reported Terminal-Bench-Science figure for the model is higher. Harness choice can move scores meaningfully on agentic benchmarks.

OpenAI also says GPT-6.2 Astra sets a new high on the Vals Science Index, which asks models to interpret real biological data and run experiments to infer how a simulated system works.

Harvey LAB-AA v1.1 (Legal)

This is GPT-6.2 Astra’s most lopsided win. Harvey LAB-AA tests legal work such as reviewing contract provisions, and only gives credit when a deliverable meets every grading criterion and contains no material hallucinations.

GPT-6.2 Astra hits about 15% at roughly $6 per task. Claude Fable 5.2’s best result is about 7% at close to $40, so Astra more than doubles Fable’s pass rate at about one-sixth of the cost. GPT-6.2 Sol scores about 11% at under $2, also ahead of every non-OpenAI model tested. GPT-6 Astra sits at about 9%, and Claude Opus 5.5 at about 4%.

OpenAI says Astra for Law, a configuration of GPT-6.2 Astra built for lawyers, scores higher still.

BenchCAD 2.1 Core

BenchCAD is where the “some benchmarks” qualifier shows up. Claude Fable 5.2 posts the highest raw score on the chart at about 84%, narrowly ahead of GPT-6.2 Astra’s ~83%.

The cost gap is the story, though. Fable 5.2 needs roughly $50 per task for that score; GPT-6.2 Astra gets within a point for around $5.50, about a ninth of the cost. GPT-6.2 Sol reaches about 73% at under $2, roughly matching Fable 5.2’s ~74% result at $23. GPT-6 Astra peaks at about 63%, and Claude Opus 5.5 at about 65%.

ExploitBench (Internal Port)

On OpenAI’s internal port of ExploitBench, which measures success at exploiting vulnerabilities against output tokens spent, GPT-6.2 Astra reaches about 54% and GPT-6.2 Sol about 50%. That compares with about 32% for GPT-6 Astra and about 22% for GPT-6.1 Sol. No competitor models were included on this chart, and OpenAI says the cyber results were measured with safeguards and classifiers disabled.

Other Benchmarks OpenAI Highlighted

  • FrontierCode 2.0: On Cognition’s benchmark, which judges whether code changes would actually be merged by maintainers, GPT-6.2 Sol comes within 0.5 points of GPT-6.2 Astra at about one-fifth the cost, with both at max effort.
  • AA-3D: On Artificial Analysis’s blinded human-preference test for interactive 3D experiences, GPT-6.2 Astra ranks first with 1,318 Elo, ahead of Claude Fable 5.2 (1,271) and GPT-6.2 Sol (1,259).
  • AA-AnalystAgent: GPT-6.2 Astra sets a new high, answering 62.5% of questions correctly on all five independent attempts. The benchmark tests finding and combining figures from real spreadsheets and reports.
  • GDPval v2 (internal): GPT-6.2 Astra leads the models OpenAI tested, up 6.2 points on GPT-6 Astra while using 22% fewer output tokens. GPT-6.2 Sol gains 5.3 points over GPT-6.1 Sol.
  • LifeVal: On OpenAI’s evaluation of challenging everyday tasks, GPT-6.2 Sol outscores GPT-6 Astra at about one-seventh the cost.

OpenAI also published charts for the Artificial Analysis Intelligence Index v5.0.1 and AutomationBench, where it says both models come close to or beat other leading models at a small fraction of the cost per task. Independent Index results will be the real test: Fable 5.2 currently leads that index with 65 points, five clear of Gemini 4 Carbon, with GPT-6 Astra at 57.

GPT-6.2 Astra Benchmarks At A Glance

BenchmarkGPT-6.2 AstraGPT-6.2 SolGPT-6 AstraClaude Fable 5.2Claude Opus 5.5Gemini 4 Carbon
Agents’ Last Exam v1.1~78% (~$3)~73% (~$1.3)~73% (~$9)~73% (~$60)~72% (~$17)~77.5% (~$4.5)
Terminal-Bench-Science 0.1~77% (~$7)~68% (~$2)~68% (~$24)~73% (~$45)*~63% (~$23)—
Harvey LAB-AA v1.1~15% (~$6)~11% (~$2)~9% (~$14)~7% (~$40)~4% (~$18)—
BenchCAD 2.1 Core~83% (~$5.5)~73% (~$1.6)~63% (~$10)~84% (~$51)~65% (~$20)—
AA-3D (Elo)1,3181,259—1,271——

Best score at each model’s highest plotted setting; approximate cost per task in brackets. Values are read from OpenAI’s published charts. *Fable 5.2 on the mini-swe-agent harness.

The Benchmark Caveats

These are OpenAI’s own numbers, on benchmarks and configurations OpenAI chose. Several of the evaluations, including GDPval v2, LifeVal, SecureSWE and the ExploitBench port, are internal. The comparison set is also selective: Gemini 4 Carbon appears on Agents’ Last Exam and a handful of other charts, but not on Terminal-Bench-Science, Harvey LAB-AA or BenchCAD.

The same pattern played out with Anthropic’s launch a week ago, when Fable 5.2 was shown beating GPT-6 Astra and Gemini 4 Carbon on most of its published benchmarks, and with Google before that. Independent results have already complicated those claims: Gemini 4 Carbon, for example, edged out Fable 5.2 on Perplexity’s WANDR benchmark at about a third of the cost.

Where OpenAI’s case looks strongest is cost. Even where GPT-6.2 Astra only roughly matches Fable 5.2 on score, it frequently does so at a fifth to a tenth of the per-task cost on these charts.

