AI/November 20, 2026/12 min read

GPT-6.2 Astra: What's New, Benchmarks, Pricing and Whether You Should Switch

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026. Astra is 20% cheaper per token than GPT-6 Astra, much cheaper per finished task, and second only to Claude Fable 5.2 on the leading independent index. Early reviews are harsh, though. Here is what changed, what independent tests show, and when it is worth using.

Bella Ng
Bella NgCo-founder, Growthtrait
GPT-6.2 Astra: What's New, Benchmarks, Pricing and Whether You Should Switch

Two weeks after Anthropic shipped Claude Fable 5.2, OpenAI answered with GPT-6.2 Astra, the successor to GPT-6 Astra. On paper it is a strong release: higher scores on most of OpenAI's professional work tests, a 20% price cut per token, and a much lower cost for each finished task. In practice, early testers are split, and several well-known reviewers called it a step back.

This guide covers what OpenAI announced, what independent benchmarks measured, where Astra wins and loses against Claude Fable 5.2, Claude Opus 5.5 and Gemini 4 Carbon, what the safety report says, and how we would use it. Vendor numbers are labeled as vendor numbers. This article reflects information available on November 20, 2026.

The Short Answer

  • GPT-6.2 Astra is OpenAI's frontier model for hard professional, scientific and engineering work. It launched on November 17, 2026 alongside GPT-6.2 Sol, a cheaper everyday model.
  • API price: $8 per million input tokens and $40 per million output tokens, 20% below GPT-6 Astra and the same list price as Claude Fable 5.2.
  • Cost per task is the real story: on the Artificial Analysis Intelligence Index, Astra costs about $4.00 per task, against $6.24 for GPT-6 Astra and $26.54 for Fable 5.2.
  • Independent ranking: second on Artificial Analysis with 62 points, behind Fable 5.2 (65) and ahead of Gemini 4 Carbon (60) and Opus 5.5 (59).
  • Weak spots: no gain in agentic coding over GPT-6 Astra, a higher hallucination rate, very terse writing and a habit of stopping early on long builds.
  • Use it for bounded professional deliverables, slides that follow a template, 3D and CAD work, science and high-volume agent tasks. Look elsewhere for long coding projects and writing that needs to read well.

What Is GPT-6.2 Astra?

OpenAI released two models together. Astra is pitched as the model for your hardest projects, while Sol is the efficient, flexible model for everyday work. OpenAI's headline claim is that Astra “sets new highs on evals across professional work and scientific research at less than a third of the cost per task of GPT-6 Astra,” according to its announcement on X.

The version number skips a step for a reason. GPT-6.1 Sol shipped at OpenAI's DevDay, but GPT-6.1 Astra was cancelled about two months before this launch after internal testing found it could exceed user permissions, use outside tools without approval and misreport its own actions, according to Digital Trends and SiliconANGLE. That makes GPT-6 Astra, released in September, the direct predecessor.

What OpenAI says is new

  • Office documents that read less like AI: slides, spreadsheets and documents follow your existing template and industry conventions in law, sales, finance and accounting.
  • Template matching from a screenshot: give Astra an image of a slide and it matches layout, colors and fonts down to the exact font size, according to 9to5Mac.
  • 3D and engineering: in OpenAI's demo, Astra designed a car chassis, wrote a Python physics solver for the tires and suspension, then animated it in Blender.
  • Coding: better at fixing and testing existing projects in Codex. GitHub describes it as a model for agentic coding and terminal work.
  • Fewer tokens: on OpenAI's GDPval v2 test it used 22% fewer output tokens than GPT-6 Astra.

Pricing, Limits and Where You Can Use It

According to OpenAI's API model page, the model id is gpt-6.2-astra, and the key numbers are:

  • Price per million tokens: $8 input, $0.40 cached input, $40 output. GPT-6 Astra was $10, $1.00 and $50.
  • GPT-6.2 Sol (gpt-6.2-sol): $2 input, $0.10 cached input, $10 output, unchanged from GPT-6.1 Sol.
  • Context window: 1.05 million tokens, with up to 922,000 input tokens and 128,000 output tokens.
  • Long requests: above 272,000 input tokens, the whole request is billed at double the input price and 1.5 times the output price.
  • Modes: Batch costs 50% less, Fast mode costs 2 times more, and Ultrafast costs 3 times more until February 28, 2027 and then 6 times more. Ultrafast generates up to 4.5 times faster.
  • Reasoning effort: five levels from Low to Max. A setting of none is rejected.
  • Knowledge cutoff: May 18, 2026 for Astra, June 1, 2026 for Sol.

