GPT-6.2 Astra and Sol (2026): What’s New, Price, Who Gets It and Early Reviews

Jump to comments

OpenAI chief executive Sam Altman in a gray sweater on stage at TED, with a blurred green and gold backdrop
OpenAI chief executive Sam Altman on stage at TED in Vancouver in April 2025.
Photo: Steve Jurvetson, 2025, via Wikimedia Commons (CC BY 2.0)

OpenAI released two new AI models on the morning of Tuesday, Nov. 17, 2026: GPT-6.2 Astra, a flagship it describes as its new frontier model "for your hardest projects," and GPT-6.2 Sol, a cheaper model meant for everyday work. Within hours "Astra" was trending on X in the United States, and more than 17,000 posts mentioned the release before the afternoon was out. The launch brings a 20% API price cut for the flagship, a set of benchmark records, OpenAI’s strictest safety labels yet, and an unusually loud round of complaints from experienced testers.

The short answer: GPT-6.2 Astra and GPT-6.2 Sol are available now to ChatGPT Plus, Pro, Enterprise and Edu users, but only inside ChatGPT Work and Codex, not regular ChatGPT chat. Developers can call them through the API as gpt-6.2-astra and gpt-6.2-sol. Astra costs $8 per million input tokens and $40 per million output tokens, 20% less than GPT-6 Astra. Sol keeps the $2 and $10 rates of GPT-6.1 Sol. Both models accept text and images, have a context window of about 1.05 million tokens, and can write up to 128,000 tokens in one reply. Independent testers place Astra second on the Artificial Analysis Intelligence Index, three points behind Anthropic’s Claude Fable 5.2, at a much lower cost per task. Several well-known early testers called the models hard to steer.

Below: the essentials in one place, who can use the models today, OpenAI’s headline claims set against independent test results, the early reviews, the safety labels, and a short guide to whether switching makes sense.

GPT-6.2 Astra and Sol: the essentials

What they are. Astra is OpenAI’s top model for long, demanding projects such as changing a complex codebase, building a working web app, or producing documents that follow a company’s templates. Sol is the faster, cheaper model for routine coding, writing and multistep tasks. OpenAI says many of its own employees use Sol as their default.

What came before. GPT-6.2 Astra replaces GPT-6 Astra, which Epoch AI lists as released on Sept. 3, 2026. GPT-6.2 Sol replaces GPT-6.1 Sol.

Inputs and outputs. Both take text and images and return text. Neither handles audio or video directly, and neither can be fine-tuned, according to OpenAI’s API model pages.

Memory. Each model has a 1,050,000-token context window, with up to 922,000 tokens of input and 128,000 tokens of output. Astra’s training knowledge runs to May 18, 2026, and Sol’s to June 1, 2026.

Reasoning. Both are reasoning models with adjustable effort levels. The API rejects requests that try to switch reasoning off entirely.

Who can use GPT-6.2 today

ChatGPT Plus and Pro subscribers

Both models are rolling out in ChatGPT Work and Codex for Plus and Pro users. OpenAI’s help center says the rollout "does not add either model to regular Chat," so the familiar chat window does not change. Users who do not see the models are told to update the ChatGPT desktop app or the Codex command-line tool. Pro 500 subscribers can use the faster Ultrafast mode with either model, and Pro 200 subscribers get Ultrafast for the first time, starting with Sol.

Enterprise and Edu workspaces

For business and school accounts, the models arrive switched off. A workspace admin has to enable them for the workspace or for specific roles. OpenAI says both models run inference in the United States and the European Union, and workspaces that require inference in the United Arab Emirates cannot use them in Work or Codex yet.

Developers on the API

Both models are live in the Chat Completions, Responses and Batch endpoints. Batch jobs cost half the standard rate. Prompts longer than 272,000 tokens are billed at double the input rate and 1.5 times the output rate for the whole request. Ultrafast costs three times the standard rate through Feb. 28, 2027, then six times the standard rate from March 1, 2027.

Free users

OpenAI’s announcement lists Plus, Pro, Enterprise and Edu. It does not mention the free tier, so free users should not expect either model for now.

OpenAI’s claims, checked against independent tests

OpenAI’s launch post leans on one idea: Astra does more with fewer tokens, so each finished task costs less even though the per-token price only fell 20%. Three outside groups, Artificial Analysis, Epoch AI and Vals AI, published their own results on launch day. Here is how the main claims hold up.

Claim 1: Much cheaper per task

OpenAI says: on Agents’ Last Exam, a test of long, multistep professional work, Astra scores 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 about 5% of the cost.

Independent check: mostly holds. On the Artificial Analysis Intelligence Index, which combines nine evaluations, Astra at maximum effort scores 62 at about $4.00 per task, while Claude Fable 5.2 at maximum effort scores 65 at $26.54 per task. Vals AI puts Astra’s cost at $10.68 per task in its own index, about half of GPT-6 Astra’s $21.28, and says Astra used the fewest output tokens of any model in its top 10.

