GPT-6.2 Astra
OpenAI's most capable model, priced 20% below GPT-6 Astra
Context window
1.05M
Input / 1M tokens
$8.00
Output / 1M tokens
$40.00
Provider
OpenAI
Cached input $0.40 per 1M (60% below GPT-6 Astra); cache writes $10 (30-minute) or $20 (12-hour). Prompts over 272K input tokens cost 2x input and cache rates and 1.5x output for the full request ($16/$60, cached $0.80). Batch is half price ($4/$20); the 2-hour and 8-hour background completion windows (Responses API, public beta) are 30% and 40% below standard. Fast mode is 2x ($16/$80). Ultrafast is 3x standard through February 28, 2027 and 6x from March 1, 2027. No Flex tier. Regional (US or EU) processing adds 10%. Max input 922K tokens, max output 128K; knowledge cutoff May 18, 2026. · Data verified 2026-11-20
GPT-6.2 Astra (November 17, 2026) is OpenAI's most capable model and the successor to GPT-6 Astra, aimed at the hardest coding, science, legal and finance work. It costs $8/$40 per 1M tokens, 20% below GPT-6 Astra and the same list price as Claude Fable 5.2, with cached input cut 60% to $0.40. It has a 1.05M-token context window (922K max input, 128K max output), accepts text and images, and has a knowledge cutoff of May 18, 2026. VentureBeat counts it ahead of GPT-6 Astra on all nine of OpenAI's head-to-head launch charts, including 78.3% vs 73.1% on Agents' Last Exam v1.1, where it also leads Claude Fable 5.2 (72.8%) at a fraction of the cost per task; Fable 5.2 stays ahead in OpenAI's own charts on FrontierCode 2.0 (49.1% vs 44.2%) and the Artificial Analysis Intelligence Index (64.6 vs 62.1). OpenAI rates it Critical for cybersecurity and says it is the first model it has released that meets its High threshold for AI self-improvement; the tighter safeguards also restrict some legitimate security research. Early hands-on reviews collected by VentureBeat were mixed, with some developers finding it hard to steer and others praising its 3D, visual and computer-use work.
How to access it
OpenAI API since November 17, 2026 (model ID gpt-6.2-astra) on the Responses, Chat Completions and Batch endpoints, with Fast and Ultrafast processing; tool calling requires the Responses API. In ChatGPT it is in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users (Enterprise and Edu admins must enable it), chosen with the Power slider in Work or /model in the Codex CLI. Ultrafast with Astra needs the $500 Pro plan or an Enterprise or Edu workspace where an owner turns it on. Also on Vercel AI Gateway (openai/gpt-6.2-astra).
Strengths
- ✓78.3% on Agents' Last Exam v1.1 vs 73.1% for GPT-6 Astra and 72.8% for Claude Fable 5.2, at $2.89 per task vs $60.05 for Fable (OpenAI-reported)
- ✓15.0% on Harvey LAB-AA v1.1 legal tasks, about double Claude Fable 5.2's 7.2% (OpenAI-reported)
- ✓77.1% on Terminal-Bench-Science 0.1 at max effort vs 68.1% for GPT-6 Astra (OpenAI-reported); first on Vals AI's Science Index and Terminal-Bench-Science leaderboards
- ✓$8/$40 per 1M tokens, 20% below GPT-6 Astra, with cached input at $0.40
- ✓Fewer output tokens per task: OpenAI reports 22% fewer than GPT-6 Astra on GDPval v2 while scoring 6.2 points higher
- ✓1.05M-token context window and 128K max output
Best for developers who...
When to choose it (and when not to)
Reach for GPT-6.2 Astra when...
- →You need OpenAI's strongest model for long, difficult projects
- →You use GPT-6 Astra today: GPT-6.2 Astra costs 20% less and scored higher on every head-to-head chart OpenAI published
- →Science, legal or finance work, where its launch benchmark leads are largest
Look elsewhere if...
