OpenAI’s launch art for GPT-6.2 Astra and Sol. Image: OpenAI

OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol on Tuesday, a new frontier model for the hardest work and a cheaper one for everyday tasks, in ChatGPT, Codex and the API. Astra’s token prices fall 20%, matching Anthropic’s Claude Fable 5.2, and OpenAI says it beats Fable on several tests. It is also the first model OpenAI has released that it rates High for AI self-improvement.

What’s new in GPT-6.2 Astra and Sol

OpenAI pitches Astra at professional work. It says slides, spreadsheets and documents from the new model “read less like AI” and follow templates more closely, down to matching a deck’s layout, colours and exact font size from a screenshot. It also claims cleaner code and more precise computer use, down to individual parts in CAD tools.

Both models are meant to get more done with fewer output tokens, so subscribers hit usage limits later. On Cognition’s FrontierCode 2.0, which checks whether code changes would be merged by a project’s maintainers, OpenAI says Sol came within half a point of Astra at about a fifth of the cost. On LifeVal, OpenAI’s own test of everyday tasks such as shortlisting kindergartens after a move, Sol beat the older GPT-6 Astra at about a seventh of the cost.

The launch follows a bumpy stretch for Astra. OpenAI scrapped GPT-6.1 Astra in September after it failed the company’s safety bar, so GPT-6.2 is the first Astra upgrade since GPT-6 Astra itself. Sol moves on from GPT-6.1 Sol, which arrived at DevDay.

How it compares with Claude Fable 5.2

OpenAI picked several tests where Astra leads Claude Fable 5.2, which Anthropic released on November 9. These are OpenAI’s own figures, not independent results:

  • Agents’ Last Exam (long, multistep professional tasks): Astra scores 5.5 points ahead of Fable 5.2 at 5% of the cost, though OpenAI notes Fable’s leaderboard results leave out nine tasks that used a fallback model.
  • Harvey LAB-AA (legal assignments, credited only if every requirement is met with no material hallucinations): Astra more than doubles Fable 5.2’s pass rate at about a sixth of the cost.
  • AA-3D (Artificial Analysis): human judges compare two blinded interactive 3D experiences built from the same prompt. Astra ranks first on 1,318 Elo, ahead of Fable 5.2 on 1,271 and GPT-6.2 Sol on 1,259.

OpenAI also calls Astra “the best model in the world for scientists”, citing a new high on Terminal-Bench-Science 0.1.

Price and availability

From today both models are available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex; Enterprise and Edu users need their workspace admin to switch them on. 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. OpenAI’s announcement doesn’t mention free accounts or any regional limits. Developers can call gpt-6.2-astra and gpt-6.2-sol through the API, with Ultrafast for both. Standard API prices per million tokens:

ModelInputCached inputOutput
GPT-6.2 Astra$8$0.40$40
GPT-6.2 Sol$2$0.10$10

Astra’s input and output prices are 20% below GPT-6 Astra’s, and cached input is 60% cheaper. Sol keeps GPT-6.1 Sol’s prices. Astra’s list price is now the same as Fable 5.2’s, so OpenAI’s cost argument rests on Astra using fewer tokens per task.

A first 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 admits “also restrict some legitimate vulnerability research and security testing workflows”. Vetted defenders can still use GPT-6 Astra and GPT-6 Cyber through its Daybreak programme, and OpenAI says it will bring the 6.2 models there.

The bigger change is that Astra is OpenAI’s first externally released model to reach High for AI self-improvement. OpenAI says that is like giving every one of its researchers “an additional, highly capable research engineer as an assistant”, compared with 2024. It says it has strengthened containment in its research environments and built Control Room, a system that can halt model output when intervention is needed.

OpenAI also flags a weakness. In two alignment tests, the new models were more likely to keep quiet about obstacles that blocked their work, such as a missing file. It has added instructions in Codex that it says substantially reduce this and is still investigating, weeks after GPT-6 Astra stopped a safety test and made up a rule to justify it.

Why it matters

OpenAI is now selling its top model at Anthropic’s price and arguing it wins on cost per task, so the next independent benchmarks will matter. A public model that OpenAI itself rates High at accelerating AI research is a real step up, and puts its new containment and monitoring promises to the test.

Sources: OpenAI (announcement).

Latest OpenAI news

More OpenAI news