Overview
GPT-6.2 Astra is the flagship OpenAI released on November 18, 2026, alongside GPT-6.2 Sol for everyday work. The headline number this time is not how much the scores moved, but that they moved up while the price moved down. OpenAI puts the cost per task at under a third of GPT-6 Astra, and Artificial Analysis recorded the Intelligence Index climbing from 57 to 62 in the same window. Earlier frontier releases tended to trade a modest capability gain for a steeper price. This generation is the first in a while where the two point the same way.
On benchmarks it closes out the abstract reasoning line. ARC-AGI-3 comes in at 100.0%, the first perfect score on that benchmark, with ARC-AGI-2 at 95.8% and ARC-AGI-1 at 98.5%. On Agents' Last Exam it lands 5.2 points above GPT-6 Astra at less than a third of the cost, and 5.5 points above Claude Fable 5.2 at roughly 5% of the cost. GDPval v2 improves 6.2 points over GPT-6 Astra while producing 22% fewer output tokens. Terminal-Bench-Science 0.1 gains 9 points. On Harvey LAB-AA v1.1 the pass rate is more than double Claude Fable 5.2 at about a sixth of the cost. AA-3D puts it first at 1,318 Elo, against 1,271 for Claude Fable 5.2.
Pricing is the most concrete part of the release. Input is $8 and output $40 per million tokens, with cached input at $0.40. That is 20% below GPT-6 Astra on standard rates and a further 60% cut on cache reads. Sol sits at $2 and $10, unchanged from GPT-6.1 Sol, with more capability at the same token price. Ultrafast drops 50% for the next three months and costs API and Enterprise users only 3x standard, and it reaches Pro 200 subscribers for the first time, starting with Sol.
Two ratings in the system card deserve their own note. AI self-improvement is rated High, the first publicly released OpenAI model to land there, and both models carry a Critical cybersecurity rating. These are grades OpenAI assigned under its own framework, describing which band the evaluation placed the model in rather than anything that has happened. The ARC-AGI-3 perfect score also comes with two run configurations: the provider adapter path costs $4.1K and the standard path costs $24.2K for the same score, so the figure means little without the run mode attached.
The limits are worth stating plainly. The launch page gives no context window or maximum output token count, and neither does any parameter table circulating in third-party write-ups, so those two numbers should not be filled in from memory. Artificial Analysis also reports coding agents flat against the previous generation, with the 5-point index gain coming from other areas. Both facts belong together: a model can gain 5 points overall while standing still in one specific direction.
Whether a team should migrate comes down to three questions. Is the work multi-step, spanning many tools, with long outputs per run? That is the case most affected by 22% fewer output tokens and $0.40 cache reads. Is the workload concentrated in coding agents? If so, run it against an internal eval first, because the third-party verdict there is flat. And is the plan Pro 200? Ultrafast was previously out of reach at that tier and is now included, which adds response speed that money could not buy before.
Key Features
- Cost per task at one third: Agents' Last Exam lands 5.2 points above GPT-6 Astra at under a third of the per-task cost, and 5.5 points above Claude Fable 5.2 at roughly 5% of it. Capability and the bill move in the same direction for once.
- First perfect ARC-AGI-3: 100.0% on ARC-AGI-3, the first model to top that benchmark, with 95.8% on ARC-AGI-2 and 98.5% on ARC-AGI-1. The abstract reasoning line is now pressed against the ceiling.
- Input, output and cache all cut: $8 input, $40 output, $0.40 cached input per million tokens. That is 20% below GPT-6 Astra on standard rates and 60% lower on cache reads, so multi-turn readback saves more than the headline rate suggests.
- 22% fewer output tokens: GDPval v2 improves 6.2 points over GPT-6 Astra while emitting 22% fewer output tokens. Bills are settled on total tokens, so writing less beats discounting more on long jobs.
- Top of AA-3D: 1,318 Elo on AA-3D against 1,271 for Claude Fable 5.2. OpenAI singles out slides, websites and 3D design as the widest gaps, with cleaner code and stronger frontend output in the same batch of changes.
