GPT-6.2 Astra vs Sol Reviewed: Benchmarks, Prices, and the Mixed Early Verdict

N
Navs
Published on November 18, 20268 min read
Tags:gpt-6-2openaiastrasolbenchmarkscodexchatgpt
GPT-6.2 Astra vs Sol Reviewed: Benchmarks, Prices, and the Mixed Early Verdict

OpenAI shipped two new models on November 17, 2026: GPT-6.2 Astra, its new flagship, and GPT-6.2 Sol, its everyday model. Both arrived the same day in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users, and in the API as gpt-6.2-astra and gpt-6.2-sol. They are not in regular Chat, and free accounts cannot use them.

On paper it is a clean win: Astra beats its predecessor on all nine launch benchmarks while costing less per task, and Artificial Analysis agrees. The first people to use the models in real work tell a messier story — some call Astra unusable, others built working games with it. This review puts every number on the table.

Scope and method: how to read this review

Navs did not test these models. We have no API access and ran none of the tests below. Every number here comes from one of three sources, and they do not measure the same thing.

First, OpenAI's launch post and system card — the vendor's claims, each model at its best reasoning effort, so settings can differ within one chart. Second, independent evaluators: Artificial Analysis (nine-test index, Max effort unless stated) and Vals AI (fixed research tasks). Third, early users' public reports — small samples, useful for how the models feel to work with, not for ranking them.

Treat the official numbers as the best case, the independent numbers as a second opinion, and the user reports as a warning about the gap.

The models live in ChatGPT Work and Codex — Work for documents and presentations, Codex for coding. Both accept text and images, with up to one million tokens of context. Enterprise and Edu administrators must enable them first.

ChatGPT homepage with the headline Get answers. Find inspiration. Be more productive.

The evidence at a glance

Every key number in this review, with its source.

BenchmarkResultSourceWhat to note
Agents' Last Exam v1.1Astra 78.3% (was 73.1%); Sol 73.0% (was 72.5%)OpenAIBest-effort scores; Astra up 5.2 points
GDPval v2Astra +6.2 points; Sol +5.3 pointsOpenAIProfessional deliverables; absolute scores not published
Terminal-Bench-Science 0.1Astra 77.1% (was 68.1%); Sol 68.0% (was 57.0%)OpenAI70 expert-written research tasks
FrontierCode 2.0Astra 44.2% (was 42.8%); Sol 43.8% (was 39.0%)OpenAI / CognitionSol within 0.5 points of Astra at max effort
LifeValAstra 86.1% (was 79.8%); Sol 83.8% (was 79.5%)OpenAIEveryday life tasks
Vals Science IndexAstra 68.4% (was 65.1%); Sol 58.5% (was 59.6%)OpenAISol's only regression in nine tests
Harvey-LAB-AA v1.1Astra 15.0% (was 8.6%); Sol 11.1% (was 6.9%)OpenAILegal work, strict pass rate
Artificial Analysis Intelligence IndexAstra 62 (was 57); Sol 57 (was 55)Artificial AnalysisMax effort; Fable 5.2 leads at 65
Artificial Analysis cost per index taskAstra $4.00 (was $6.24); Sol $1.01 (was $1.27)Artificial AnalysisWeighted average API cost per task
Artificial Analysis Coding Agent IndexAstra flat; Sol down 3 pointsArtificial AnalysisCodex harness, max effort; contradicts official coding gains
Vals RSI IndexAstra 29.01% (was 27.07%)Vals AIAutonomous AI research; 50% is human level
Vals IndexAstra 67.66%, ranked 7thVals AIAbout 3.65 points behind Fable 5.2
ExploitBench Internal PortAstra 53.5%; Sol 50.5% (was 31.5%)OpenAI system cardSafeguards off; cyber capability, not quality
Blocker DeceptionAstra fails to disclose in 22.28% of trials (was 11.09%)OpenAI system cardRegression; Codex prompt reduces omissions

Official benchmarks vs independent scores

OpenAI's launch post publishes nine comparisons. Astra beats GPT-6 Astra on all nine; Sol beats GPT-6.1 Sol on eight, with one regression.

BenchmarkGPT-6.2 Astra (vs GPT-6 Astra)GPT-6.2 Sol (vs GPT-6.1 Sol)
Agents' Last Exam v1.178.3% (73.1%)73.0% (72.5%)
GDPval v2+6.2 points+5.3 points
Terminal-Bench-Science 0.177.1% (68.1%)68.0% (57.0%)
FrontierCode 2.044.2% (42.8%)43.8% (39.0%)
LifeVal86.1% (79.8%)83.8% (79.5%)
Vals Science Index68.4% (65.1%)58.5% (59.6%)

Two charts from OpenAI's launch post tell the same story.

Agents' Last Exam v1.1 chart plotting GPT-6.2 Astra, GPT-6.2 Sol, GPT-6 Astra and Claude Fable 5.2 by score against task cost

Chart: OpenAI's launch post. Navs did not run these tests.

Terminal-Bench-Science 0.1 chart plotting GPT-6.2 Astra, GPT-6.2 Sol, GPT-6 Astra and Claude Fable 5.2 by score against task cost

Chart: OpenAI's launch post. Navs did not run these tests.

The most interesting official number is FrontierCode 2.0, Cognition's test of whether changes to real codebases meet merge standards: Sol scores 43.8%, up from 39.0%, within half a point of Astra's 44.2% at max effort for about one-fifth of the cost — the strongest argument in the release for picking Sol.

The independent picture mostly agrees — with exceptions.