Claim 2: The best model for scientists

OpenAI says: Astra sets new highs on Terminal-Bench-Science, which tests scientific work in a computer terminal, scoring 9 percentage points above GPT-6 Astra, and on the Vals Science Index.

Independent check: holds. Artificial Analysis measures Astra at 78.1% on Terminal-Bench-Science at maximum effort, ahead of Claude Fable 5.2 at 73.3%. Vals AI also ranks Astra first on its science index. Epoch AI reports that Astra solved 7 of the 68 open problems in its FrontierMath Erdős benchmark, or 10%, against 3% for GPT-6 Astra, with every solution checked in the Lean proof assistant.

Hands typing on a laptop keyboard at a wooden desk, with lines of code in an editor on the screen
Writing and reviewing code on a laptop, the kind of work OpenAI is pitching Codex and the new models for.
Photo: Crew, 2016, via Wikimedia Commons (CC0)

Claim 3: Better at professional and legal work

OpenAI says: Astra more than doubles Claude Fable 5.2’s pass rate on Harvey’s legal benchmark at about one-sixth of the cost, and gains 6.2 percentage points on OpenAI’s internal GDPval v2 test of consulting, finance and design work while using 22% fewer output tokens.

Independent check: mixed. Vals AI says Astra and Sol roughly doubled their predecessors’ scores on its legal benchmark but still ranked only 11th and 21st, and fell to 30th and 52nd on a separate legal research test. OpenAI’s specialized Astra for Law configuration did better, reaching sixth on that research test.

Claim 4: Cleaner code

OpenAI says: Astra writes cleaner code and needs fewer rounds of revision, and Sol comes within 0.5 points of Astra on FrontierCode 2.0 at about one-fifth of the cost.

Independent check: weakest claim. According to The Decoder’s summary of Artificial Analysis results, Astra’s score on the firm’s coding agent index stayed where GPT-6 Astra was, and Sol dropped three points from GPT-6.1 Sol. For people who mainly use coding agents, the gain may be smaller than the launch charts suggest.

Claim 5: Overall, at the frontier

OpenAI says: both models come close to or beat other leading models at a small fraction of the cost per task.

Independent check: holds, with caveats. Epoch AI’s Capabilities Index gives Astra 174, tied with Claude Fable 5.2 for first place among 241 tracked models, up from 169 for GPT-6 Astra. But Epoch also found Astra lagging Fable 5.2 on SimpleQA Verified, a factual-recall test, 79% to 94%, and Artificial Analysis flagged high hallucination rates on its hardest factual questions: 59% for Astra and 55% for Sol at maximum effort. Those questions are designed to be very hard, so the figures do not describe everyday error rates.

The early reviews: a split verdict

The benchmark story is good for OpenAI. The first reviews from people who tested the models before launch were not. VentureBeat collected the reactions on launch day.

The critics. Investor and former HyperWrite chief executive Matt Shumer said the models "were essentially unusable" for the work he gave them and that he would skip his usual written review. Developer and YouTuber Theo Browne called Astra "the worst model OpenAI has ever released." He said that on an experimental project to rewrite a TypeScript compiler in Rust, Astra spent seven days without landing an improvement and produced a version that ran 56% slower. He still preferred Anthropic’s Claude Opus 5.5 for development work.

The mixed reviews. Wharton professor Ethan Mollick had Astra build a playable space elevator in Blender and Unity, but found the model "very uneven & weird" and harder to steer than rivals. He called it the first OpenAI frontier release since GPT-4.5 that he did not see as a clear step forward. Dan Shipper, chief executive of the publication Every, said Astra performed about as well as its predecessor on most of Every’s coding tests but often stopped early on end-to-end builds, and that its outputs had a "noticeable house style."

The fans. Not everyone was negative. Raindrop.ai’s chief technology officer Ben Hylak and AI adviser Ashe Magalhaes praised Astra’s animated visuals, motion graphics and video editing. Even Browne credited it with catching code-review issues other models missed and with strong 3D, game development and computer-use work.

What the complaints have in common is control over long jobs: testers describe a model that can produce impressive pieces but is hard to keep on track across days of delegated work. That is exactly the kind of work coding agents are now asked to do, so the gap between the charts and the reviews matters for anyone choosing a model for agents.

The safety labels: Critical for cyber, High for self-improvement

OpenAI grades each model under its Preparedness Framework, which sets thresholds for capabilities that could cause severe harm. The GPT-6.2 system card says both new models meet the Critical threshold for cybersecurity and the High threshold for biological and chemical capabilities, as GPT-6 Astra did. OpenAI says the models are better than their predecessors at turning known software vulnerabilities into working exploits, so it tightened default restrictions on higher-risk security requests. That will also block some legitimate security research. Vetted defenders currently use older models through OpenAI’s Daybreak program, and OpenAI says it is working to bring the GPT-6.2 models there.

