The brand new GPT-5.6 household: Luna, Terra, Sol
ninth July 2026
OpenAI’s newest flagship mannequin hit common availability this morning, and is available in three sizes: Luna, Terra, and Sol (from smallest to largest).
The brand new fashions are priced per 1M enter/output tokens as Luna $1/$6, Terra $2.50/$15, Sol $5/$30. For comparability, the Claude Opus collection are $5/$25 and the Claude Fable 5 is $10/$50, however price-per-million tokens doesn’t inform us a lot now that the variety of reasoning tokens can differ a lot between fashions for a similar activity.
All three fashions have a February sixteenth 2026 data cutoff, 1,000,000 token context window, and 128,000 most output tokens.
OpenAI’s largest benchmark declare issues long-running agentic efficiency, with one benchmark exhibiting all three fashions outperforming Claude Fable 5:
We educated GPT-5.6 to get extra helpful work from each token. On Brokers’ Final Examination, an analysis of long-running skilled workflows throughout 55 fields, GPT-5.6 Sol units a brand new excessive of 53.6, eclipsing Claude Fable 5 (adaptive reasoning) by 13.1 factors. Even at medium reasoning, it beats Fable 5 by 11.4 factors at roughly one-quarter the estimated value. That effectivity extends to smaller fashions, that are important to creating intelligence extra considerable and inexpensive: GPT-5.6 Terra and GPT-5.6 Luna outperform Fable 5 at round one-sixteenth the price.
Amusingly, one self-reported benchmark that Fable 5 crushed the GPT-5.6 household on was SWE-Bench Professional, the place Fable 5 received 80% in comparison with GPT-5.6 Sol getting 64.6%. This may increasingly assist clarify why OpenAI selected to publish this text yesterday particularly calling out SWE-Bench Professional for issues they discovered whereas auditing that benchmark:
In mild of those outcomes, we estimate that ~30% of SWE-bench Professional duties are damaged, and advise that mannequin builders rigorously study outcomes
I’ve had some early entry to GPT-5.6 Sol—it’s undoubtedly very competent, although to this point it hasn’t struck me as higher than Fable on the type of advanced coding duties I’ve been utilizing with Anthropic’s mannequin.
As common, the mannequin steerage for utilizing GPT-5.6 has probably the most attention-grabbing particulars. There are a bunch of latest API options that I have to discover (and doubtless add assist for in LLM), together with:
Programmatic Device Calling permits the fashions to “compose and run JavaScript that orchestrates device calls”—which sounds to me prefer it may assist bridge the hole between MCPs and full terminal classes that may compose CLI utilities in helpful methods. Additionally harking back to the dynamic filtering mechanism Anthropic added to their internet search device, which permits code execution towards internet outcomes as a part of a single mannequin flip.
Multi-agent lets the mannequin “spin up subagents for parallel, centered work”—the sub-agent sample now baked into the core API.
Immediate cache breakpoints brings the Claude mannequin of immediate caching to OpenAI, letting you be specific about the place the cache breakpoints are relatively than counting on the API to detect them robotically. Personally I a lot choose automated detection (nonetheless supported by OpenAI), however presumably there are optimization value financial savings available right here in the event you put the work in.
Now you can set element: authentic on picture requests to keep away from resizing the picture in any respect earlier than it’s processed.
Right here’s a full web page with 18 totally different pelicans—for reasoning efforts none, low, medium, excessive, xhigh, and max throughout the three totally different fashions. It additionally lists their token and calculated prices—the least costly was gpt-5.6-luna at effort none for 0.71 cents, the costliest was gpt-5.6-sol at max reasoning stage for 48.55 cents.
In additional pelican information, in the event you bounce to 17:50 of their livestream from this morning you’ll see OpenAI’s personal demo of 3D pelicans driving a tricycle, a bicycle, a pony, and one other pelican!