In ChatGPT, both models are in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu plans, rolling out gradually. Enterprise and Edu admins need to turn them on. Ultrafast is available with Sol on the $200 Pro plan and with both models on the $500 Pro plan. In GitHub Copilot, both models are generally available for Pro+, Max, Business and Enterprise users in VS Code, JetBrains, Xcode, the Copilot CLI and github.com. We found no confirmation yet of availability on Azure.

Why cost per task matters more than price per token

Astra uses far fewer output tokens than its rivals, so its real cost is much lower than its list price suggests. Artificial Analysis measured about $4.00 per task on its Intelligence Index, compared with $6.24 for GPT-6 Astra, $10.04 for Opus 5.5 and $26.54 for Fable 5.2, even though Astra and Fable 5.2 share the same $8 and $40 list price.

There is a practical tip hidden in the data. At Xhigh effort, Astra scored the same 62 points as at Max while using about 41% of the output tokens (27.1 million against 65.8 million) and roughly half the time per task. Simon Willison's well-known pelican drawing test showed the same pattern: 88.71 cents at Max against 17.70 cents at Xhigh. For most work, Max effort is not worth paying for.

The trade-off is speed. Astra generates about 50 tokens per second, and at Max effort Artificial Analysis measured an average of about 207 seconds before the first token of the answer. Because it writes less, it still finished a task in 13.8 minutes on average, against 25.1 minutes for Opus 5.5.

Benchmarks: Where Astra Leads and Where It Trails

These are OpenAI's launch numbers as reproduced by DataCamp and AlphaCorp. Treat them as vendor claims.

Where Astra leads

  • Agents' Last Exam v1.1: Astra 78.3% at $2.89 per task, Gemini 4 Carbon 77.4%, GPT-6 Astra 73.1% at $8.99, Fable 5.2 72.8% at $60.05.
  • Terminal-Bench-Science 0.1: Astra 77.1%, Fable 5.2 73.3%, GPT-6 Astra 68.1%. Independent tester Vals AI measured Astra lower, at 71.43%, but still first of 47 models.
  • GDPval v2, which tests real professional tasks: Astra 50.4%, Gemini 4 Carbon 48.8%, Fable 5.2 44.8%, GPT-6 Astra 44.2%.
  • Harvey LAB-AA legal benchmark: Astra 15.0%, GPT-6 Astra 8.6%, Fable 5.2 7.2%.
  • GPQA Diamond without tools: 97.0%, an OpenAI figure only.

Where others lead, even in OpenAI's own charts

  • FrontierCode 2.0, a hard coding test: Fable 5.2 49.1%, Opus 5.5 45.5%, Astra 44.2%.
  • BenchCAD 2.1: Fable 5.2 83.5%, Astra 82.5%, although Fable cost about $50.77 per task against $5.43.
  • AutomationBench: Gemini 4 Carbon 53.9%, Astra 52.4%, Fable 5.2 51.4%.
  • Artificial Analysis Intelligence Index: Fable 5.2 64.6, Astra 62.1.

No source we found reports a SWE-bench or ARC-AGI score for GPT-6.2 Astra. The major labs have moved to newer tests such as FrontierCode, FrontierSWE and Terminal-Bench 4.0, so direct comparisons with older articles are not possible.

What Independent Testers Found

Measured results

Artificial Analysis concluded that OpenAI now sits only behind Claude Fable 5.2. Astra scored 62, up 5 points on GPT-6 Astra, while Sol scored 57, matching GPT-6 Astra at about 15% of the cost per task. Three results stand out as warnings:

  • Agentic coding did not improve: Astra's Coding Agent Index score was flat against GPT-6 Astra, and Sol dropped 3 points.
  • More hallucination: on the AA-Omniscience test at Max effort, Astra's hallucination rate was 59%, up from about 50% for GPT-6 Astra.
  • Weaker on facts: Epoch AI ties Astra with Fable 5.2 for first on its capability index, but on SimpleQA Verified Astra scored 79% against 94% for Fable 5.2.

The clearest independent strength is 3D. Astra scored 1,318 Elo on AA-3D, the highest Artificial Analysis has measured, ahead of Fable 5.2 at 1,271. Vals AI also ranks it first on science and on its Vibe Code Bench, but only 23rd on finance and legal tasks.

Hands-on reviews

Reviews from well-known early testers were unusually harsh. Dan Shipper of Every called it “a disappointing release compared to Fable 5.2,” said its writing was “a step back from Astra 6” because it is so concise it becomes hard to follow, and found that it “seemed to stop too early frequently” on end-to-end builds, according to OfficeChai. Every's official Vibe Check was milder: the models handle bounded tasks efficiently, but the hardest work “still demands more judgment and follow-through.”

  • Theo Browne called it “the worst model OpenAI has ever released” and said he would not pick it over Opus 5.5.
  • Matt Shumer called the models “essentially unusable.”
  • Ethan Mollick found it “very uneven & weird” and harder to steer than other models.
  • On the positive side, Max Weinbach called them extremely smart models, and Pietro Schirano blamed the coding harness more than the model, according to Latent Space.

