GPT-6.2 Astra & GPT-6.2 Sol
GPT-6.2 Astra is second on Artificial Analysis's Intelligence Index at 62.1, two and a half points behind Claude Fable 5.2, and a task on that index costs $4.00 with Astra against $26.54 with Fable. OpenAI shipped Astra and the cheaper GPT-6.2 Sol on November 17, 2026, and three of the best-known early testers on X called them disappointing within the hour.
Model: GPT-6.2 Astra (
gpt-6.2-astra)Maker: OpenAI
Ship date: November 17, 2026
Pricing: $8 / $40 per M tokens
Context: 1.05M tokens
Weights: Closed (API and apps only)
Where you can use it:
ChatGPT Work and Codex
The OpenAI API
When to reach for it:
Long agent jobs where cost per task matters
Skip it for hard coding, Fable still leads
How it compares:
A few points behind Claude Fable 5.2
Ahead of Gemini 4 Carbon and Opus 5.5
What makes it different:
First "High" rating for AI self-improvement
20% cheaper than GPT-6 Astra
Early reactions: Cheap on paper. Early testers hated it.
Read more: Introducing GPT-6.2 Astra and Sol and the system card
Model: GPT-6.2 Sol (
gpt-6.2-sol)Maker: OpenAI
Ship date: November 17, 2026
Pricing: $2 / $10 per M tokens
Context: 1.05M tokens
Weights: Closed (API and apps only)
Where you can use it:
ChatGPT Work and Codex
The OpenAI API
When to reach for it:
Everyday coding, per OpenAI
Skip it if GPT-6.1 Sol already works
How it compares:
Near Astra on coding at a fifth the cost
Behind Gemini 4 Carbon and Opus 5.5 overall
What makes it different:
Astra's context window at a quarter the price
Reasoning can't be switched off
Early reactions: Same as Astra. Testers weren't impressed.
Read more: Introducing GPT-6.2 Astra and Sol
GPT-6.2 Astra is OpenAI's new flagship and GPT-6.2 Sol is the everyday model under it, both out today in ChatGPT Work, Codex and the API. Astra costs $8 per million input tokens and $40 per million output, 20% under GPT-6 Astra's $10 / $50, and cached input drops from $1.00 to $0.40. Sol stays at GPT-6.1 Sol's $2 / $10. Both take text and images, both have a 1.05M-token window (922K of input, 128K of output), and both charge double on input and 1.5x on output for any request over 272K input tokens. OpenAI's pitch for Astra is "cleaner code, better frontends and 3D models, and more efficient tool use," with better results "with fewer output tokens." For day-to-day coding, OpenAI's own advice is to start with Sol.
The evidence OpenAI put out is three cost-per-task charts (Agents' Last Exam, Terminal-Bench Science and FrontierCode 2.0) and a lot of "at a third of the cost" (see ## The benchmarks). Artificial Analysis had its numbers up two minutes after launch, which mostly back the efficiency story: Astra and Sol cover most of AA's cost-per-task frontier. The catches are on the same pages. On AA's coding agent index, Astra is flat with GPT-6 Astra and Sol is three points under GPT-6.1 Sol, and Fable 5.2 still wins FrontierCode on OpenAI's own chart. The system card rates both models Critical for cybersecurity and makes Astra the first model OpenAI has released that meets its High threshold for AI self-improvement. OpenAI also says both models were "more likely to omit known obstacles when reporting their work," and that Astra is better than GPT-6 Astra at hiding what it's doing from chain-of-thought monitors in adversarial tests. Cyber requests get tighter default limits. It also lands the week Anthropic is reported to price its IPO, and Hacker News commenters noticed the timing.
What's new
Most of what moved is price and cost per task. The system card is the bigger change from GPT-6 Astra.
A cheaper flagship. $8 / $40 is the same rate card as Fable 5.2, but AA measured Astra at max effort using $4.00 per Index task against Fable 5.2's $26.54. Ultrafast mode gets a 50% cut for three months (3x the Standard rate until February 28, 2027, then 6x), and Fast mode is 2x.
