OpenAI announced two new members of the GPT‑6 family: GPT‑6 Sol and GPT‑6 Luna. The company said both models were trained using similar methods to GPT‑6 Astra and are intended to provide improved cost efficiency while carrying forward many of Astra’s advances in professional work, factuality, coding, computer use, and alignment.
Performance and benchmark claims
OpenAI published benchmark results that compare Sol and Luna to prior models and competitors. On AutomationBench, OpenAI reported GPT‑6 Sol (xhigh effort) scored 33.2% at an estimated cost per task of $0.27, while GPT‑6 Astra (low) scored 30.3% at roughly 3.9× the cost. The company also said Sol outperformed Claude Opus 5 at far lower cost and exceeded Claude Fable 5.1 in several comparisons, though it noted missing fallback costs for some competitors.
On Agents’ Last Exam, OpenAI reported GPT‑6 Sol (max effort) scored 56.4% and ran at about 60% lower cost per task than Claude Opus 5. In coding evaluations, OpenAI said GPT‑6 Sol achieved 68.8% on DeepSWE v1.1, within 1.1 percentage points of Claude Fable 5’s top score, at about 80% lower cost per task; GPT‑6 Luna scored 66.6% and was described as comparable to Claude Opus 5 and Fable 5 at lower cost.
Factuality, alignment and system improvements
OpenAI said its internal factuality evaluation showed GPT‑6 Sol makes about half as many mistakes as its predecessor in selected error-inducing conversations, and that both Sol and Luna show alignment improvements versus GPT‑5.6, including lower rates of misleading claims in coding tasks. The company cautioned these evaluations target challenging scenarios and do not represent typical failure rates.
OpenAI also described improvements to prompt caching for GPT‑6 that it says increase cache hit rates and offer discounts of up to 90% on cached input‑token reads. GitHub was cited as reporting that these caching changes reduced the share of prompt tokens needing fresh processing by more than 50% across billions of requests.
OpenAI said GPT‑6 Sol and GPT‑6 Luna are available in ChatGPT Work and Codex starting today for Plus, Pro, Business, Enterprise, and Edu users; Free and Go users can access GPT‑6 Luna in the desktop app. The models are not yet available in Chat. In the OpenAI API, they are listed as gpt-6-sol and gpt-6-luna, and OpenAI said it will roll the models out in ChatGPT gradually throughout the day.
Original source: OpenAI News