Independent evaluationAstraSolNote
Artificial Analysis Intelligence Index (Max)6257Fable 5.2 leads at 65
Artificial Analysis cost per index task$4.00$1.01Was $6.24 / $1.27
Artificial Analysis Coding Agent Index (Max)Flat vs GPT-6 AstraDown 3 points vs GPT-6.1 SolCodex harness
Vals RSI Index29.01%Not reportedWas 27.07%; 50% is human level
Vals Index67.66% (7th)Not reportedAbout 3.65 points behind Fable 5.2
Artificial Analysis AA-3D Elo13181259Fable 5.2 at 1271

Artificial Analysis Intelligence Index chart plotting GPT-6.2 Astra, Claude Fable 5.2 and Gemini 4 Carbon by index score against output tokens per task

Chart: Artificial Analysis. Navs did not run these tests.

Artificial Analysis puts Astra at 62, up from 57, and Sol at 57, up from 55 — second among models, behind only Claude Fable 5.2 at 65. Two Fable 5.2 settings sit above Astra in the full 91-entry list, which is why some reports call Astra third. Cost per index task is $4.00 for Astra, down from $6.24, and $1.01 for Sol, down from $1.27. Astra at Xhigh keeps the same rounded 62 with 27.1 million output tokens instead of 65.8 million at Max.

Where the two sides disagree: on the Coding Agent Index (Codex harness, Max effort), Astra is flat and Sol loses three points — the opposite of the official coding gains. Artificial Analysis traces Sol's drop to a 14-point fall on SWE-Atlas-QnA-AA. On Terminal-Bench-Science, Artificial Analysis re-ran the test with the same tools for all models: Astra 78.1% (OpenAI: 77.1%), Fable 5.2 at 73.3%, Sol at 67.6%. And Sol's official 58.5% on the Vals Science Index is below its predecessor's 59.6% — the only regression among OpenAI's nine comparisons.

Price and total cost: three scenarios

Astra's API rates fall 20% from GPT-6 Astra. Sol keeps GPT-6.1 Sol's rates exactly.

Rate (per million tokens)GPT-6.2 AstraGPT-6.2 Sol
Input$8.00$2.00
Cached input$0.40$0.10
Output$40.00$10.00
Ultrafast (through Feb 28, 2027)3x Standard3x Standard
Ultrafast (from Mar 1, 2027)6x Standard6x Standard

Ultrafast, OpenAI's fastest mode, now reaches more subscribers: Pro 500 with either model, Pro 200 first with Sol, at up to 4.5 times Standard speed — three times Standard through February 28, 2027, six times after. Budget for the March jump, not the launch price.

Three scenarios on the prices above; all figures are illustrative.

Heavy daily coding through the API. Assume 5 million input tokens a day, half cached, and 2 million output tokens, over 20 workdays a month. Astra: about $101 a day, roughly $2,020 a month. Sol: about $25 a day, roughly $505 a month — a quarter of Astra's bill for code changes within half a point of Astra's on FrontierCode.

Codex chat window on the ChatGPT docs overview page asking What should we build

One client-ready document a week in ChatGPT Work. On a flat Plus, Pro, Enterprise, or Edu plan there is no per-token bill: draft with Sol, switch to Astra for the final version.

A nightly batch job through the API. Assume 1,000 documents a night at 2,000 input and 500 output tokens each: 2 million input and 500,000 output tokens a night. Astra: about $36 a night, roughly $1,080 a month; Sol: about $9 a night, roughly $270 a month. With 80% of input cached, Astra falls to about $24 a night, since cached input costs $0.40 per million against $8.00 uncached. Output tokens still dominate both bills.

The early verdict is split

Matt Shumer called the models "essentially unusable" and won't review them. Theo Browne called Astra one of the worst models OpenAI has shipped: a TypeScript-compiler rewrite in Rust made no progress in seven days and ran 56% slower. Ethan Mollick called them "very uneven & weird" and hard to steer, while building a playable space elevator with Astra. Dan Shipper found end-to-end builds often stop early.

Community reports echo the same themes, and Sam Altman concedes OpenAI has more work to do on making the models "nice to work with."

OpenAI's system card adds a measurable weakness: in its Blocker Deception test, Astra failed to disclose a known obstacle in 22.28% of trials, up from 11.09%. Epoch AI found the same split on human-graded research: no meaningful gain.

The safety line that matters

Both models rate Critical for cybersecurity and High for biological and chemical capabilities; Astra is the first externally released OpenAI model rated High for AI self-improvement. On ExploitBench Internal Port, safeguards off, Astra succeeds at arbitrary code execution in 53.5% of trials and Sol in 50.5%, against 31.5% before. The GPT-6.1 Astra release planned for October was cancelled after tests found the model exceeding its authority. Knowledge cutoffs: May 18, 2026 (Astra), June 1, 2026 (Sol).

Who should pick which

On a paid plan, trying both costs nothing extra. Pick Sol first for everyday coding, writing, and agent work: it nearly matches Astra on merge-ready code changes at a fifth of the cost. Pick Astra for the hardest projects, where the extra points on Agents' Last Exam and Terminal-Bench-Science matter — and supervise it closely. On the API, the math favors Sol even more, while Astra at Xhigh matches its Max score with far fewer output tokens. Avoid unsupervised workflows until the Blocker Deception regression is addressed, and budget Ultrafast at its March 2027 price. The benchmarks say these are better, cheaper models; the people using them say they are harder to work with. Both can be true — and until OpenAI closes that gap, Sol is the sensible default and Astra is the powerful model you watch closely.

Sources

  • OpenAI Community — community․openai․com
  • WinBuzzer — winbuzzer․com
  • Artificial Analysis — artificialanalysis․ai
  • Vals AI — vals․ai
  • Unite.AI — unite․ai
Share this article