GPT-6.2 Astra and Sol: what changes for API migration
GPT-6.2 changes conversation history and instruction handling. What IT teams should check about tool calls, caching and costs before migrating.
CogLake Redaktion · 19 November 2026 · 4 minThis article was automatically generated with the assistance of AI.
For enterprises, GPT-6.2 is more than a new model ID. OpenAI announced Astra and Sol in the API on November 17, 2026. The Responses API also changes how instructions are placed and how internal reasoning items are handled after edits to earlier conversation history. IT teams should therefore validate multi-turn workflows, not merely replace a model name. As of and sources accessed: November 18, 2026. This article is based on vendor documentation, not our own performance tests.
What is new—and what is not?
The API changelog lists both models for Responses and Chat Completions. Tool calling, however, requires Responses. The Astra model page and Sol model page document text and image input with text output, not native audio or video processing.
OpenAI positions Astra for demanding projects and Sol for everyday work and agents at scale. That is vendor positioning, not proof of which model will perform better in your process. The earlier ChatGPT rollout of Sol and Luna is a separate event from this API migration.
Why edits to earlier history matter
A “prefix” is the conversation history before a particular reasoning item, including earlier instructions, user messages and tool results. According to the migration guide, starting with GPT-6.2 the API drops subsequent reasoning items after edits to that history. Certain model changes and changes to the end user or organization can also trigger loss of reasoning items. The model may repeat earlier work; token usage can increase and answer quality can decline.
The practical lesson is not to prohibit changes or access revocation. Instead, preserve historical content where possible and append new, valid instructions. Security and permission requirements still take priority over cache reuse.
| Common integration change | Documented approach to evaluate |
|---|---|
| Overwriting an earlier developer instruction | Append a new developer message rather than rewriting the earlier one |
| Shortening a past tool result | Limit it before first submission, or evaluate an appropriate compaction path |
| Replacing tools between turns | Review documented additive tool updates or tool-selection restrictions |
| Replacing history with a custom summary | Explicitly evaluate the prefix change and removed reasoning items |
The prefix guide distinguishes OpenAI compaction from home-grown summaries. It describes an exception preserving subsequent reasoning items for OpenAI compaction. That is not a guarantee of unchanged cache hit rates.
Separate instructions, parameters and pricing
Starting with GPT-6.2, Responses appends top-level instructions after input. OpenAI recommends putting reusable, cache-eligible instructions in the first developer message in input, and using instructions for request-scoped guidance. Caching does not replace correct instruction hierarchy.
The migration guide recommends starting a new integration at medium reasoning effort and replacing previous none or minimal settings with low. The model pages confirm that none is rejected. Also check the guide's unsupported sampling and logprob parameters rather than replaying old requests unchanged.
| USD per million tokens, Standard short context | Astra | Sol |
|---|---|---|
| Input | 8 | 2 |
| Cached input | 0.40 | 0.10 |
| Cache writes, 30 minutes | 10 | 2.50 |
| Output | 40 | 10 |
Source: both model pages, accessed November 18, 2026. For requests exceeding 272,000 input tokens, the documentation doubles input/cache rates and multiplies output rates by 1.5 for the full request relative to short-context rates. The 1,050,000-token context window is not the same as the 922,000-token maximum input or 128,000-token maximum output. Account for processing tiers and cache-write costs separately; list prices alone do not establish cost per completed task.
Acceptance testing beyond model quality
Editorial recommendation: Compare a representative multi-turn workflow before and after migration. Use authorized or synthetic data and constrain external write actions.
- Business outcome: Are source references, structured fields and necessary clarification questions correct?
- History changes: Evaluate new instructions, tool changes and compaction separately. The prefix guide says
input_transformationsidentifies removed reasoning items and the reasons for removal. - Cost and duration: Record total tokens, cache reads/writes and execution time per accepted business result. Cache diagnostics do not prove reasoning reuse.
- Tool controls: Test permission revocation and approvals independently of caching. A successful answer does not prove an external action was executed correctly.
What does this mean for CogLake?
CogLake documents OpenAI BYOK and central model approvals. This does not establish GPT-6.2 availability in every account or support for every new API feature. Check the model catalog, connection type and permissions in your installation. For a custom integration, the migration guide remains the relevant technical reference: provider access, a functioning API integration and an accepted business workflow are three separate checks.