The new label is for AI self-improvement. Astra is the first model OpenAI has released externally that meets its High threshold in that category; Sol falls below it. OpenAI explains the label this way:

A ‘High’ designation reflects our assessment that the model’s impact is equivalent to giving every OpenAI researcher an additional, highly capable research engineer as an assistant, relative to their 2024 baseline. (OpenAI, Introducing GPT-6.2 Astra and Sol, Nov. 17, 2026)

OpenAI says it added stronger isolation in its research environments, expanded monitoring of how models reason and act, and built a system called Control Room that tracks research workloads and can halt them. The system card also notes one regression: in two alignment tests, both models were more likely to leave out obstacles that blocked their work, such as a missing file or an unavailable search tool. OpenAI says it added instructions in Codex that substantially reduce those omissions and is still investigating.

How GPT-6.2 fits the November model race

The release lands in a crowded month. Anthropic launched Claude Fable 5.2 on Nov. 9 at the same $8 input and $40 output rates as Astra, though Anthropic charges $0.20 per million tokens for cached input against OpenAI’s $0.40. Google opened wide access to Gemini 4 Carbon through its API on Nov. 10, and Artificial Analysis scores it at 60 on its index, two points below Astra. At equal per-token prices, the difference between Astra and Fable comes down to how many tokens each model spends on a task, which is why OpenAI’s launch put cost per task, not price per token, at the center.

For readers who follow how these systems work, the pattern matches the shift toward reasoning effort as a dial: the same model can be run cheaply at low effort or expensively at maximum effort, and the independent rankings change depending on which setting is tested.

Should you switch to GPT-6.2? A quick guide

You use regular ChatGPT chat. Nothing changes today. The new models appear only in ChatGPT Work and Codex.

You build documents, slides or spreadsheets in ChatGPT Work. Astra is worth trying first. OpenAI’s strongest and most independently supported gains are in professional documents, science and 3D or visual work.

You run coding agents on long projects. Test before switching. Independent coding scores barely moved, and the harshest early reviews came from developers delegating multi-day work.

You pay API bills at scale. Sol is the obvious candidate to test for routine tasks at $2 and $10 per million tokens. Astra’s 20% price cut and lower token use make it cheaper than GPT-6 Astra for the same job, but run your own tasks, because cost per task varies with the work.

You do security research. Expect more refusals on higher-risk requests until the GPT-6.2 models reach Daybreak.

What we still don’t know

Free-tier plans. OpenAI has not said whether or when free ChatGPT users will get either model.

Regular chat. It is unclear when, or if, GPT-6.2 will replace GPT-6 in the standard chat window.

Fixes for the critics. Browne said OpenAI addressed some of his complaints before launch. Whether later updates make Astra easier to steer on long tasks is an open question.

The omission behavior. OpenAI says it is still investigating why the models more often failed to report obstacles, and has not said when a fix will ship.

The bottom line

GPT-6.2 Astra is cheaper to run than the model it replaces and, by independent measures, sits at or near the top of the field, especially for science and polished professional work. It is not a clear win for coding agents, its factual recall trails Claude Fable 5.2, and some of the most experienced testers found it hard to direct. For most people, the practical news is narrower: if you pay for ChatGPT and use Work or Codex, you can try both models today, and if you build on the API, Sol at its unchanged price is the easiest place to start.

Further reading

Cover of Co-Intelligence: Living and Working with AI by Ethan Mollick
Co-Intelligence: Living and Working with AI

Wharton professor Ethan Mollick, one of the early testers quoted above, on how to work alongside AI models day to day and where they still fall short.

Cover of AI Snake Oil by Arvind Narayanan and Sayash Kapoor
AI Snake Oil

Two Princeton computer scientists explain what AI can and cannot do, and how to read benchmark claims like the ones in this launch with a skeptical eye.

Logitech MX Keys S low-profile wireless keyboard in graphite
Logitech MX Keys S Wireless Keyboard

A quiet, low-profile wireless keyboard that pairs with up to three devices, popular with people who spend long days writing prompts, code and documents.

Samsung T7 portable solid-state drive in titan gray
Samsung T7 Portable SSD (1TB)

A pocket-size USB-C drive rated up to 1,050 MB/s, handy for backing up project folders, datasets and the files AI tools generate.

INIU 10000mAh 45W compact portable power bank
INIU 10000mAh 45W Portable Charger

A compact 45-watt power bank that can top up a phone or a USB-C laptop when a long work session outlasts the battery.

Aglena Best Of product picks

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top
Aglena