AI models · OpenAI

GPT-6.2 Astra

GPT-6.2 Astra came out on 17 November 2026 as the successor to GPT-6 Astra. It always thinks before answering and takes in about a million tokens. OpenAI tuned it for professional work: slides, spreadsheets, legal and financial documents, coding and scientific research. In the API it costs 20% less than its predecessor, and OpenAI says it uses fewer tokens for the same task. In ChatGPT you only get it in Work and Codex, not in regular chat.

Strengths

  • OpenAI reports 77.1% on Terminal-Bench-Science 0.1, a test of scientific work in the terminal. GPT-6 Astra scores 68.1%, and Claude Fable 5.2 scores 73.3% as measured by Artificial Analysis.
  • On OpenAI's internal GDPval v2 test of long real-world projects, for example in finance and consulting, it scores 50.4%. GPT-6 Astra scores 44.2% and Claude Fable 5.2 44.8%.
  • Cheaper than its predecessor: $8 per million input tokens and $40 per million output tokens instead of $10 and $50. Cached input costs $0.40.
  • It scores 62 on the independent Artificial Analysis index v5.0.1, against 57 for GPT-6 Astra. One test task costs $4.00 with it and $6.24 with its predecessor.
  • OpenAI says it follows slide templates and industry conventions more closely. On the Harvey LAB-AA v1.1 legal test it scores 15.0%, more than double Claude Fable 5.2 (7.2%).

Weaknesses

  • Not in regular ChatGPT chat, only in ChatGPT Work and Codex. OpenAI still lists GPT-6 Pro mode in chat as powered by the older GPT-6 Astra.
  • Still expensive, four times the price of GPT-6.2 Sol, which OpenAI says nearly matches it at coding (FrontierCode 2.0: 43.8% vs 44.2%).
  • It does not win everywhere. It scores 62 on the Artificial Analysis index v5.0.1 against 65 for Claude Fable 5.2, and OpenAI reports 44.2% on FrontierCode 2.0 against 49.1% for Fable 5.2.
  • Slow to answer: about 49 tokens per second according to Artificial Analysis, and at the highest thinking level it takes roughly 3.5 minutes before it starts writing.
  • Stricter cybersecurity filters also hold back some legitimate vulnerability research, OpenAI says. OpenAI also admits that in some tests it more often left out an obstacle it had run into when reporting its work.

Best for

  • demanding professional work: analyses, slides, spreadsheets and memos
  • the hardest coding and autonomous agent work
  • scientific and technical research
  • 3D design and interactive websites
  • very long documents, up to about a million tokens

Not ideal for

  • high volumes of simple tasks, where GPT-6.2 Sol and GPT-6 Luna are cheaper
  • chat where you need an instant reply

How to use it in Europe

In ChatGPT it comes with Plus, Pro, Business, Enterprise and Edu, but only in ChatGPT Work and Codex, not in regular chat. On Enterprise and Edu an admin has to switch it on. Among personal plans, only Pro 500 gets the faster Ultrafast mode. Codex CLI needs updating to the latest version. Developers use it through the OpenAI API, which is available in Slovakia and Czechia. It supports EU data processing, which costs 10% extra in the API.

Price compared with similar models

1 3 10 30 Beam 1.01 Mistral Large 42.1 DeepSeek V4 Pro3.96 Qwen4-Max4.5 Qwen3.8-2.4T-A95B6 Grok 4.76 Gemini 4 Carbon10 Kimi K3.1 Pro10 Gemini 4 Neon15 Kimi K315 Claude Opus 5.520 Muse Wonder20 GPT-6.2 Astra40
Output price in USD per 1M tokens (log scale). Lighter bars are open-weight models.

Benchmark results

BenchmarkScoreSource
Artificial Analysis Intelligence Index v5.0.1 (max effort)62artificialanalysis.ai ↗, November 2026
Terminal-Bench-Science 0.1, max effort77.1%openai.com ↗ (vendor-reported), November 2026
GDPval v2, max effort50.4%openai.com ↗ (vendor-reported), November 2026
Agents' Last Exam v1.178.3%openai.com ↗ (vendor-reported), November 2026
FrontierCode 2.0, max effort44.2%openai.com ↗ (vendor-reported), November 2026

Benchmarks cover only a narrow slice of ability and vendors often report them at their best settings. Treat them as a rough guide.

More models from this vendor

  • GPT-6.2 Sol: OpenAI's balanced model for everyday work, coding and agents. Stronger and faster than GPT-6.1 Sol at the same price.
  • GPT-6.1 Terra: OpenAI's fast, cheaper model for working with images, operating a computer and everyday office work.
  • GPT-6 Luna: OpenAI's cheapest model, made for high volumes of simple tasks, sorting text and low-cost chatbots.
  • gpt-oss-120b: OpenAI's open model that a company can run on its own server when data must not leave its infrastructure.
  • gpt-oss-20b: OpenAI's smaller open model that runs on a computer with 16 GB of memory, even without internet access.

Similar models from other vendors

  • Claude Opus 5.5 (Anthropic): Anthropic's main model for demanding work, coding and long documents, and the one Anthropic recommends starting with.
  • DeepSeek V4 Pro (DeepSeek): DeepSeek's largest open-weight model for demanding text tasks at a low price. It does not understand images.
  • Gemini 4 Carbon (Google): Google's most capable model, available since 10 November 2026 in the paid API, for businesses and in Google AI Ultra. It scores 60 on the independent index.
  • Gemini 4 Neon (Google): Google's powerful model between Carbon and Hydrogen, out since 18 November 2026. In the app it comes with Google AI Pro in place of Argon.
  • Muse Wonder (Meta): Meta's strongest model for long-running agents and coding, still released as a preview.
  • Mistral Large 4 (Mistral AI): Mistral's largest model, now with open weights. Built for demanding tasks, long documents and agents, with data kept in the EU and a cheaper API than in preview.

Sources

Prices and models change often. The data comes from official vendor pricing pages and documentation, and every page shows when it was last checked. Before a larger purchase, confirm the price with the vendor.

← All AI models

Not sure which model fits your business?

I can help you pick a model and plan based on what you actually need, set it up safely for company data and train your team to use it well. Let’s start with an intro consultation.