OpenAI halves GPT-6 Sol and Luna API prices, and says Sol beats Opus 5 at 9% of the cost
OpenAI has cut the API price of GPT-6 Sol to $2 per million input tokens and $10 per million output, and GPT-6 Luna to $0.10 and $0.50. On its own run of AutomationBench, OpenAI says Sol at its highest effort setting scores 33.2% at $0.27 per task, against Claude Opus 5 at 26.9% and 11.1 times Sol's cost per task.
OpenAI credits caching and inference work for the lower prices, and says GitHub measured a drop of more than 50% in prompt tokens needing fresh processing. Image: OpenAI
OpenAI calls that a 50% cut, measured against GPT-5.6's promotional pricing, which is its own wording. Both models went live on September 22 and are called gpt-6-sol and gpt-6-luna in the API.
Every benchmark number here is OpenAI's own run
OpenAI ran the tests itself and took competitor scores "from publicly available reports", using Claude Fable 5 figures where Fable 5.1 numbers were unavailable. The cost-per-task column is the honest part of the comparison, because it is the only one matched on price rather than capability.
Claude Fable 5.1 with an Opus 5 fallback scores 31.4% on AutomationBench 1.0.6, below Sol's 33.2% and at more than 8.9 times the cost per task. OpenAI notes that figure understates Fable's real cost, because it leaves out the Opus 5 fallbacks, which fired on about 40% of tasks.
On DeepSWE v1.1, Sol at maximum effort scores 68.8%, within 1.1 points of Claude Fable 5's best of 69.9%, at roughly 80% lower cost per task. Luna scores 66.6% there, which OpenAI puts at 93% less per task than Opus 5 and 96% less than Fable 5.
On the offline set of OSWorld 2.0, Sol at its highest effort lands at 60.5% against Opus 5 at medium effort on 60.3%, again at about 80% lower cost per task. On Agents' Last Exam V1, which covers 55 sub-industries, Sol at maximum effort scores 56.4%, which OpenAI says beats Opus 5's best score in that evaluation at 60% lower cost per task.
Claude Opus 5.5, which undercut Opus 5 at $4 per million tokens on the same day, does not appear in any of the tables. Anthropic grades its own models the same way, and scored Opus 5.5 at 66.4% on Terminal-Bench on a test it ran itself.
ChatGPT Chat does not get the new models
Sol and Luna arrived in ChatGPT Work and Codex on September 22 for Plus, Pro, Business, Enterprise and Edu users, and Free and Go users get Luna in the desktop app. OpenAI says they "are not yet available in Chat", so the model behind everyday ChatGPT conversations has not changed.
OpenAI also says Sol makes about half as many mistakes as its predecessor on an internal factuality evaluation built from de-identified conversations where users had flagged an error. Both models inherit the communication style OpenAI introduced with Astra: "more clarity, less jargon, fewer odd turns of phrase, fewer low-value details".
Developers on much older endpoints have a separate deadline. The models gpt-3.5-turbo-instruct, babbage-002, davinci-002 and gpt-3.5-turbo-1106 shut down on September 28, with gpt-5.6-terra as the recommended replacement, a retirement OpenAI announced on September 26, 2025.
One number OpenAI volunteers puts the price cut in context: valued at API prices, daily token use inside the company has passed $600 for the median researcher and $7,000 at the 90th percentile. Sol is not the top of the range either. GPT-6 Astra at low effort scores 30.3% on AutomationBench, below Sol, at 3.9 times the cost per task, and OpenAI still rates Astra its best model for computer use.
