OpenAI’s flagship reasoning model for complex professional work can implement changes across code modules, follow document templates, and analyze scientific data in multistep projects.
OpenAI
| Item | Specification |
|---|---|
| Input | Text, images |
| Output | Text |
| Context window | 1,050,000 tokens |
| Max input | 922,000 tokens |
| Max output | 128,000 tokens |
| Vision | ✓ |
| Function calling | ✓ |
OpenAI released GPT-6.2 Astra on November 17, 2026. It is a flagship reasoning model for complex coding, scientific research, and professional work that requires several rounds of verification.
Compared with GPT-6 Astra, this version improves code implementation and testing, professional deliverables, and precise interaction with graphical interfaces. OpenAI reports a gain of 5.2 points on Agents’ Last Exam, which evaluates multistep professional work, and 9 percentage points on Terminal-Bench-Science 0.1, which evaluates scientific work in a terminal.
| Feature | Benefit |
|---|---|
| Long context | Review specifications, code, and test records together to check dependencies across files. |
| Adjustable reasoning effort | Set different levels of reasoning for clearly defined and exploratory tasks within the same application. |
| Structured Outputs | Constrain successful responses to a JSON Schema so downstream software can read the result. |
| Code Interpreter | Run Python in a tool environment and review the code behind calculations. |
| Apply patch tool | Make precise file changes in an application-provided execution environment and review the resulting diff. |
| Streaming | Display generated content incrementally so users can start reading before the response is complete. |
Keep existing messages in order and append the change as a new message. The OpenAI API drops prior reasoning items when it detects edits to the conversation prefix, which can include earlier instructions or tool calls.
Use Astra for changes across modules, work that needs to adapt to test results, or projects whose acceptance criteria need clarification during implementation. OpenAI’s selection guide suggests starting at medium for complex projects, then using representative task results to decide whether to increase effort.
Define deliverables, tool permissions, and completion criteria, and preserve records for recovery. GPT-6.2 changes how edits to earlier conversation prefixes are handled: these edits can discard previous reasoning items, so review how your application maintains conversation history.
Query LinkModel’s visible catalog at GET /v1/models, which requires no Bearer API key. Verify the model ID and version in the response, then look up the operation in the LinkModel API Reference.
Look up the exact catalog ID in the LinkModel API Reference and check support for your current endpoint and required parameters. LinkModel documents the OpenAI-compatible POST /chat/completions and the Responses-compatible POST /responses separately, with model support specified in its documentation.
Tiered by input prompt tokens (incl. cache): once over the threshold, the whole request is billed at the higher tier.
| Token Type | Short context ≤272K | Long context >272K |
|---|---|---|
| Input | $6.00$8-25% | $12.00$16-25% |
| Output | $30.00$40-25% | $45.00$60-25% |
| Cached input | $0.300$0.4-25% | $0.600$0.8-25% |
| Cache write | $7.50$10-25% | $15.00$20-25% |
| Cache write (12 hours) | $15.00$20-25% | $30.00$40-25% |