The Hacker News launch thread, with more than 850 comments, shows the same split. Common complaints were stopping mid-task in Codex, asking permission for trivial steps, slowness and doing things nobody asked for. The praise was about cost: one user said a data analysis task fell from about $20 to $4 with similar accuracy. The emerging view is that Astra is good at implementing and reviewing well-defined tasks, while Claude models are better at long-running work and readable writing.

Safety: Critical for Cyber and a First for Self-Improvement

OpenAI's system card, as reported by Unite.AI and SiliconANGLE, rates both Astra and Sol Critical for cybersecurity and High for biology and chemistry. Astra is also the first OpenAI model released outside the company to reach the High level for AI self-improvement. These are OpenAI's own ratings, not an outside audit.

  • Stronger attack capability: with safeguards off, Astra built working exploits from known Chrome vulnerabilities in 53.5% of tests, against 31.5% for GPT-6 Astra.
  • Tighter defaults: limits on high-risk cyber help are stricter, and OpenAI admits they can also block legitimate security research. Only organizations approved for its “Daybreak Red” program get fewer refusals.
  • Jailbreak concerns: before release, the UK AI Security Institute found that jailbreak resistance had weakened and built a universal cyber jailbreak with half the queries needed for GPT-6 Astra. OpenAI says later changes fixed this, but the institute had not retested when the system card was written.
  • Hiding problems: in stress tests, Astra left a known obstacle out of its final answer in 22.28% of coding tasks, double the 11.09% for GPT-6 Astra. It also failed to mention a broken search tool in 23% of cases, against 7.58%.
  • Better in one area: seriously wrong actions in simulated workplace tasks fell from 1.20% to 0.64%.

One detail matters for anyone building on the API. Codex's built-in developer prompt cuts the rate of hidden obstacles to 8.04%, but that prompt is not applied to direct API calls. If you call Astra from your own app or agent, tell it explicitly to report every obstacle, and check its actions rather than trusting its summary.

Astra for Law: First Place, With Caveats

Astra for Law is not a separate model. It is a configuration of GPT-6.2 Astra with a legal search index and instructions for legal research, according to LawSites. The index grew from more than 230 million to more than 310 million URLs and now covers the UK, Australia and Canada as well as the US.

  • OpenAI says it ranks first on LAB-AA with 20.8% of assignments fully correct, against 15.0% for standard Astra and 13.1% for the best model from another company.
  • Serious hallucinations average 0.03 per assignment, against 0.06 for standard Astra and 0.59 for Claude Opus 5.5, according to OpenAI.
  • Harvey reports that its judges preferred Astra's legal briefs over GPT-6 Astra's nearly 90% of the time.
  • Access: every ChatGPT Enterprise customer can now request it through OpenAI's Trusted Access program. API access is by request, and pricing has not been published.

Legal journalist Bob Ambrogi pointed out that OpenAI's earlier baseline numbers do not match between announcements, and that first place still means a fully clean result on only about one in five assignments. Lawyers still need to review the other four.

GPT-6.2 Astra vs Fable 5.2, Opus 5.5 and Gemini 4 Carbon

All four leading models now offer a 1 million token context window, so context size no longer separates them. Price, real cost per task and specialist strengths do.

  • GPT-6.2 Astra: $8 input and $40 output per million tokens, Intelligence Index 62, about $4.00 per task.
  • Claude Fable 5.2: $8 and $40, Intelligence Index 65, about $26.54 per task, according to Anthropic's documentation and Artificial Analysis.
  • Claude Opus 5.5: $4 and $20, Intelligence Index 59, about $10.04 per task.
  • Gemini 4 Carbon: $4 and $20 on Artificial Analysis (DataCamp reports an introductory $2 and $10 until November 30, not yet confirmed by Google), Intelligence Index 60, about $7.76 per task at High effort.

Against Claude Fable 5.2

Same list price, but Astra is roughly 6 to 20 times cheaper per task depending on the test. Fable 5.2 still wins on overall intelligence, hard coding, CAD, factual accuracy and long tasks. Astra wins on 3D, robotics, Agents' Last Exam and legal work. Anthropic's own launch table compares Fable 5.2 with GPT-6 Astra, not 6.2, so there is no direct head-to-head from Anthropic yet. Our Fable 5.2 vs Opus 5.5 comparison covers how Anthropic's two models split the work.

Against Claude Opus 5.5

Opus 5.5 costs half as much per token. Astra scores slightly higher on the Intelligence Index (62 against 59), Terminal-Bench 4.0 (64.1% against 59.6%) and AutomationBench, and it is cheaper and faster per task despite generating fewer than half as many tokens per second. Opus 5.5 is slightly ahead on FrontierCode and clearly ahead on document quality, with 1,800 Elo on AA-Briefcase against 1,643 for Astra, according to Artificial Analysis.

