GPT-6 Sol vs Luna vs Astra: Which Model Should You Use?

OpenAI's GPT-6 Family Moves Down the Price Ladder

OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, filling out the GPT-6 family that began with Astra earlier this month. OpenAI says both were trained with methods similar to Astra's, and both cost half the promotional API rates of their GPT-5.6 predecessors. The useful questions are where each fits and what changes in your API calls.

• Pricing: GPT-6 Sol costs $2 input and $10 output per million tokens; GPT-6 Luna costs $0.10 and $0.50, cuts of 50% or more from GPT-5.6's promotional rates
• Context: both offer a 1.05M-token window (922K input, 128K output) with reasoning effort from none to max
• Availability: the API as gpt-6-sol and gpt-6-luna, ChatGPT Work and Codex on paid plans, Luna for Free and Go users in the desktop app, and GitHub Copilot
• Caching: cached input reads are discounted 90%, and reasoning effort can change mid-conversation without breaking the cache

GPT-6 Sol vs Luna vs Astra: What Does Each Cost?

Prices are per million tokens at standard processing. Past 272K tokens, input and cache rates double and output rises 1.5x; Batch and Flex cost half.

Sol costs one fifth of Astra and Luna one twentieth of Sol, a hundredfold spread from top to bottom. That spacing rewards routing each request to the cheapest model that handles it well. An OpenAI spokesperson told VentureBeat the new rates are permanent rather than introductory.

What OpenAI's Benchmarks Show for Sol and Luna

These figures are OpenAI's own; none of the launch coverage we reviewed included an independent evaluation.

On factuality, OpenAI says Sol makes about half as many mistakes as GPT-5.6 Sol on an internal test built from de-identified ChatGPT conversations where users had flagged errors, which OpenAI notes are not representative of typical use. It adds that Luna at higher effort matches GPT-5.6 Sol's factuality at about a hundredth of the cost.

For agentic work, OpenAI reports Sol at 68.8% on DeepSWE v1.1 at max effort, with Luna close behind at 66.6%. On AutomationBench 1.0.6, which tests business workflows across 47 tools, Sol at xhigh effort scored 33.2% at $0.27 per task, edging out Astra at low effort (30.3%). Both models also made fewer misleading claims about their coding work than their GPT-5.6 counterparts in OpenAI's alignment tests.

Astra remains OpenAI's pick for the most demanding projects; our GPT-6 Astra pricing breakdown covers where its computer-use lead earns the premium.

How Do You Migrate API Calls to GPT-6?

The OpenAI model guidance boils migration down to a short checklist:

• Reasoning effort: use reasoning.effort in the Responses API or reasoning_effort in Chat Completions. Sol and Luna accept none; Astra does not, so use low. Requests using minimal should start at low
• Tool calling: use the Responses API. Sol and Luna support function calling in Chat Completions only with reasoning_effort: "none"
• Sampling parameters: when reasoning effort is anything other than none, remove temperature, top_p and top_logprobs
• Prompt caching: teams moving from GPT-5.5 or earlier should replace prompt_cache_retention with prompt_cache_options.ttl set to "30m"

Codex users can hand the job to the OpenAI Docs skill with $openai-docs migrate this project to the GPT-6 model family.

Async Tools and Mid-Turn Steering for GPT-6 Agents

Three API features change agent design. Async tool calling lets the model keep working while a slow tool runs: set async: true on a function or custom tool and return the result later with its original call_id. Mid-turn steering accepts corrections while the model works, and over a WebSocket connection the Responses API keeps the work already completed. Configuration updates raise reasoning effort for a hard step, or lower it for a routine follow-up, through a configuration_update input item that leaves the cached prompt prefix intact.

OpenAI says toggling tools now preserves the cache too, and GitHub saw the caching work cut the share of prompt tokens needing fresh processing by more than 50%.

Where GPT-6 Sol and Luna Are Available Today

In ChatGPT, both are rolling out gradually to Work and Codex for Plus, Pro, Business, Enterprise and Edu users; they are not in regular Chat yet. GitHub made both generally available in Copilot the same day, with Sol on Pro+, Max, Business and Enterprise plans and Luna on those plus Pro.

My working recommendation: default to Sol, route high-volume extraction and summarization to Luna, and reserve Astra for tasks where its extra depth is measurably worth five times the price. Anthropic also shipped Claude Opus 5.5 the same day, so there are two fresh defaults worth testing on your own workloads. Follow more model releases in our artificial intelligence coverage.

Sources: OpenAI — September 22, 2026; OpenAI API model guidance — September 22, 2026; TechCrunch — September 22, 2026; VentureBeat — September 22, 2026; GitHub Changelog — September 22, 2026.