What was released and at what prices
On 17 November 2026 OpenAI released two language models. GPT-6.2 Astra targets demanding professional work and GPT-6.2 Sol targets everyday use, as described by The Decoder and MarkTechPost. In the API they are available as gpt-6.2-astra and gpt-6.2-sol. Simon Willison reports that he released version 0.38 of the LLM tool with support for both models.
For Astra, the company lists $8 per million input tokens, $0.40 for cached input and $40 for output. Willison recalls that GPT-6 Astra cost $10 and $50, with $1 for cached input. He also notes that the new price is the same as Claude Fable 5.2, which also became 20% cheaper than its predecessor. MarkTechPost describes the change as a 20% reduction.
Sol stays at $2 for input, $0.10 cached and $10 for output, like GPT-6.1 Sol. According to Willison, the same was true of GPT-6 Sol, while GPT-5.6 Sol cost $4 and $20.
In ChatGPT Work and in Codex the models are available to Plus, Pro, Enterprise and Edu users, according to The Decoder. Pro-500 subscribers get Ultrafast mode with both. Pro-200 subscribers get access to Ultrafast for the first time, starting with Sol.
Results: what OpenAI states and what Artificial Analysis measures
The figures in the first group below are OpenAI's own results for its own models, as relayed by The Decoder. On Agents' Last Exam, a benchmark for multi-step professional workflows, Astra beats GPT-6 Astra by 5.2 points at less than one third of the cost. It leads Claude Fable 5.2 by 5.5 points at 5% of the cost. On the legal benchmark Harvey LAB-AA, OpenAI says the success rate is more than double that of Claude Fable 5.2, at about one sixth of the cost. On the internal GDPval-v2 the gain is 6.2 percentage points with 22% fewer output tokens. On Terminal-Bench-Science the difference from GPT-6 Astra is 9 percentage points.
On AA-3D, where human judges compare two blind interactive experiences, Astra has 1,318 Elo, Claude Fable 5.2 has 1,271 and Sol 1,259. On FrontierCode 2.0, which tests whether code changes meet the standard of real-world projects, Sol trails Astra by 0.5 points at one fifth of the cost.
On cybersecurity, OpenAI states that both models reach the “Critical” threshold of the Preparedness Framework, as GPT-6 Astra did. It adds stricter safeguards, which will also restrict legitimate vulnerability research. On the internal ExploitBench, Astra exceeds a 50% success rate at creating exploits for known browser vulnerabilities. The measurements were run with the safeguards disabled.
Artificial Analysis provides an independent picture. Its Intelligence Index combines nine evaluations across knowledge, coding and science. Astra rises five points to 62, second behind Claude Fable 5.2 (65). Sol gains two points and reaches 57, the level of GPT-6 Astra at about 15% of the cost per task. In robotics (AA-Robotics v0.1) Astra scores 79, Sol 78 and Claude Fable 5.2 75.
Artificial Analysis Intelligence Index, scores as given by Artificial Analysis
| Model | Intelligence Index points |
|---|---|
| Claude Fable 5.2 | 65 |
| GPT-6.2 Astra | 62 |
| Gemini 4 Carbon | 60 |
| Claude Opus 5.5 | 59 |
| GPT-6.2 Sol | 57 |
Limits of the measurements and reception
Artificial Analysis sees no improvement in coding. On the Coding Agent Index, with the Codex harness and maximum compute effort, Astra stays where GPT-6 Astra was. Sol falls three points from GPT-6.1 Sol, due to a 14-point drop on SWE-Atlas-QnA-AA, where it more often omits key elements in its answers.
The AA-Omniscience Hallucination Rate measures how often a model guesses instead of saying it does not know. At maximum effort Astra reaches 59% and Sol 55%. The questions are unusually difficult, so the rates do not reflect everyday use. The Decoder also relays that Epoch AI tested Astra before release and measured 174 points on the Epoch Capabilities Index.
Simon Willison describes the reception as mixed. Astra does not score as high as Claude Fable 5.2 on several benchmarks and there are complaints about a confusing communication style. Theo Browne, in the note he sent to OpenAI before release, describes a model that handles many subagents well on large tasks but asks about implementation details. In his own test, after two hours with the /goal command for a new level in Outer Wilds, the model had “blocked” on a question about a video file name.
In Willison's test with a three.js animation, in which Yonkers becomes a giant mech and walks through Manhattan, Queens and Brooklyn, with openstreetmap data and the Ultra effort level in Codex CLI, neither model did well enough. The results were worse than those of Claude Fable 5.2 and Gemini 4 Carbon.