- Legal configuration leads Harvey LAB-AA: Astra for Law 6.2 reaches SOTA on Harvey LAB-AA v1.1, with a pass rate more than double Claude Fable 5.2 at about a sixth of the cost. High-value legal review gets a dedicated configuration.
- Ultrafast reaches Pro 200: Ultrafast is 50% cheaper for three months and costs API and Enterprise users 3x standard. Pro 200 subscribers get access for the first time, beginning with GPT-6.2 Sol.
- Available everywhere on day one: Live in ChatGPT Work, Codex and the API on launch day across Plus, Pro, Enterprise and Edu. OpenRouter, Perplexity, OpenClaw and Conductor picked it up the same day.
Use Cases
- Multi-step professional work: analyzing campaigns, building decks and assembling financial tables, the exact tasks in Agents' Last Exam
- Scientific workflows: 9 points above the prior generation on Terminal-Bench-Science 0.1, for work that runs in a shell and fits models to data
- Legal review and contract comparison: Astra for Law 6.2 leads Harvey LAB-AA v1.1, suited to reading clauses in bulk and flagging conflicts
- Interactive 3D and frontend pages: first on AA-3D, with websites and 3D design named explicitly in the release
- High-volume agent backends: $0.40 cache reads plus 22% fewer output tokens for services running thousands of long-context calls a day
- Latency-sensitive products: Ultrafast at 3x standard for three months, for interfaces that need fast first tokens without dropping to a smaller model
Pros
- Scores and cost move together for the first time: index up 5 points, per-task cost down to a third of GPT-6 Astra
- First perfect ARC-AGI-3 score opens real distance on abstract reasoning
- $0.40 cached input cuts multi-turn readback by 60%, which matters more than the headline rate
- 22% fewer output tokens lowers long-job bills by writing less rather than by discounting
- Ultrafast drops 50% and reaches Pro 200, lowering the speed floor for subscription tiers
- ChatGPT Work, Codex and the API all available on launch day, with third-party platforms following the same day
Pricing
API standard rates: $8 per million input tokens, $40 per million output, $0.40 for cached input. That is 20% below GPT-6 Astra at $10 and $50, and a further 60% cut on cache reads. Sister model: GPT-6.2 Sol is $2 input, $10 output and $0.10 cached, unchanged from GPT-6.1 Sol. Ultrafast: 50% off for three months after launch, charged at 3x standard for API and Enterprise users. Pro 500 can use it on both models; Pro 200 starts with GPT-6.2 Sol. Subscriptions: ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu, with Enterprise and Edu requiring an admin to enable it in the workspace. Channels: OpenAI API under the model name gpt-6.2-astra, ChatGPT and Codex; OpenRouter lists it at the same $8 and $40, with Sol at $2 and $10.
Summary
GPT-6.2 Astra is the flagship OpenAI shipped on November 18, 2026, built around a rare alignment of capability and cost: the Intelligence Index rose from 57 to 62 while cost per task fell to a third of GPT-6 Astra. It scores the first perfect 100% on ARC-AGI-3, 95.8% on ARC-AGI-2, sits 5.2 points above its predecessor on Agents' Last Exam at under a third of the cost, and doubles Claude Fable 5.2's pass rate on Harvey LAB-AA. Pricing is $8 input and $40 output with $0.40 cached input, 20% below the previous generation. The system card rates AI self-improvement High and cybersecurity Critical, both official grades. No context window or maximum output token count has been published, and third-party evaluation puts coding agents flat.
Version History
- GPT-6.2 Astra and Sol launch (2026-11-18): Flagship Astra and everyday Sol ship together. Astra reaches an Intelligence Index of 62 for second place, posts the first perfect 100% on ARC-AGI-3, and beats GPT-6 Astra by 5.2 points on Agents' Last Exam at under a third of the cost. Priced at $8 and $40 with $0.40 cached input, 20% below GPT-6 Astra; Ultrafast drops 50% and reaches Pro 200 for the first time.