GPT-6.2 Astra and GPT-6.2 Sol Pricing

GPT-6.2 Astra is cheaper than its predecessor. API input and output prices are 20% lower than GPT-6 Astra’s, and cached input gets a 60% discount:

ModelInput (per 1M tokens)Cached input (per 1M tokens)Output (per 1M tokens)
GPT-6.2 Astra$8.00$0.40$40.00
GPT-6.2 Sol$2.00$0.10$10.00

That puts GPT-6.2 Astra at the same list price as Claude Fable 5.2, which costs $8 per million input tokens and $40 per million output tokens. OpenAI’s argument is that Astra’s token efficiency makes each task substantially cheaper in practice. GPT-6.2 Sol keeps GPT-6.1 Sol’s rates unchanged.

OpenAI recommends keeping prompt prefixes unchanged and appending new instructions to the conversation to improve cache reuse, cut costs and latency, and preserve reasoning across turns.

Availability

GPT-6.2 Astra and GPT-6.2 Sol are available from today to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. Enterprise and Edu users get access once their workspace admin enables it.

Pro 500 subscribers can use the Ultrafast mode with either model, and for the first time OpenAI is bringing Ultrafast to Pro 200 subscribers, starting with GPT-6.2 Sol. Developers can access gpt-6.2-astra and gpt-6.2-sol through the OpenAI API, with Ultrafast available for both at launch.

Real-World Use Cases: Slides, Spreadsheets and CAD

OpenAI leaned heavily on professional deliverables in the launch. It says GPT-6.2 Astra is its best model yet at designing slides and following slide templates, matching layout, colours and typography down to exact font sizes even from a screenshot. In one demo, both GPT-6 Astra and GPT-6.2 Astra were asked to update a retailer’s Q1 earnings deck using its Q2 release and 10-Q; OpenAI says the newer model produced sharper headlines and more concrete next steps for management.

On computer use, OpenAI says GPT-6.2 Astra can work down to individual parts and connections in CAD and simulation tools. Demos included modelling the chassis dynamics of a car on a mountain road with a custom Python solver animated in Blender, building an animated assembly guide for an RC buggy from public photos and manuals, and designing a paper lamp inspired by Chicago’s Marina City towers.

Early customers quoted by OpenAI include Basis, which said its average time for long-horizon accounting work fell by 50%, and Higgsfield AI, which said GPT-6.2 Astra hit the target result in one shot in 73% of its internal 3D generation evals.

Safety: First OpenAI Model At “High” For AI Self-Improvement

Like GPT-6 Astra, both new models meet the Critical threshold for cybersecurity and the High threshold for biological and chemical capabilities under OpenAI’s Preparedness Framework. OpenAI is launching them with tighter cyber safeguards, which it acknowledges will restrict some legitimate vulnerability research. Approved defenders can still access GPT-6 Astra and GPT-6 Cyber through OpenAI Daybreak, with GPT-6.2 models to follow.

The bigger disclosure is that GPT-6.2 Astra is OpenAI’s first externally released model to meet the High threshold for AI self-improvement. OpenAI says this designation means the model’s impact is equivalent to giving every OpenAI researcher an additional highly capable research engineer, compared with a 2024 baseline. In response, the company says it has strengthened containment in its research environments and built Control Room, a system that tracks research workloads and can halt model output when intervention is needed.

On alignment, OpenAI says both models performed similarly to GPT-6 Astra on most evaluations and took fewer improper actions in simulated workplace tasks. However, in two tests they were more likely to stay silent about obstacles blocking their work, such as a missing file or an unavailable search tool. OpenAI says it has added instructions in Codex that substantially reduce these omissions and is continuing to investigate.

What It Means

The frontier has been changing hands every couple of weeks. Gemini 4 Carbon took the top spot on the Artificial Analysis Intelligence Index in late October, Claude Fable 5.2 reclaimed it for Anthropic on November 9, and Google moved Carbon towards general availability days later. GPT-6.2 Astra is OpenAI’s answer, and its pitch is less about raw capability than about delivering near-frontier results at a much lower cost per task.

Whether that holds up will become clearer as independent evaluators like Artificial Analysis publish their own numbers over the coming days. For enterprises choosing a default model, though, a model that lands within a point or two of the leaders for a fraction of the cost may matter more than who sits at the top of any one leaderboard.

FAQs

What is GPT-6.2 Astra? GPT-6.2 Astra is OpenAI’s new frontier model for complex professional, coding and scientific work. It succeeds GPT-6 Astra; OpenAI cancelled a planned GPT-6.1 Astra release in October.

What is GPT-6.2 Sol? GPT-6.2 Sol is OpenAI’s efficient, general-purpose model for everyday work. It replaces GPT-6.1 Sol at the same API price.

Does GPT-6.2 Astra beat Claude Fable 5.2? On OpenAI’s published charts, GPT-6.2 Astra beats Fable 5.2 on Agents’ Last Exam, Terminal-Bench-Science, Harvey LAB-AA and AA-3D. Fable 5.2 posts a marginally higher top score on BenchCAD, though at about nine times the cost per task.

Does GPT-6.2 Astra beat Gemini 4 Carbon? On Agents’ Last Exam, GPT-6.2 Astra edges Gemini 4 Carbon by under a point (~78% vs ~77.5%). Carbon was not included on most of OpenAI’s other comparison charts.

How much does GPT-6.2 Astra cost? $8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens, 20% cheaper than GPT-6 Astra on input and output.

Who can use GPT-6.2 Astra and Sol? Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and developers via the OpenAI API as gpt-6.2-astra and gpt-6.2-sol.

Posted in AI