GPT-5.6 is rolling out in the API. Sol is our flagship model, leading in coding, knowledge work, cybersecurity, and science. Terra delivers performance competitive with GPT-5.5 at lower cost. Luna is our fastest, most affordable model for high-volume tasks.

GPT-5.6 is built for greater token efficiency. On the Artificial Analysis Coding Agent Index, Sol with max reasoning sets a new high of 80 while using 54% fewer output tokens and completing tasks in 57% less time than the next-highest-scoring model. Across knowledge-work evaluations, Terra surpasses GPT-5.5 at lower cost, while Luna nearly matches GPT-5.5’s peak performance at well under half the estimated API cost.

Programmatic Tool Calling lets GPT-5.6 write and run JavaScript to coordinate complex tool workflows, process results, and decide what to do next. That means fewer model round trips, fewer tokens, and less orchestration for developers. Programs run in isolated hosted V8 runtimes, so developers get stateless code-driven orchestration without managing containers.

Multi-agent lets GPT‑5.6 spawn concurrent subagents within a single request. They can divide complex work, explore different approaches in parallel, and synthesize their findings into one result. This feature is initially available in beta via the Responses API.

Explicit prompt caching gives developers more control over cache boundaries, with explicit cache breakpoints and a 30-minute minimum cache life. Starting with GPT-5.6, both explicit and implicit cache writes are billed at 1.25x the uncached input rate, while cache reads retain the 90% discount.

GPT-5.6 is our most capable model yet on cyber and bio-related tasks. Some API calls may be blocked or paused mid-stream while our safety systems conduct additional review, particularly in dual-use areas such as cybersecurity where defensive and offensive activity can initially look similar. We’ll keep tuning safeguards as we learn from real-world use.

GPT‑5.6 Sol, Terra, and Luna are rolling out in the OpenAI API. Programmatic Tool Calling is available in the Responses API, with Multi-agent initially available in beta.

The only flagship you have is 4o #keep4o #4oForAll

so to sum up it's good at only coding but not even as good as fable5 ? 🙂

#BringBack4o #keep4o #OpenSource4o

i plugged sol into @opendottrade the second it came out. my spare mac is finally employed

Not interested! Bring Back GPT-4o✨️ @OpenAI #keep4o #BringBack4o #chatgpt

#keep4o #OpenSource4o

Where is Sol ?

Learn when to use which model before you start using it

watch the full breakdown of GPT 5.6

Return these excellent models. #Keep4o #Keep51 #Keep45 #Keep41 #keepo3

Why only Terra and Luna in Codex 0.144? Where is Sol?

@OpenAI @OpenAIDevs Brazil is ready. 🇧🇷 Excited for GPT-5.6 and the future of AI. As a builder from Brazil, I can’t wait to create new solutions, businesses, and opportunities with Sol, Terra, and Luna. The Brazilian community is ready to build. 🚀

Codex Pets

@OpenAIDevs Where is Sol?

Can @OpenAIDevs confirm when it will roll out completely for Indian users!

So which app do I use now? ChatGPT, Codex or ChatGPT Work... bad enough keeping up with Model names!

No sol for plus users?

5hours limits decreased so they take 15% of the weekly limit now. Used to be 20% ... That sucks... Sol Ultra cannot even finish 1 prompt with that, not even on standard mode. With Fast mode it's just a joke :S ... And tokens are much more expensive than with using fable :S

Blind auf "lower cost" schauen ist die klassische Falle. Wer kontrolliert die Infra? Wo laufen die Modelle? Terra klingt perfekt — bis die Prompts komplexer werden und Sol nötig ist. Dann sitzt du fest.

Why only Terra and Luna in Codex 0.144? Where is Sol?

who is checking? └ This request requires additional safety checks, which can take extra time. Hang tight or retry with a faster model for a quicker response, though it may be less capable of handling complex requests.

With the Multi-agent API allowing concurrent subagents, how do we prevent "agent loop inflation" where they burn tokens on dead ends? Can we set explicit budget caps per subagent, or a shared context window limit for parallel paths before synthesis?

This looks like a meaningful update. Better performance, more flexibility, and lower costs should make it easier to build AI applications at scale. Excited to see the benchmarks.

The three-lane split reads clean: frontier capability, lower-cost throughput, and high-volume speed. Useful shape for agent routing, where one model rarely needs to do every job.

Exciting direction. Performance matters, but giving developers the right model for the right job matters even more. Looking forward to seeing what people build with this lineup.

luna is the interesting one tbh most apps don’t need genius, they need cheap boring calls that don’t fall over

GPT-5.6 with three sub-models less than a week after everyone's still processing Fable 5 pricing. The timing is not an accident.