GPT-6.2 Astra at $8 / $40: OpenAI cuts its flagship price by a fifth, matches Claude Fable 5.2 and ships the first model it rates 'High' for AI self-improvement. What changed and what to test
The headline is a price cut; the system card is the story. OpenAI has released a model it says can materially help build the next one, under added safeguards rather than restricted access.

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Monday afternoon to the API, to Codex and to ChatGPT Work for Plus, Pro, Enterprise and Edu customers. The developer model index lists Astra at $8 per million input tokens and $40 per million output, against $10 and $50 for the GPT-6 Astra it replaces, and Sol at $2 and $10, unchanged from GPT-6.1 Sol. Both have a 1.05 million-token context window, a 128,000-token output limit, and are offered in the Ultrafast speed tier. OpenAI's launch page on its main site could not be reached from our build environment, so the figures here come from the model index and the community announcement.
The price
At $8 / $40 Astra now sits exactly on Claude Fable 5.2, which Anthropic priced at the same level on November 9, and under Gemini 4 Carbon, whose launch price of $2 / $10 steps up to $4 / $20 on December 1. Ten weeks ago OpenAI's top tier cost $50 per million output tokens; the new price is a 20% cut delivered alongside a capability increase, which is the pattern the industry had lost during the GPT-6 generation. TestingCatalog reports cached input at $0.40 for Astra and $0.10 for Sol, down from $1.00; we have not confirmed those on OpenAI's pricing page and have left the cached column blank in our tracker until we can.
Sol's price did not move. The gap between the two tiers is therefore four times on input and output, the same ratio as Anthropic's Fable-to-Opus step, and the choice between them is now purely about capability.
The system card
The card, dated November 17, classifies both models as Critical under OpenAI's cybersecurity category and High for biology and chemistry, the same tiers as GPT-6 Astra. The new line is AI self-improvement: Astra is the first externally deployed OpenAI model rated High, the threshold at which the preparedness framework requires added safeguards for the model's use in AI research. The card reports SEC-Bench Pro scores of 86.75% for Astra and 88.34% for Sol against 85.4% for GPT-6 Astra, KernelGen at 76.92% against 66.72%, and GPQA Diamond at 97.0% and 96.0%. Secondary coverage adds an Artificial Analysis Elo of 1,318 for Astra against 1,271 for Fable 5.2 and ARC-AGI-2 at 95.8% for $0.50 a task; treat those as unconfirmed.
Two things are notable. The smaller Sol model beats Astra on the security benchmark, which suggests the Critical rating is about the tier, not the parameter count, and that the cheaper model is the one an attacker would rent. And the self-improvement rating is applied to a model released on a public API with no vetting, where GPT-6 Cyber was held behind the Daybreak programme. OpenAI's framework distinguishes between capabilities that warrant restriction and capabilities that warrant mitigation; it has put self-improvement in the second group.
The counter-argument
Vendor benchmark deltas of one to ten points on tests the vendor selected are not evidence that your workload improves. The KernelGen jump is large and specific; the GPQA numbers are saturated. The cut in Astra's price may also say as much about OpenAI's inference costs falling as about competitive pressure, and a price that falls 20% in ten weeks can fall again, which argues against long reserved-capacity commitments at today's rate.
What to test
Run your existing eval set on both models with the same prompts and compare cost per completed task, not per token; Sol at a quarter of the price may finish more of your tasks per dollar. Check anything version-sensitive against the new cutoffs. If you use prompt caching heavily, wait for OpenAI's pricing page to confirm the cached rate before migrating. And if you run a red team, the Critical cybersecurity rating on a $2 model is your new baseline for what an adversary can afford.
Sources
OpenAI developer community announcement ↗OpenAI model index ↗GPT-6.2 Astra and Sol system card (PDF) ↗The story so far
- Nov 17, 2026 Qwen4-Max at $1.50 / $4.50: 2.4 trillion parameters, 71 billion active, a 1M window, and a benchmark story that depends …
- Nov 15, 2026 Microsoft's Lazarus: seven edited neurons out of 12,000 backdoor a coding agent 98.8% of the time and pass every standar…
- Nov 15, 2026 AlphaGenome 2: a conditional DNA model trained on 107 million cell profiles finds nearly twice the expression-changing v…
- Nov 14, 2026 OpenAI disbands its standalone robotics team for the second time, betting GPT-6 Astra's 74.8% simulation score makes a s…
Key terms in this story
- Preparedness framework (OpenAI) OpenAI's internal policy for measuring a model's dangerous capabilities before release and deciding which safeguards or restrictions apply.
- Context window The maximum text a model can consider in one request: prompt, documents, conversation and answer combined.
- Prompt caching Charging a reduced rate for the repeated prefix of a prompt, such as a system prompt or a document reused across requests.
- Inference Running a trained model to produce output. What you pay for when you call an API.
- Benchmark A standard test used to compare models, such as a set of coding tasks or exam questions.
More in Models & research
All →
Gemini 4 Neon: $3 / $15 per million tokens, 1M context, four thinking levels and a 32% cost-per-task claim against Carbon. Where it sits in Google's four-model lineup and what to test
Google's fourth Gemini 4 release in a month is priced to become the default for paying users and the orchestrator for Hydrogen sub-agents; the benchma…

Qwen4-Max at $1.50 / $4.50: 2.4 trillion parameters, 71 billion active, a 1M window, and a benchmark story that depends on whose table you read
Alibaba's new flagship is 25% cheaper than its predecessor and competitive on agentic coding, but independent indices put it mid-pack; the model to wa…

Microsoft's Lazarus: seven edited neurons out of 12,000 backdoor a coding agent 98.8% of the time and pass every standard eval. What ML teams should change about model provenance
A model file can now carry a trigger-activated payload the way a binary can, and the only reliable detector is a diff against the original weights; th…
Get the brief in your inbox
One email every morning: the stories that mattered in AI, each with a one-line "why it matters" for your role. Free, no ads, unsubscribe any time.
Published daily by AI Briefing. We never share your address.