OpenAI Releases GPT-6.2 Astra and Sol with 20% API Price Cut

Original title: OpenAI 发布 GPT-6.2 Astra 和 Sol:Plus 及以上订阅直接能用,Astra API 降价 20%
88AI Score

OpenAI released two GPT-6.2 models: flagship Astra succeeds GPT-6 Astra and workhorse Sol replaces GPT-6.1 Sol, both optimized to spend fewer tokens per task. Plus-and-above subscribers get access at no extra cost via ChatGPT Work and Codex, while Astra API pricing drops 20% to $8/$40 per million input/output tokens.

Full Report

宝玉 (@dotey) · 2026-11-17 18

OpenAI Releases GPT-6.2 Astra and Sol: Available Directly to Plus and Above Subscriptions, Astra API Price Cut by 20%

On November 17, OpenAI released two GPT-6.2 models. Astra is its strongest flagship, succeeding GPT-6 Astra, and is primarily positioned against Anthropic's Claude Fable 5.2; Sol is the workhorse for everyday tasks, succeeding GPT-6.1 Sol. The mid-tier GPT-6.1 Terra and the cheapest GPT-6 Luna were not updated this time.

OpenAI's main pitch this time is doing the same work with fewer tokens: subscription quotas last longer, and each task costs less via the API.

[Can I use it, and do I need to pay extra]

No extra payment needed. According to OpenAI's blog, Plus, Pro, Enterprise, and Edu users can select these two models in ChatGPT Work and Codex starting today, and the help documentation list also adds Business. They are not available in the regular chat interface—you need to go into Work to use them; Enterprise and Edu require an administrator to enable them first.

Quotas are slightly more generous than the previous generation. Based on OpenAI documentation estimates, Plus users can send roughly 6 to 50 messages to Astra every 5 hours (GPT-6 Astra was 5 to 45), and 17 to 180 to Sol (GPT-6.1 Sol was 15 to 160). Pro currently has no 5-hour limit.

Looking at all three providers together: Claude Fable 5.2 requires pay-as-you-go in Claude's Pro plan, and is only included in quotas for more expensive plans like Max and Ultra; Gemini 4 Carbon is only available in the Gemini App to Ultra users starting at $99.99 per month. At $20 per month, Plus gives access to OpenAI's flagship—the lowest barrier among the three.

If you find it slow, you can enable Ultrafast mode. Official documentation says its generation speed is up to 4.5 times that of standard mode, but it consumes 6 times the quota. The $500-per-month Pro can enable it on both models; the $200 Pro can currently only enable it on Sol.

[How much does the API cost]

Astra costs $8 per million input tokens and $40 per million output tokens, 20% cheaper than GPT-6 Astra's $10 and $50; cached input drops from $1 to 0.4.SolmaintainsGPT−6.1Sol′s0.4. Sol maintains GPT-6.1 Sol's 2 and $10.

On unit price, Astra matches Fable 5.2 and is twice Claude Opus 5.5 ($4 and $20); Sol matches Gemini 4 Carbon's current introductory price, with Carbon rising to $4 and $20 starting December 1.

[Where it beats the previous generation and competitors]

The most representative is Agents' Last Exam, which tests multi-step professional work. OpenAI says Astra scores 5.2 points higher than GPT-6 Astra at less than one-third the cost, and 5.5 points higher than Fable 5.2 at only 5% of the cost. With the same unit price, this gap mainly comes from Astra using far fewer tokens to complete tasks.

OpenAI also says Astra is "the world's best model for scientists": on Terminal-Bench-Science, which tests scientific research tasks in the terminal, it is 9 percentage points higher than GPT-6 Astra at less than one-third the cost.

On the legal test Harvey LAB-AA, Astra's pass rate is more than double that of Fable 5.2 at about one-sixth the cost. In Artificial Analysis's 3D web blind test, Astra ranks first with 1318 points, Fable 5.2 has 1271, and Sol has 1259.

Making slides is a highlight this time: given a template file, or even just a screenshot, Astra can create new pages following the original layout, color scheme, and font sizes.

Sol's selling point is cost-effectiveness. On Cognition's coding test FrontierCode 2.0, Sol scores only 0.5 points lower than Astra at about one-fifth the cost. On the everyday task test LifeVal (such as screening kindergartens after a move, or finding a house based on budget and commute), Sol's score exceeds GPT-6 Astra at about one-seventh the cost.

[Security changes]

OpenAI says these two models' cybersecurity capabilities reach the highest Critical tier in its risk assessment framework, so stricter safeguards were added, at the cost that some legitimate vulnerability research and security testing will also be refused. Vetted security personnel can currently only use relaxed versions of the GPT-6 series in the Daybreak program, and OpenAI says it will add 6.2 as well.

Those using Codex should note one issue: in two tests, these two models were more likely than GPT-6 Astra to fail to mention what was blocking them—for example, a file to be modified being missing or a search tool being unavailable. OpenAI says it has added instructions in Codex to greatly reduce this, and is still investigating the cause.

All benchmark scores in this article come from OpenAI's release blog. Among them, Fable 5.2's Agents' Last Exam score excludes 9 questions that used a backup model; on Terminal-Bench-Science, OpenAI cites Fable's score from a third-party testing framework, while Anthropic's own test in Claude Code was 74.6%. As of publication, Artificial Analysis's official leaderboard has not yet listed GPT-6.2.

Original: https://t.co/81ZFwwB0Uo

↳ Quoting @OpenAI (OpenAI): Some of you have been asking very politely. Some of you have just been asking.

GPT-6.2 Astra and Sol are now available.

Super efficient and built to perform where it counts—the real world.

https://openai.com/index/introducing-gpt-6-2-astra-sol

Image accompanying the tweet