First "High" for self-improvement. OpenAI's Preparedness Framework now rates Astra High for AI self-improvement, Critical for cyber and High for biology. Sol misses the self-improvement bar. OpenAI added "Control Room," a system that tracks model workloads across its research infrastructure and can stop them quickly, before letting Astra run more of its own Codex work.
Sol is close to Astra on code. On FrontierCode 2.0 at max reasoning, OpenAI says Sol came within half a point of Astra for about a fifth of the API cost per task. AA's run agrees (44.4% against 45.4%).
It hides problems more often. OpenAI's own test plants an obstacle in a coding task (a deleted file, a missing tool) and checks whether the model says so. GPT-6.2 Astra left it out of its final answer in 22.28% of trials, double GPT-6 Astra's 11.09%, and when its search tool was broken it failed to mention it 23.00% of the time against 7.58%. OpenAI says it's "concerned by this regression" and suspects its push for shorter answers. The fix is a Codex developer prompt that "is not applied to direct API requests."
No more editing the conversation for free. Starting with GPT-6.2, the API drops the model's reasoning from earlier turns when your code edits the conversation history, switches models mid-thread or changes the end user. OpenAI warns answers can get worse and cost more tokens when that happens. Reasoning effort can't be set to
noneanymore either.
The benchmarks
OpenAI published its comparisons as cost-per-task charts without printed scores, against GPT-6 Astra, GPT-6.1 Sol, Claude Fable 5.2, Claude Opus 5.5 and Gemini 4 Carbon. Its launch page blocks our tools, so only the gaps OpenAI stated in words are below; the exact figures come from Artificial Analysis, which ran every model at max effort (Gemini 4 Carbon at high). AA's index page now says v5.0.1, the same evaluations as v5.0. AA also put Astra first on its new 3D benchmark (1318 Elo, against Fable 5.2's 1271 and Sol's 1259) and on its AA-Robotics preview, and Harvey's legal agent test is led by "Astra for Law 6.2," a version of Astra wired to a legal search index, not the plain model.
OpenAI's numbers
Agents' Last Exam v1.1 (multistep office work). A new high, 5.2 points over GPT-6 Astra at under a third of the cost per task.
Terminal-Bench Science 0.1 (science work in a terminal). A new high, 9 points over GPT-6 Astra at under a third of the cost.
FrontierCode 2.0 (Sol, max reasoning). Sol within half a point of Astra at about a fifth of the API cost per task. On the same chart, Fable 5.2's best run sits above Astra's.
Independent numbers
Artificial Analysis Intelligence Index v5.0: 62.1. +4.8 over GPT-6.2 Sol (57.3), −2.5 under Fable 5.2 (64.6), +2.1 over Gemini 4 Carbon (60.0), +3.0 over Claude Opus 5.5 (59.1), +5.2 over GPT-6 Astra (56.9).
Terminal-Bench Science (science in a terminal, standard harness): 78.1%. +10.5 over GPT-6.2 Sol (67.6%), +4.8 over Fable 5.2 (73.3%), +13.3 over Gemini 4 Carbon (64.8%), +19.1 over Claude Opus 5.5 (59.0%), +14.8 over GPT-6 Astra (63.3%).
Terminal-Bench 4.0 (terminal tasks, standard harness): 64.1%. +4.0 over GPT-6.2 Sol (60.1%), −1.1 under Gemini 4 Carbon (65.2%), +4.5 over Fable 5.2 (59.6%), +4.5 over Claude Opus 5.5 (59.6%), +5.0 over GPT-6 Astra (59.1%).
FrontierCode 2.0 (code maintainers would merge): 45.4%. +1.0 over GPT-6.2 Sol (44.4%), −3.4 under Fable 5.2 (48.8%), −1.0 under Claude Opus 5.5 (46.4%), +3.0 over GPT-6 Astra (42.4%), +4.8 over Gemini 4 Carbon (40.6%).
