OpenAI introduces GPT-6 Sol and Luna at lower API prices
OpenAI has launched GPT-6 Sol and GPT-6 Luna, two new models positioned below GPT-6 Astra for work that requires lower cost and faster deployment. The company reduced API prices by 50% compared with GPT-5.6 promotional pricing: GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens, while GPT-6 Luna costs $0.10 for input and $0.50 for output.
GPT-6 Sol and Luna are available from today in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Luna is also available to Free and Go users in the desktop app. In the API, the models are named gpt-6-sol and gpt-6-luna. OpenAI said the ChatGPT rollout will occur gradually during the day to maintain service stability.
Lower costs across professional and coding work
OpenAI says the new models use methods similar to those behind GPT-6 Astra, extending its improvements in professional work, factuality, coding, computer use and alignment to less expensive tiers. Astra remains OpenAI’s highest-performing model overall, while Sol and Luna are intended for workloads with different budgets and operating scales.
On AutomationBench, OpenAI reported that GPT-6 Sol at xhigh effort scored 33.2% and cost $0.27 per task. The company compared this with 26.9% for Claude Opus 5 at maximum effort, at 11.1 times Sol’s per-task cost. GPT-6 Luna at high effort improved on its predecessor by 5.4 percentage points at 58% lower cost per task.
OpenAI also reported a 56.4% result for GPT-6 Sol at maximum effort on Agents’ Last Exam, describing it as above Claude Opus 5’s highest result in that evaluation at 60% lower cost per task. These are provider-reported evaluations conducted in research environments or through the API, where system prompts and available tools may differ from production ChatGPT.
Factuality, code and computer-use claims
For factuality, OpenAI said GPT-6 Sol makes about half as many mistakes as its predecessor in an internal evaluation based on de-identified conversations where users had flagged factual errors. GPT-6 Luna, at higher effort levels, matched GPT-5.6 Sol at roughly one hundredth of its cost. OpenAI cautioned that these deliberately error-inducing conversations are not representative of typical use.
In software engineering, GPT-6 Sol reached 68.8% on DeepSWE v1.1, within 1.1 percentage points of the 69.9% result cited for Claude Fable 5 at xhigh effort. OpenAI said Luna scored 66.6% at maximum effort and cost 93% less per task than Opus 5 and 96% less than Fable 5 in the stated comparisons. On OSWorld 2.0 offline, Sol at xhigh effort reached 60.5%, compared with 60.3% for Claude Opus 5 at medium effort.
Prompt caching becomes part of the cost equation
The release also expands prompt-caching capabilities for agents and long conversations. OpenAI said cached input-token reads can receive discounts of 90%, while a dashboard and diagnostics tool help developers measure cache use and missed opportunities. Developers can now adjust reasoning effort and enable or disable tools without breaking reuse of earlier context; explicit breakpoints also let them choose where a cached prompt prefix ends.
GitHub reported that these caching improvements cut the share of prompt tokens needing fresh processing by more than 50% across billions of requests to OpenAI models. For organisations considering the new models, the access picture includes GPT-5.6 Luna free desktop access for desktop users, while the operational decision should assess workflow quality, reasoning settings, tool use and cache-hit rates alongside headline token prices.