OpenAI has released a guide to the new GPT-6 model family. It explains when to choose the most powerful Astra, when the cheaper GPT-6.1 Sol is sufficient, and where the fast Luna is best suited. It also advises how to formulate prompts correctly, how to let the models work for hours, and how to optimize costs in the process. For most developers and companies, GPT-6.1 Sol will likely be the default choice, as its performance comes close to Astra while costing only one-fifth as much.
Three models for three types of work
GPT-6 Astra is the most intelligent model OpenAI currently offers and is suitable for the most demanding tasks. According to the company, it achieved better results in a range of tests with significantly fewer output tokens. As a result, a single task via the API ultimately costs less than with older models, even though each individual token is more expensive. The pricing lists $10 for input, $50 for output, and $1 for cached input.
GPT-6.1 Sol, introduced at the end of September, offers intelligence almost on par with Astra at one-fifth of the price. Input costs $2, output $10, and cached input ten cents. This model is aimed at complex programming, research, and advanced computer control. OpenAI recommends testing both Sol and Astra on your own tasks before deciding whether the difference in quality justifies the higher cost.
The cheapest model, GPT-6 Luna, costs 10 cents for input and 50 cents for output. It is designed for narrowly defined, high-volume work, such as extracting data from invoices, categorizing requests, or writing structured summaries.
How to enable and configure Sol
In a Responses API request, simply set the model parameter to gpt-6.1-sol. The older Chat Completions interface supports this model only for queries that do not involve external tools. As soon as the model needs to call functions, the request must go through the new Responses API.
The amount of reasoning is controlled by the reasoning.effort parameter, which offers five levels: low, medium, high, xhigh, and max. The default is medium, while Sol, like Astra, does not support the none and minimal levels. A low level is sufficient for looking up facts or making minor text edits. The medium level is suitable for planning features or comparing options, while the high level is useful for demanding debugging and deeper analysis. It makes sense to try the highest levels only when the high level is insufficient, and they should remain enabled only when the better result offsets the longer wait and higher cost.
The amount of internal reasoning can now also be changed during a conversation. The developer inserts a configuration update item into the input, allowing the model to increase its reasoning intensity for a complex section and reduce it again for routine queries. The beginning of the prompt remains unchanged, so the cache is not cleared. In the Codex programming tool, OpenAI recommends starting with the model's default setting and adjusting it as needed.
Fast and ultra-fast modes
When a quick response matters, such as in a chat application, the API offers a fast mode with quicker and more consistent responses at a higher price per token. Ultra-fast mode accelerates token generation itself, regardless of the configured reasoning level. OpenAI currently lists it for the Astra model and recommends it for cases where paying extra for speed is worthwhile.
European companies may encounter limitations, however. Fast mode is unavailable for Astra, GPT-6 Sol, and Luna if data must remain within the European Union for legal reasons. Ultra-fast mode works exclusively with data stored in the United States or processed globally.
Fewer instructions and better results
Eric Provencher from OpenAI's developer experience team notes that current models understand nuances and ambiguities much better. Overly detailed instructions may therefore do more harm than good, even though they helped with older versions.
According to OpenAI, a properly formulated prompt clearly defines the desired outcome, the target user, applicable constraints, and the criteria for completing the work. System skills should briefly describe the conditions under which they are triggered, while the AGENTS.md file should explicitly authorize safe routine steps, such as local tests without access to production. Blanket rules requiring constant questions should be replaced by a clear boundary between actions the model may take autonomously and situations that require human approval.
Astra stays on track better during long-running tasks, but it asks questions more often in situations where older models would simply assume the next step. The process can therefore sometimes stop even though the user expected it to continue smoothly. OpenAI consequently provides sample instructions telling the model to infer intent from context and complete the work independently. Phrases expressing possibility should be understood by the system as direct instructions to act, and it should request approval only after producing a finished result that can be reviewed. These practices are based on Astra's behavior, but OpenAI also recommends trying them with Sol.
The new models also respond more sensitively to instructions hidden in skills and configuration files, so an unclear or contradictory instruction can unnecessarily block the work. A suitable solution is to state that direct user instructions take precedence over general skills. Astra also tends to produce long responses full of lists and tables, often repeating the same phrases. Users who require continuous prose in simple language must request it explicitly. In programming, OpenAI recommends restraining the model's testing behavior because it generates unnecessarily extensive test suites even for minor code changes.
Tasks lasting hours or even days
According to OpenAI, GPT-6 models can be assigned tasks with processing times measured in hours or even days. Ongoing steering makes it possible to send the model clarifications while it is working through a version of the Responses API built on the WebSocket protocol. The message is placed in a queue and does not stop running tools or invalidate work that has already been completed. Asynchronous tool calls then allow the model to continue with another activity while the application runs tests, for example. Simply set the function's asynchronous mode to true and return the result with the original call_id identifier.
GPT-6.1 Sol can also divide independent tasks among subagents, each of which can examine a different part of the code, and then combine their findings into a single response. Multi-agent support is currently in beta, and OpenAI recommends defining in the prompt when and to what extent the model should divide the work.
A model that clicks for you
All three models can directly control a computer and work with websites and applications, including those without their own API. The model can therefore fix code, open the product in a browser, and verify that the fix works. Nevertheless, OpenAI recommends choosing the simplest reliable approach. If a task can be handled by a standard API or a connected tool, that option should take priority. Computer control should be used only when the model needs to read the screen, click, or fill out forms. Developers can integrate it into their own applications through a code execution tool, use Playwright for browsers, and use the PyAutoGUI interface for desktop applications.
How to save money in production
The company recommends removing everything from the context that the task does not immediately require and running independent tasks concurrently. Prompt caching delivers the greatest savings because cached input tokens are up to 95 percent cheaper. Fixed instructions and background materials should therefore be placed at the beginning of the prompt, followed by variable details. Cache writes and long-context surcharges should also be included in the budget. For extensive conversations, compaction helps by compressing the context and retaining only the data needed to continue. Before production deployment, companies should accurately measure the success rate, response speed, and cost per completed task using representative tasks.
Migrating from older models
Codex itself can migrate an entire project using the OpenAI Docs skill. Developers modifying code manually should move from the none or minimal levels to low and then compare the results. When internal reasoning is enabled, the temperature, top_p, and top_logprobs parameters must be removed from the request. Users migrating from GPT-5.5 or older models should replace the prompt_cache_retention item with the new prompt_cache_options.ttl setting with a time value of 30 minutes.
How companies use the models
OpenAI demonstrates practical applications through customers already building on the Astra model. The legal platform Harvey combines case law, law firm documents, and the preferences of a specific lawyer to prepare complex drafts of contracts and legal filings. Cognition deployed Astra in its autonomous agent Devin. While testing an iPhone game, it analyzed a recording from the simulator and clearly separated successfully verified steps from areas that had not yet been tested in its final report. With the help of this model, the analytics platform Hex transforms questions about sales channel performance into written conclusions and interactive dashboards. Invideo uses the system to plan edits on a video's timeline and create custom effects. In this context, the company reports roughly three times the success rate for color adjustments and corrections, while several of its video editors were able to create around fifty new effects in a single day.
Sources: openai.com and openai.com
