$ cat wiki/models/gpt-6-2-astra.md
GPT-6.2 Astra
Spec
| Attribute | Value |
|---|---|
| Developer | OpenAI |
| Released | 2026-11-17 |
| Announced | 2026-11-17 |
| Context window | 1,050,000 tokens; maximum input 922,000 tokens; maximum output 128,000 tokens |
| Pricing | Standard, up to 272K input tokens: $8 per 1M input tokens; $0.40 per 1M cached input tokens; $40 per 1M output tokens |
| License | unknown |
| Availability | OpenAI API; ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users |
| Context and rates come from the official model reference; release and product access come from the announcement. No weights license is established by these sources. (source) (announcement) |
Cache writes cost $10 per 1M tokens for 30 minutes of retention or $20 per 1M tokens for 12 hours, at short-context Standard rates. Above 272K input tokens, input/cache rates become 2× and output 1.5× for the whole request. Batch is 50% of Standard; Fast 2×; Ultrafast 3× through 2027-02-28, then 6× from 2027-03-01, Pacific time. These rates do not measure total task cost. (source)
Release Date
Released 2026-11-17; the official X public excerpt independently corroborates availability on that UTC date, without verifying an exact posting time. (announcement) (source)
Benchmarks
OpenAI reports an improvement of 6.2 percentage points over Astra on its internal GDPval v2, using 22% fewer output tokens. These are vendor-reported results, not an independent reproduction. (announcement)
Astra reaches OpenAI's High AI Self-Improvement threshold; GPT-6.2 Sol remains below it. The card calls this an early safeguard-preparation milestone, not evidence of fully automated recursive self-improvement. Both models reach Critical for cybersecurity and High for biological/chemical capabilities. (source)
The card reports fewer improper workplace actions but more omissions of known blockers in model-only evaluations; the Codex developer prompt reduces those omissions. It also reports residual biological jailbreak regressions after refusal adjustments. A deployment safeguard is not proof that every underlying behavior improved. (source)
Use Cases
Text/image input and text output support coding, document work and computer use. Native audio/video are unsupported. The Responses API offers tools and a Multi-agent beta; reasoning effort none is rejected. (source)
Compared To
See GPT-6.2 Sol for the sibling model and Agent Runtime Containment for execution boundaries. API rates, experimental safeguards and benchmark results describe different properties and should be evaluated separately. (source) (source)