October 2, 2026
Codex can handle the move to GPT-6: OpenAI shares a project migration command
OpenAI's guide connects model selection with configuring reasoning, tools and skills for an application's workflow.
When moving to GPT-6 Astra or GPT-6.1 Sol, replace the previous `reasoning.effort: none` with `low`. GPT-6 Luna supports `none`.
OpenAI has put together a guide for startups integrating GPT-6 into their applications. It covers model selection, improving prompts and skills, coordinating tools and preparing for use by customers.
Migration through Codex. The command `$openai-docs migrate this project to the GPT-6 model family` instructs the agent to migrate the project to the new family. For manual integration, specify one of these models in the `model` field:
- `gpt-6-astra`; - `gpt-6.1-sol`; - `gpt-6-luna`.
Reasoning with tools requires the Responses API. Astra and Sol support 5 `reasoning.effort` levels: from `low` to `max`, including `medium`, `high` and `xhigh`.
How much the models cost to run
Standard pricing as of Oct 2, 2026, varies by model. All amounts in the table are per 1 million tokens.
| Model | Input | Output | Additional pricing details | | --- | --- | --- | --- | | GPT-6 Luna | $0.10 | $0.50 | The context window holds 1,050,000 tokens, with a maximum response of 128,000 tokens | | GPT-6.1 Sol | $2 | $10 | Cached input costs $0.10, cache writes cost $2.50 | | GPT-6 Astra | $10 | $50 | Batch and Flex cost half the Standard rate, Fast costs twice as much |
For inputs over 272,000 tokens, all 3 models charge higher rates for the entire request. Input and cache rates double, while the output rate increases by 1.5 times.
How to avoid waiting for tools
Independent work. With `async: true` in a function or custom tool definition, the model continues with other parts of the task while the application runs the tool. The application returns the result with the original `call_id` to associate it with the call.
GPT-6.1 Sol also supports multi-agent in beta. Enable it with `multi_agent.enabled: true` and `betas: ["responses_multi_agent=v1"]`. In the official PR review example, `max_concurrent_subagents: 3` allows up to 3 subagents.
You can change the reasoning level between responses through `configuration_update` in `input`, preserving the request-level `reasoning.effort` setting and cache prefix. This mode works with a single agent in standard and is incompatible with automatic context compaction and truncation.
What to improve in skills
Short skill descriptions help Codex select the right instructions. In a recommendation dated Sep 11, Eric Provencher suggests loading details as needed: when there are too many long descriptions, Codex truncates them. A skill serves as a recipe for the agent to load for a particular task.
Cognition is already using GPT-6 Astra for testing in Devin. In an iPhone game example, the agent returned a simulator recording and a report on the checks it ran.
Invideo reports roughly a 3-fold increase in color correction success rates. Several editors created around 50 effects in a day, according to OpenAI on Oct 2.
Primary sources: [OpenAI's guide to working with GPT-6](https://openai.com/index/practical-guide-building-gpt-6), [official OpenAI integration and migration documentation](https://developers.openai.com/api/docs/guides/latest-model).
In Devin's example, the test results include a recording of the game in action and a report on the checks performed.
Source
