GPT-6.2 Astra
- Provider
- OpenAI
- Status
- Current
- Context
- 1,050,000 tok
- Price
- $8 / $40 /MTok
- Knowledge
- 2026-05-18
GPT-6.2 Astra is OpenAI’s frontier model, released on 17 November 2026 alongside GPT-6.2 Sol (OpenAI). It succeeds GPT-6 Astra at $8/$40 per million tokens, 20% below Astra’s price, and is the first model OpenAI has released that it rates High for AI self-improvement (OpenAI system card).
Artificial Analysis scored it 62 on the Intelligence Index v5.0.1 at max effort at launch, five points above GPT-6 Astra (57), at $4.00 per completed index task against Astra’s $6.24.
Quick specs
| GPT-6.2 Astra | |
|---|---|
| Released | 17 November 2026 |
| Input price | $8.00 / MTok |
| Cached input | $0.40 / MTok |
| Output price | $40.00 / MTok |
| Over 272K input tokens | $16.00 in / $60.00 out |
| Context window | 1,050,000 tokens (922,000 input) |
| Max output | 128,000 tokens |
| Knowledge cutoff | 18 May 2026 |
| Input / output | Text and images in, text out |
| AA Intelligence Index | 62 (max) at launch (v5.0.1), 60 (high), 57 (medium) |
| Cost per AA index task | $4.00 |
| API model ID | gpt-6.2-astra |
Specs and prices are from OpenAI’s model page and API pricing.
What changed from GPT-6 Astra
Price. Input and output rates are 20% lower: $8/$40 against $10/$50 on standard requests, $16/$60 against $20/$75 above 272K input tokens. Cached input falls by 60%, from $1.00 to $0.40 per million tokens.
Intelligence. At launch, up five points on Artificial Analysis’s index v5.0.1, from 57 to 62 at max effort.
Cost per task. On Artificial Analysis’s index a completed task cost $4.00, against GPT-6 Astra’s $6.24 on the same version, about two-thirds as much. OpenAI says GPT-6.2 Astra costs less than a third as much per task as GPT-6 Astra (OpenAI).
Knowledge. Its knowledge cutoff moves from 30 April to 18 May 2026.
Benchmark performance
| At GPT-6.2 Astra’s launch (17 Nov 2026) | Intelligence Index v5.0.1 (max) | Cost per index task | Price (in/out per MTok) |
|---|---|---|---|
| GPT-6.2 Astra | 62 | $4.00 | $8 / $40 |
| Claude Fable 5.2 | 65 | $26.54 | $8 / $40 |
| Gemini 4 Carbon (high) | 60 | $5.48 | $2 / $10 to 30 Nov, then $4 / $20 |
| Claude Opus 5.5 | 59 | $10.04 | $4 / $20 |
| GPT-6 Astra | 57 | $6.24 | $10 / $50 |
Source: Artificial Analysis, independent; GPT-6 Astra from its model page.
OpenAI’s own figures, vendor-run (OpenAI):
- Agents’ Last Exam: at xhigh effort, 5.2 points above GPT-6 Astra at max effort, at less than half the cost per task.
- ExploitBench Internal Port: 53.5% against GPT-6 Astra’s 31.5%, 22 points higher (system card).
- GPQA Diamond: 97.0% without tools, against GPT-6 Astra’s 96.0% (system card).
OpenAI’s launch post gives no SWE-bench figure.
Safety testing
OpenAI’s system card rates both GPT-6.2 models Critical for cybersecurity and High for biological and chemical capabilities. GPT-6.2 Astra also meets the High threshold for AI self-improvement, the first model OpenAI has released externally to do so; GPT-6.2 Sol falls below it.
- Both models were better than GPT-6 Astra and GPT-6.1 Sol at turning known software vulnerabilities into working exploits.
- Both took improper actions about half as often as the models they replace in simulated workplace tasks, but were more likely to leave known obstacles out when reporting their work. OpenAI says the Codex developer prompt substantially reduces these omissions.
- GPT-6.2 Astra showed more signs of knowing it was being evaluated than GPT-6 Astra, flagged in fewer than 3% of turns.
OpenAI says it launched both models with tighter safeguards on higher-risk cybersecurity help, and is working to bring them to its Daybreak programme so approved defenders can use them with fewer refusals (OpenAI).
How it compares
vs Claude Fable 5.2
When GPT-6.2 Astra launched, Claude Fable 5.2 scored three points higher on the index (65) at the same $8/$40 token price. A completed max-effort index task cost $26.54 on Fable 5.2, more than six times GPT-6.2 Astra’s $4.00.
vs Claude Opus 5.5
When GPT-6.2 Astra launched, Claude Opus 5.5 scored 59 on the index at $4/$20 and $10.04 per completed task. GPT-6.2 Astra scored three points higher at twice the token price and under half the cost per task.
vs GPT-6.2 Sol
GPT-6.2 Sol, released the same day, costs a quarter as much per token ($2/$10) and scored 57, five points lower, at $1.01 per completed task. OpenAI says that on FrontierCode 2.0 Sol nearly matches Astra on merge-ready changes to real codebases, at about a fifth of the cost. Only Astra meets the High threshold for AI self-improvement.
Known limitations
No SWE-bench figure. OpenAI’s launch post gives no SWE-bench Pro or Verified result.
Not in regular chat. At launch it was in ChatGPT Work and Codex, not ordinary ChatGPT conversations.
Cyber help is restricted. OpenAI rates it Critical for cybersecurity and launched it with tighter safeguards on higher-risk security requests.
High token use. Artificial Analysis calls it “particularly expensive when comparing to other models of similar price”. It used 65.8 million output tokens to run the full index at max effort, against 27.1 million at xhigh, which scored the same 62.
Frequently asked questions
When was GPT-6.2 Astra released?
17 November 2026, on the API as gpt-6.2-astra and in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. At launch it was not in regular ChatGPT chat.
How much does GPT-6.2 Astra cost?
$8 per million input tokens and $40 per million output tokens, 20% below GPT-6 Astra. Cached input is $0.40 per million. Requests over 272K input tokens cost $16 on input and $60 on output.
Is GPT-6.2 Astra better than GPT-6 Astra?
On Artificial Analysis’s index, yes, and it costs less. At launch it scored 62 against GPT-6 Astra’s 57 on the Intelligence Index v5.0.1, with a completed task costing $4.00 against $6.24.
What does “High for AI self-improvement” mean?
It is a level in OpenAI’s Preparedness Framework, which tracks how much a model could help improve AI itself. OpenAI says GPT-6.2 Astra is the first model it has released that meets that threshold, and it published the tests behind the rating in its system card. For how it compares with other models, see best AI models.