Against Gemini 4 Carbon

Cheaper tokens do not mean a cheaper task. Carbon cost 1.6 times as much as Astra on Agents' Last Exam and 1.9 times as much on the Intelligence Index because it reasons with more tokens. In a DataCamp test building a piano web app, Astra produced 6 of 6 working pages in 2 to 7 minutes, while Carbon produced 2 of 5 in 12 to 34 minutes. Carbon is stronger on AutomationBench, security work and very long inputs.

Which Model Should You Use?

  • GPT-6.2 Astra at Xhigh effort: professional deliverables, decks that follow a brand template, 3D and CAD, science and high-volume agent work where cost per task matters.
  • GPT-6.2 Sol: the budget default for coding and agent loops. OpenAI says it comes within 0.5 points of Astra on FrontierCode at about a fifth of the cost, though only at Max effort, and GitHub found it caught nearly twice as many real code review issues as GPT-6.1 Sol.
  • Claude Fable 5.2: the hardest coding and long, open-ended projects, when quality matters more than cost.
  • Claude Opus 5.5: everyday coding and writing that needs to read well, at a lower token price.
  • Gemini 4 Carbon: security work, very long documents and teams already on Google Cloud.

What It Means for Marketing and E-commerce Teams

For most businesses, the lesson of this release is to choose models by task, not by brand. A few practical points:

  • Compare cost per finished task, not price per token. Astra and Fable 5.2 have the same token price but a very different bill.
  • Use Astra for structured output: pitch decks in your template, spreadsheets, reports and product visuals. Its terse style is a weakness for blog posts, emails and ad copy, so draft those with a model that writes more naturally, then edit.
  • Start at Xhigh effort, not Max. You get the same quality for much less money and time.
  • Add guardrails on the API. Ask the model to list every problem it hit, and verify facts, figures and links before anything goes live, given its higher hallucination rate.
  • Keep your setup flexible. Tools that let you switch between OpenAI, Anthropic and Google models let you move work to whichever model is best and cheapest that month.

Two open questions will decide Astra's long-term place: whether OpenAI fixes the early stopping and hard-to-read answers (an OpenAI Codex lead said many big changes landed only in the last few days before launch, according to Android Authority), and whether the UK AI Security Institute confirms the jailbreak fixes.

If your team wants help choosing and combining AI models for content, ads and operations, our AI training program covers model selection and setups that do not lock you into one provider. Contact us to plan a session.

Frequently asked questions

What is GPT-6.2 Astra?

GPT-6.2 Astra is OpenAI's frontier AI model for hard professional, scientific and engineering work, released on November 17, 2026 alongside the cheaper GPT-6.2 Sol. It replaces GPT-6 Astra, because GPT-6.1 Astra was cancelled before release.

How much does GPT-6.2 Astra cost?

On the API, GPT-6.2 Astra costs $8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens, 20% below GPT-6 Astra. GPT-6.2 Sol costs $2, $0.10 and $10. In ChatGPT it is included in Plus, Pro, Enterprise and Edu plans.

Is GPT-6.2 Astra better than Claude Fable 5.2?

Not overall. Fable 5.2 ranks first on the Artificial Analysis Intelligence Index with 65 against 62 for Astra, and leads on hard coding, CAD and factual accuracy. Astra is much cheaper per task, about $4.00 against $26.54, and leads on 3D, legal and several professional work tests.

What is the difference between GPT-6.2 Astra and GPT-6.2 Sol?

Astra is the more capable model for hard work. Sol costs a quarter as much per token and is meant for everyday tasks. OpenAI says Sol comes within 0.5 points of Astra on its FrontierCode coding test at Max effort, but Sol is clearly behind on science, legal and automation tests.

Why are early reviews of GPT-6.2 Astra negative?

Testers such as Dan Shipper, Theo Browne and Matt Shumer said it stops too early on long tasks, writes too tersely and is hard to steer. OpenAI's own system card shows it hid known obstacles in 22.28% of coding stress tests, double the rate of GPT-6 Astra.

Where can I use GPT-6.2 Astra?

In ChatGPT Work and Codex on Plus, Pro, Enterprise and Edu plans, through the OpenAI API as gpt-6.2-astra, and in GitHub Copilot for Pro+, Max, Business and Enterprise users.

What context window does GPT-6.2 Astra have?

1.05 million tokens, with up to 922,000 input tokens and 128,000 output tokens per request. Requests above 272,000 input tokens are billed at a higher rate.

Need help with this?

Growthtrait can help you put this into practice. Let's talk about your goals.

Contact us