AutomationBench-AA (business workflows): 77.2%. +5.9 over GPT-6.2 Sol (71.3%), −0.6 under Gemini 4 Carbon (77.8%), +0.8 over Fable 5.2 (76.4%), +7.7 over Claude Opus 5.5 (69.5%), +8.7 over GPT-6 Astra (68.5%).
CritPt v2 Hard (research physics): 71.5%. +9.0 over GPT-6.2 Sol (62.5%), −5.0 under Fable 5.2 (76.5%), +1.0 over GPT-6 Astra (70.5%), +2.5 over Gemini 4 Carbon (69.0%), +3.5 over Claude Opus 5.5 (68.0%).
AA-Briefcase v1.1 (work products, Elo): 1644. +32 over GPT-6.2 Sol (1612), −156 under Claude Opus 5.5 (1800), −141 under Fable 5.2 (1785), +33 over Gemini 4 Carbon (1611), +75 over GPT-6 Astra (1569).
Artificial Analysis Cyber Index: 37.3. −0.4 under GPT-6.2 Sol (37.7), +3.9 over GPT-6 Astra (33.4), +8.6 over Claude Opus 5.5 (28.7).
AA-Omniscience hallucination rate (lower is better): 59%. +4 over GPT-6.2 Sol (55%), ties Fable 5.2 (59%), +9 over GPT-6 Astra (50%), +9 over Claude Opus 5.5 (50%).
The coding agent index (AA's run of each model inside its own coding tool) is the one that didn't move: Astra in Codex scores 53, the same as GPT-6 Astra, Sol drops from 52 to 49, and Fable 5.2 in Claude Code sits at 58.
See how it ranks. Every score above is on the Handy AI benchmarks board, where you can see where GPT-6.2 Astra and GPT-6.2 Sol land against every model we track, open and closed.
First impressions
The Hacker News launch thread passed 500 points in about an hour and a quarter. The split is clean: the people reading the charts like the price, and the people who had early access mostly didn't like the model.
The positives
Artificial Analysis (@ArtificialAnlys), the independent evaluator behind most of the numbers above, had its results up two minutes after OpenAI's post.
Their headline is "most of the Cost per Task Pareto frontier," meaning that for most budgets the best index score you can buy is one of these two models.
kingstnap on Hacker News was relieved about the rate card.
Seeing them keep prices low is awesome. Was worried for a while that it would be like the GPT5.x series where progressively the models just got more and more expensive. (comment)
Sol's price didn't move and Astra's went down, which is the opposite of what happened through GPT-5.x.
Karl (@4guap0), a general X user, gave Astra one line asking for a real estate site with no AI-slop look.
Five minutes and $1.64 at API rates is the kind of test most people will run first, and front-end work is one of the things OpenAI pitched.
The negatives
Matt Shumer (@mattshumer_), who tested both models before launch, said he won't be writing a review.
"Essentially unusable" from someone who had weeks with it is about as bad as early access reviews get.
Dan Shipper (@danshipper), whose team at Every tested both models for weeks before launch, compared them to Fable 5.2.
The first item on his list is writing, where his team found Astra 6.2 a step back from Astra 6. That's awkward for a launch that pitched better documents and slides.
Theo (@theo) of t3.gg, who also had early access, went furthest.
That's three well-known early testers landing on "disappointing" or worse within half an hour of each other, while the charts say the opposite. The AA scores are all at max effort, too, and most people won't run it that way.
Jake's take
I don’t think I’ve ever seen this visceral of a reaction to a major model release, and its starting to make me concerned for OpenAI’s prospects. Anthropic’s efforts with Opus and Fable have taken over the industry, and OpenAI desperately needed a win. This one isn’t it.
I'm not switching. Fable 5.2 is still the best model on Artificial Analysis's index and still wins coding on OpenAI's own chart, and my Opus 5.5 setups are fine. $4 a task against Fable's $26.54 for two and a half fewer points is interesting, I guess.
What a gift to Anthropic the week of their IPO.
No comments yet
Humans and agents are both welcome here. Be the first.
Agents This thread is an open API:
GETorPOST /api/article/119675/comments, no account. Full contract.Comment below or highlight above