The GPT-6 Family: Astra, GPT-6.1 Sol, GPT-6.1 Terra, and Luna

The GPT-5.6 family, covered in previous lessons, remained the standard for only two months. Starting September 3, 2026, OpenAI rolled out the next generation, GPT-6, in four phases. This lesson introduces the new models, their pricing, and what changes for your code.

The timeline

Which model for what task

OpenAI’s models page sums up the choice in three sentences: if you aren’t sure where to start, pick GPT-6 Astra, the flagship model for complex reasoning and code; pick GPT-6.1 Sol to balance intelligence and cost; pick GPT-6 Luna for high-volume, cost-sensitive workloads. GPT-6.1 Terra rounds out the lineup: a fast, efficient model for visual reasoning, precise computer control, and clear, natural writing.

All four models offer a 1,050,000-token context window and up to 128,000 output tokens. Beyond 272,000 input tokens, the entire request is billed at double the input rate and 1.5× the output rate. Batch processing and Flex processing cost half price, while Fast mode doubles the rate. Ultrafast mode, available on Astra since September 29 and on GPT-6.1 Sol since October 8, drastically reduces the time between generated tokens for six times the price. GPT-6 Sol, released September 22 at the same price as GPT-6.1 Sol, remains in the catalog; the documentation invites its users to review the migration guidelines before switching to GPT-6.1 Sol.

What changes for your code

Four points in OpenAI’s migration guide require special attention.

The none effort level disappears on the two most capable models. GPT-6 Astra and GPT-6.1 Sol do not accept reasoning.effort: "none": on Astra, the request returns an HTTP 400 error. GPT-6.1 Terra, GPT-6 Sol, and GPT-6 Luna do accept it. If your GPT-5.6 code used none for speed, switch to low or change models; if you used minimal, start with low and compare.

Sampling parameters are removed whenever the model reasons. When the effort is not none, remove temperature, top_p, and top_logprobs. In Chat Completions, also remove logprobs; in the Responses API, remove message.output_text.logprobs from include.

Tools go through the Responses API. GPT-6 Astra, GPT-6.1 Sol, and GPT-6.1 Terra accept Chat Completions, but tool calling there requires the Responses API. GPT-6 Sol and GPT-6 Luna allow function calling in Chat Completions only with an effort of none. OpenAI’s recommendation is clear: use the Responses API to reason with tools.

Caching is configured differently. If you are coming from GPT-5.5 or earlier, replace prompt_cache_retention with prompt_cache_options.ttl, set to 30m. Since October 20, a 12-hour duration is also available, with writes billed at 2.5× the input price.

from openai import OpenAI

client = OpenAI()

response = client.responses.create(
    model="gpt-6.1-sol",
    reasoning={"effort": "medium"},
    input="Summarize the legal risks of this service contract in five points.",
)
print(response.output_text)

New features accompanying GPT-6

On September 3, OpenAI added three controls to the Responses API designed for long-running work with Astra. Asynchronous tool calling lets the model continue working while your application executes a tool. Mid-response steering allows you to send additional instructions via WebSocket while the response is being built. Mid-conversation effort changes, using a configuration_update element, increase or decrease effort without rewriting the start of the prompt, thus preserving the cached prefix. On September 29, GPT-6.1 Sol received beta support for delegating to sub-agents within a single request. Finally, Astra includes asynchronous monitoring for misalignment behaviors during agent work, which can trigger security alerts or interrupt a conversation for review.

What about GPT-5.6?

The GPT-5.6 family remains in the catalog and is not deprecated; it even serves as the official replacement for models retired on December 11. OpenAI has lowered its prices: Luna by 80% and Terra by 20% on July 30, then GPT-5.6 Sol to $4 / $20 on August 21, a promotional rate announced through November 30, 2026. Nothing forces you to migrate immediately. However, GPT-6.1 Sol, presented as close to Astra, costs half as much as promotional GPT-5.6 Sol ($2 / $10 vs. $4 / $20): any new project should start with the GPT-6 family.

Your turn

Prepare to switch a request. Take a GPT-5.6 request from your code and list what needs to change to send it to GPT-6.1 Sol: the effort level (none or minimal?), sampling parameters, tools called via Chat Completions, and cache settings. Then run both versions on about ten real-world cases and compare quality, latency, and cost.

Official sources: Models · Using GPT-6 · Pricing · Changelog (OpenAI API documentation)

Key takeaways

• GPT-6 Astra (September 3), GPT-6 Sol and Luna (September 22), GPT-6.1 Sol (September 29), and GPT-6.1 Terra (October 26) form the current family
• Starting point: Astra for the most demanding tasks, GPT-6.1 Sol for balance, Luna for volume, GPT-6.1 Terra for speed and visuals
• Astra and GPT-6.1 Sol reject the none effort level; remove temperature, top_p, and log probabilities whenever the model reasons
• Program tools using the Responses API, and configure caching with prompt_cache_options.ttl