OpenAI announced GPT-6 Sol and GPT-6 Luna on September 22, 2026 (local time), describing them as fast, low-cost models trained with a method similar to GPT-6 Astra. They are available through ChatGPT Work, Codex, and the API. Sol is designed for complex coding and sequences of work handled by AI agents; Luna is aimed at processing large volumes of narrower tasks. Both models support a 1.05-million-token context window and up to 128,000 output tokens.

On OpenAI’s workflow evaluation, AutomationBench, GPT-6 Sol scored 33.2% with the “xhigh” reasoning setting. Claude Opus 5 scored 26.9% with “max.” Sol costs $0.27 per task, about 9% of Opus 5’s cost under those comparison conditions. A note says the cost for Claude Fable 5.1 excludes Claude Opus 5 usage for tasks handed off to it.

OpenAI’s internal evaluations found that GPT-6 Sol produced answers containing factual errors at about half the rate of the previous-generation GPT-5.6 Sol. The evaluation used anonymized ChatGPT conversations in which users had previously flagged factual errors, so it does not represent the factual error rate in typical use.

Standard API prices per 1 million tokens are $2.00 for GPT-6 Sol input, $10.00 for output, and $0.20 to read cached input. GPT-6 Luna costs $0.10 for input, $0.50 for output, and $0.01 to read cached input. Sol’s input and output rates are each half of GPT-5.6 Sol’s promotional prices of $4 and $20. These standard rates apply to shorter contexts. For requests with more than 272,000 input tokens, input and cached-input charges double, while output charges rise by 1.5 times. Batch, Flex, and fast modes have separate pricing.

OpenAI also improved prompt caching for the GPT-6 series that day. The system reuses processed instructions and context shared across API requests; discounts apply when a shared prompt prefix is reused within 30 minutes. Developers get a usage dashboard and diagnostic tools to investigate why a cache could not be reused. Existing caches can also be retained when the amount of processing devoted to reasoning changes.

GPT-6 Sol and Luna are rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. Free and Go users can access GPT-6 Luna in the desktop app, but regular Chat was not included at launch. The API model names are “gpt-6-sol” and “gpt-6-luna.” OpenAI said the rollout in ChatGPT would be gradual on launch day to maintain service stability.

OpenAI announced GPT-6 Sol and GPT-6 Luna on September 22, 2026 (local time), describing them as fast, low-cost models trained with a method similar to GPT-6 Astra. They are available through ChatGPT Work, Codex, and the API. Sol is designed for complex coding and sequences of work handled by AI agents; Luna is aimed at processing large volumes of narrower tasks. Both models support a 1.05-million-token context window and up to 128,000 output tokens.

On OpenAI’s workflow evaluation, AutomationBench, GPT-6 Sol scored 33.2% with the “xhigh” reasoning setting. Claude Opus 5 scored 26.9% with “max.” Sol costs $0.27 per task, about 9% of Opus 5’s cost under those comparison conditions. A note says the cost for Claude Fable 5.1 excludes Claude Opus 5 usage for tasks handed off to it.

OpenAI’s internal evaluations found that GPT-6 Sol produced answers containing factual errors at about half the rate of the previous-generation GPT-5.6 Sol. The evaluation used anonymized ChatGPT conversations in which users had previously flagged factual errors, so it does not represent the factual error rate in typical use.

Standard API prices per 1 million tokens are $2.00 for GPT-6 Sol input, $10.00 for output, and $0.20 to read cached input. GPT-6 Luna costs $0.10 for input, $0.50 for output, and $0.01 to read cached input. Sol’s input and output rates are each half of GPT-5.6 Sol’s promotional prices of $4 and $20. These standard rates apply to shorter contexts. For requests with more than 272,000 input tokens, input and cached-input charges double, while output charges rise by 1.5 times. Batch, Flex, and fast modes have separate pricing.

OpenAI also improved prompt caching for the GPT-6 series that day. The system reuses processed instructions and context shared across API requests; discounts apply when a shared prompt prefix is reused within 30 minutes. Developers get a usage dashboard and diagnostic tools to investigate why a cache could not be reused. Existing caches can also be retained when the amount of processing devoted to reasoning changes.

GPT-6 Sol and Luna are rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. Free and Go users can access GPT-6 Luna in the desktop app, but regular Chat was not included at launch. The API model names are “gpt-6-sol” and “gpt-6-luna.” OpenAI said the rollout in ChatGPT would be gradual on launch day to maintain service stability.