- ✕Everyday coding and writing: GPT-6.2 Sol costs a quarter as much and came within half a point on FrontierCode 2.0 at max effort
- ✕Repository-level coding where Claude Fable 5.2 leads in OpenAI's own chart (49.1% vs 44.2% on FrontierCode 2.0)
- ✕Security research and penetration testing: OpenAI says the tighter safeguards restrict some legitimate workflows (approved teams can use GPT-6 Cyber through Daybreak)
- ✕Code that sets temperature, top_p, logprobs or reasoning effort none, which these models reject
How to use it
- ›Start at medium reasoning effort; if older code used none or minimal, move to low (OpenAI GPT-6.2 guide)
- ›Use the Responses API for tool calling
- ›Keep earlier turns unchanged so reasoning carries over and the prefix stays eligible for prompt caching
- ›To change reasoning effort mid-conversation, append a configuration_update item before the next user message
Quickstart
Pythonfrom openai import OpenAI
client = OpenAI()
response = client.responses.create(
model="gpt-6.2-astra",
reasoning={"effort": "medium"},
input="Your prompt here"
)
print(response.output_text)Install with pip install openai. OpenAI's GPT-6.2 guide suggests starting at medium effort; tool calling needs the Responses API.
API model id: gpt-6.2-astra
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| Agents' Last Exam v1.1 | 78.3% | OpenAI launch chart as reported by DataCamp and VentureBeat; GPT-6 Astra 73.1%, Claude Fable 5.2 72.8% (max), Gemini 4 Carbon 77.4%, GPT-6.2 Sol 73.0% |
| GDPval v2 | 50.4% | OpenAI internal evaluation, as reported by VentureBeat; GPT-6 Astra 44.2%, Claude Fable 5.2 44.8%, Gemini 4 Carbon 48.8% |
| Harvey LAB-AA v1.1 | 15.0% | OpenAI launch chart as reported by DataCamp and VentureBeat; GPT-6 Astra 8.6%, Claude Fable 5.2 7.2% |
| FrontierCode 2.0 | 44.2% (max) | OpenAI launch chart as reported by DataCamp and VentureBeat; GPT-6.2 Sol 43.8%, GPT-6 Astra 42.8%, Claude Fable 5.2 49.1%, Claude Opus 5.5 45.5% (max) |
| Terminal-Bench-Science 0.1 | 77.1% (max) | OpenAI launch chart as reported by DataCamp and VentureBeat; GPT-6 Astra 68.1%, Claude Fable 5.2 73.3% (Artificial Analysis harness; Anthropic reports 74.6% in Claude Code) |
| AutomationBench 1.0.6 | 52.4% | OpenAI launch chart as reported by DataCamp and VentureBeat; GPT-6 Astra 41.4%, Claude Fable 5.2 51.4%, Gemini 4 Carbon 53.9% |
| BenchCAD 2.1 Core | 82.5% | OpenAI launch chart as reported by DataCamp and VentureBeat; GPT-6 Astra 62.6%, Claude Fable 5.2 83.5% |
| Artificial Analysis Intelligence Index v5.0.1 | 62.1 | Per OpenAI's launch chart; GPT-6 Astra 56.9, Claude Fable 5.2 64.6. Artificial Analysis ranks it second, behind Fable 5.2 |
| GPQA Diamond | 97.0% (no tools) | OpenAI GPT-6.2 system card; GPT-6 Astra 96.0%, GPT-6.2 Sol 96.0% |
| SEC-bench Pro | 86.75% | OpenAI GPT-6.2 system card, pass@1; GPT-6 Astra 85.4%, GPT-6.2 Sol 88.34% |
| Terminal-Bench 4.0, Vals AI run | 63.13% | Vals AI leaderboard (mini-SWE-agent harness), ranked 5th; not an OpenAI-published figure |
Tools powered by GPT-6.2 Astra
Compare GPT-6.2 Astra
Compare GPT-6.2 Astra with any other model
Build a comparison →All model comparisons →