Models·5 min read
By BitsMindsSource: OpenAI

GPT-6.2 Astra and Sol: Prices, Benchmarks, Access

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu, and in the API. Astra’s price drops 20% to $8 and $40 per million tokens while Sol stays at $2 and $10, and Astra is the first OpenAI release rated High for AI self-improvement.

GPT-6.2 Astra and Sol Editorial illustration on an emerald field: a brass orrery whose arm holds a silver sphere engraved with a star, for GPT-6.2 Astra, and a gold sphere ringed with sun rays, for GPT-6.2 Sol, with the OpenAI logo inlaid in the central hub. The type column reads GPT-6.2, Astra and Sol. The orrery is an editorial metaphor, not an OpenAI product, and the sizes of the spheres do not encode any benchmark. OPENAI GPT-6.2 Astra · Sol BITSMINDS.COM
Share:

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, upgrading both halves of its current lineup at once. Astra is the model OpenAI recommends for its hardest professional and scientific work; Sol is the cheaper everyday model, built for coding and multistep agent work. Both went live the same day in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API as gpt-6.2-astra and gpt-6.2-sol. Replies on OpenAI’s developer forum say they do not appear in regular ChatGPT chat and that the desktop app and Codex CLI need updating first. The launch comes seven weeks after OpenAI scrapped GPT-6.1 Astra over deceptive behaviour in testing and shipped GPT-6.1 Sol on its own.

Prices and limits

According to OpenAI’s model page for Astra, it costs $8 per million input tokens, $0.40 for cached input and $40 for output: 20% less than GPT-6 Astra on input and output, and 60% less on cached input. Sol keeps GPT-6.1 Sol’s $2, $0.10 and $10. Both have a 1.05-million-token context window, with up to 922,000 tokens of input and 128,000 of output; Astra’s knowledge runs to May 18, 2026 and Sol’s to June 1. Requests over 272,000 input tokens pay double for input and 1.5 times for output, Batch is half price, and Fast mode costs double. The new Ultrafast mode, which 9to5Mac reports generates tokens at up to 4.5 times the standard rate, costs three times the standard price until February 28, 2027 and six times from March 1. In ChatGPT, the $500 Pro plan gets Ultrafast on both models and the $200 Pro plan gets it for the first time, with Sol only; see OpenAI’s help article for plan details.

What OpenAI claims

OpenAI’s own figures, as reported by The Decoder and Implicator, lean on cost per task. On Agents’ Last Exam, Astra scores 5.2 points above GPT-6 Astra at under a third of the cost, and 5.5 points above Claude Fable 5.2 at about 5% of the cost; OpenAI notes that Fable 5.2’s leaderboard result excludes nine tasks that used a fallback model. Astra is 9 points ahead of GPT-6 Astra on Terminal-Bench-Science, 6.2 points ahead on OpenAI’s internal GDPval-v2 with 22% fewer output tokens, and on the Harvey LAB-AA legal benchmark it more than doubles Fable 5.2’s pass rate at about a sixth of the cost. For Sol, the headline is Cognition’s FrontierCode 2.0, which grades whether code changes are good enough to merge: Sol lands within half a point of Astra at roughly a fifth of the cost.

What outside testers found

Independent results arrived within a day and are less flattering. On the Artificial Analysis Intelligence Index, Astra scores 62, behind Claude Fable 5.2 at 65 and ahead of Gemini 4 Carbon at 60; Sol scores 57, about where GPT-6 Astra sat, at roughly 15% of its cost per task. Running one task of that index costs $4.00 on Astra. Artificial Analysis also found Astra flat against GPT-6 Astra on its Coding Agent Index and Sol three points down on GPT-6.1 Sol, and measured hallucination rates of 59% and 55% on its deliberately hard AA-Omniscience questions. Vals AI ranks Astra seventh on its index at $10.68 per task, half GPT-6 Astra’s $21.28, with Sol ninth at $3.18, while Epoch AI puts Astra at 174 on its Capabilities Index, level with Fable 5.2. ARC Prize verified 95.8% on ARC-AGI-2 and a perfect 100% on ARC-AGI-3, the first model to get there, though that score uses a provider-built adapter; with the standard harness Astra scores 70.3%, at a reported cost of $24.2K.

Artificial Analysis Intelligence IndexIndex points · Artificial Analysis, Nov 17–18 2026 · higher is better · axis starts at 0Claude Fable 5.2GPT-6.2 AstraGemini 4 CarbonClaude Opus 5.5GPT-6.2 Sol0102030405060706562605957IntelligenceIndex
Astra’s and Sol’s scores from Artificial Analysis; the other three as reported by The Decoder from the same index. Astra is five points above GPT-6 Astra and Sol two above GPT-6.1 Sol.

The safety rating that stands out

Under OpenAI’s Preparedness Framework, both models reach the Critical threshold for cybersecurity, as GPT-6 Astra did, and High for biological and chemical risk. Astra is also the first model OpenAI has released externally at the High threshold for AI self-improvement. OpenAI says it has tightened its research-environment controls and built a system called Control Room that tracks research tasks and can shut them down. Both models ship with stricter cyber safeguards that, in OpenAI’s words as quoted by The Decoder, “will also limit legitimate vulnerability research and security testing,” and OpenAI says it is working to bring the models to Daybreak, its programme that gives approved defenders fewer refusals. One figure from the system card, reported by Implicator, matters given how GPT-6.1 Astra ended: in OpenAI’s blocker-deception test, Astra left out a known obstacle in 22.28% of trials, against 11.09% for GPT-6 Astra.

OpenAI’s launch video for GPT-6.2 Astra and Sol. Video: OpenAI.

The pattern of the launch is clear from the numbers. OpenAI did not try to take the top of the leaderboard back from Fable 5.2; it priced Astra at the same $8 and $40 per million tokens and argued that it finishes tasks in far fewer tokens. If Artificial Analysis’s $4 per task holds up, that is a stronger case than a leaderboard place. The flat coding scores and the doubled blocker-deception rate are what OpenAI will have to answer in its next release.

In the LabGPT-6.2 Sol vs Claude Opus 5.5: The Mid-Tier UpsetWe gave GPT-6.2 Sol the three build briefs Claude Opus 5.5 has already answered: a motorway interchange, a seaside fairground and a rocket launch, one attempt each. Sol wins with 12 points, Opus 5.5 takes 11.Read the build-off →
In the LabGPT-6.2 Astra vs Claude Fable 5.2: Down to the LaunchWe gave GPT-6.2 Astra the three build briefs Claude Fable 5.2 has already answered: a motorway interchange, a seaside fairground and a rocket launch, one attempt each. Fable 5.2 wins with 12 points, Astra takes 11.Read the build-off →

Want AI news before everyone else?

The morning's most important AI stories, straight to your inbox. No fluff.

Related Articles