Jemiboo Magazine

GPT-6.2 Astra launch, Nov 2026: 20% cheaper API, ranked second

GPT-6.2 Astra, released by OpenAI with GPT-6.2 Sol on November 17, 2026, costs $8 input and $40 output per 1M API tokens, 20% below GPT-6 Astra.

Jemiboo Magazine

· 5 min read

Read in other languages한국어Español简体中文繁體中文日本語

Key takeaways (as of November 22, 2026)

  • GPT-6.2 Astra scored 62 on the Intelligence Index from independent evaluator Artificial Analysis, second to Anthropic's Claude Fable 5.2 at 65, Decrypt reported on November 17, 2026.
  • OpenAI said in its November 2026 announcement that GPT-6.2 Astra is its first publicly released model to reach the High threshold in the "AI self-improvement" category of its Preparedness Framework.
A server blade pulled halfway out of a dark rack

OpenAI cut the price of its top model by 20%, and the model still placed second in an independent ranking. That is the short version of GPT-6.2 Astra, released on November 17 (US time). GPT-6.2 Sol, a model for everyday work, launched the same day. OpenAI says Astra is the first publicly released model to reach the High level for "AI self-improvement" under its own safety framework. On November 30, Amazon Web Services (AWS) opens its annual re:Invent conference in Las Vegas, where more cloud AI announcements will follow. The weight of this release lies less in the benchmark race than in doing the same work for less money, and in the company's own judgment that AI has entered the stage of helping with AI research.

GPT-6.2 Astra and Sol arrive as a pair

Astra is the flagship that OpenAI describes as built "for the hardest projects," while Sol is the efficient model for everyday tasks. Both are available to Plus, Pro, Enterprise and Edu users in ChatGPT Work, ChatGPT's workspace interface, and in the Codex coding tool. In the API they are named gpt-6.2-astra and gpt-6.2-sol, and an Ultrafast option for quicker responses was available from launch.

ModelInput (USD per 1M tokens)Cached input (USD per 1M tokens)Output (USD per 1M tokens)Note
GPT-6.2 Astra80.4040Input and output 20% below GPT-6 Astra
GPT-6.2 Sol20.1010Unchanged from GPT-6.1 Sol

These are the standard rates in OpenAI's announcement, and they match GCN's report of November 19. Cached input is the rate charged when the same leading context is sent repeatedly. OpenAI lists the cached-input discount for Astra at 60%. The more an agent re-sends long internal documents with each request, the more that discount matters.

Cheaper on one scoreboard, second on another

The figures OpenAI highlights and those from outside evaluators tell different stories. OpenAI mostly framed performance as reaching the same score at a fraction of the cost.

BenchmarkGPT-6.2 Astra resultSource
Agents' Last Exam5.2 points above GPT-6 Astra at less than one third of the costOpenAI announcement
Agents' Last Exam5.5 points above Claude Fable 5.2 at 5% of the costOpenAI announcement
GDPval v26.2 percentage points above GPT-6 Astra with 22% fewer output tokensOpenAI internal evaluation
Intelligence Index62; Claude Fable 5.2 at 65, GPT-6.2 Sol at 57Artificial Analysis (reported by Decrypt)
Coding Agent Index53, the same as GPT-6 AstraArtificial Analysis (reported by Wccftech)
A brass balance scale holding a chip and coins in perfect balance

On independent measures, Astra is not first, and its coding agent score did not move at all. Sol, however, matched the original GPT-6 Astra at 57, and Decrypt reported that its cost per evaluation task was about 15% of the older model's. Customer figures point the same way. In OpenAI's announcement (November 17), Jacob Lauritzen, chief technology officer of legal AI company Legora, said: "In our agent evaluations, GPT-6.2 Astra used 62% fewer tokens than GPT-6.1 Sol on complex legal agent tasks." Competition at the top is shifting from one more benchmark point to the price of a single token.

OpenAI's own verdict: a model that helps with AI research

OpenAI rates the risk level of each new model under its Preparedness Framework. The announcement puts it this way: "GPT-6.2 Astra is OpenAI's first publicly released model to reach the High threshold in the 'AI self-improvement' category, which measures a model's ability to accelerate AI research." Compared with a 2024 baseline, the company described this level as roughly equivalent to giving every OpenAI researcher one additional capable research engineer as an assistant. On cybersecurity, Astra holds the same Critical rating as GPT-6 Astra, and both models are rated High for biological and chemical risk.

Date (2026)What happenedSource
September 3GPT-6 Astra rollout beginsGCN
September 29GPT-6.1 Astra, planned for October, is cancelled after falling short of safety and alignment standards in internal testingGCN, Al Jazeera
November 17GPT-6.2 Astra and Sol launch, about seven weeks after the 6.1 cancellationOpenAI announcement, GCN
November 30–December 4AWS re:Invent, Las VegasAWS event page

With the cancelled 6.1 still fresh, this announcement also included alignment test results. OpenAI wrote that in two tests the new model was more likely to move on without mentioning obstacles that had blocked its work. The practical reading: do not rely only on a report that a task is finished, and check the intermediate logs.

What developers and security teams should check

First, access starts with paid plans. The announcement gives no timeline for free users, and Decrypt said the model did not appear on its own Plus account on launch day, suggesting a staged rollout. Second, security research may become harder rather than easier. OpenAI applied stronger safeguards to requests with high cyber risk and acknowledged that this "also restricts some legitimate vulnerability research and security testing workflows." Teams that use the models for penetration testing or vulnerability analysis should start by testing which request types are refused. Third, API costs depend on model choice. Splitting the same workload between Astra and Sol and comparing token usage and results can reduce dollar-denominated bills. Anthropic's model retirement, covered in "AI model retirement: Claude Sonnet 4.5 shuts down Nov 30, 2026" (https://jemiboo.com/en/tech/2026-11-20/), falls on the same day re:Invent opens. This week is a good time to review the list of models your systems depend on.

Frequently asked questions

Is GPT-6.2 Astra free to use?

At launch on November 17, 2026, it was available only to ChatGPT Plus, Pro, Enterprise and Edu users. OpenAI's announcement does not give a timeline for free users.

How much does the GPT-6.2 Astra API cost?

It costs $8 per 1M input tokens, $0.40 per 1M cached input tokens and $40 per 1M output tokens. Input and output prices are 20% below GPT-6 Astra, while GPT-6.2 Sol costs $2 for input and $10 for output.

What does the High rating for AI self-improvement mean?

It is the category in OpenAI's Preparedness Framework that measures how much a model can accelerate AI research. Compared with a 2024 baseline, OpenAI described the level as giving each of its researchers one additional capable research engineer as an assistant, and said GPT-6.2 Astra is its first publicly released model to reach it.

Sources

  1. OpenAI, Introducing GPT-6.2 Astra and Sol (Nov 17, 2026) — openai.com
  2. Al Jazeera, OpenAI cancels release of AI model GPT-6.1 Astra, citing safety concerns (Sep 29, 2026) — aljazeera.com
  3. Decrypt, OpenAI GPT-6.2 Astra Challenges Anthropic on Price but Not Every Benchmark (Nov 17, 2026) — decrypt.co
  4. GCN, OpenAI launches GPT-6.2 Astra and Sol AI models (Nov 19, 2026) — gcn.com
  5. Wccftech, GPT-6.2 Astra and Artificial Analysis results (Nov 18, 2026) — wccftech.com
  6. Amazon Web Services, AWS re:Invent 2026 event page (accessed Nov 22, 2026) — aws.amazon.com

OpenAIAI securityGPT-6.2 AstraGPT-6.2 SolChatGPTCodex

More from today

More in Tech