Artificial Intelligence
OpenAI launches GPT-6 Sol and Luna
OpenAI said GPT-6 Sol and GPT-6 Luna were generally available, with claimed gains in accuracy, benchmark performance and pricing over GPT-5.6, alongside changes to response style and caching.
- OpenAI announced general availability of GPT-6 Sol and GPT-6 Luna, less than three months after GPT-5.6 Sol and Luna in July.
- The company said GPT-6 Sol made about half as many mistakes as GPT-5.6 Sol.
- OpenAI said GPT-6 Luna matched GPT-5.6 Sol at about one-hundredth of the cost at higher effort.
- OpenAI said it was cutting API prices for Sol and Luna by 50% against GPT-5.6 promotional pricing.
- The release said prompt caching could cost up to 90% less than evaluating an uncached prompt.
OpenAI announced GPT-6 Sol and GPT-6 Luna as generally available mainstream models. It said the release came less than three months after GPT-5.6 Sol and Luna, which it had launched in July.
The company said GPT-6 Sol cut the mistake rate to about half of GPT-5.6 Sol’s level. It also said GPT-6 Luna matched GPT-5.6 Sol at roughly one-hundredth of the cost when run at higher effort.
OpenAI compared the models with Anthropic’s Claude family across several benchmarks. On AutomationBench, it said Luna beat GPT-5.6 by 5.4% and cost 58% less per task, while Sol scored 6.3% better than Claude Opus 5 at 9% of the cost.
On Agents’ Last Exam, OpenAI said Sol achieved a slight performance lead over Claude Opus 5 and used 61% less per task. It also said Sol matched Claude Fable 5.1 xhigh on FrontierCode at lower cost, and came within 1.1 percentage points of Claude Fable 5 on DeepSWE 1.1.
OpenAI said GPT-6 Luna compared favourably with Claude Opus 5 and Claude Fable 5 on DeepSWE 1.1 at medium effort, with costs 93% and 96% lower respectively. It also said Luna overtook GPT-5.6 Sol on OSWorld at about 11% of the cost per task, and that prompt caching and some alignment measures had improved.
Named in this story
Companies
- OpenAI
- announced GPT-6 Sol and GPT-6 Luna
- Anthropic
- rival company whose models were compared with OpenAI’s
Products and systems
- GPT-6 Sol
- new mainstream model that OpenAI said made fewer mistakes
- GPT-6 Luna
- new mainstream model that OpenAI said matched earlier performance at lower cost
- GPT-5.6 Sol
- previous OpenAI model used for comparison
- Claude Opus 5
- Anthropic model used in benchmark comparisons
- Claude Fable 5.1 xhigh
- Anthropic model that OpenAI said GPT-6 Sol matched on coding at lower cost
- Claude Fable 5
- Anthropic model that OpenAI said GPT-6 Sol came close to on software engineering tasks
- AutomationBench
- benchmark used for professional work comparisons
- Agents’ Last Exam
- benchmark used for complex workflow comparisons
- FrontierCode
- benchmark used for coding comparisons
- DeepSWE 1.1
- benchmark used for software engineering comparisons
- OSWorld benchmark
- benchmark used for computer-use comparisons
How the source tells it
The register was opinionated and wry, with the author amplifying OpenAI’s claims while repeatedly undercutting them with ironic asides.
- hype loaded metaphors and repeated emphasis on large gains and lower costs
- scepticism ironic asides and rhetorical questions questioned the practical meaning of the claims
- urgency the short gap since GPT-5.6 and the focus on immediate price cuts gave the piece a rushed feel