GPT-5.6 family: OpenAI launches Luna, Terra, Sol with prices

OpenAI released the GPT-5.6 family to general availability on 9th July 2026, shipping three sizes named Luna, Terra, and Sol (from smallest to largest) and new API capabilities. The models are priced per 1M input/output tokens as Luna $1/$6, Terra $2.50/$15, and Sol $5/$30.

What did OpenAI release?

GPT-5.6 is a three-model family offered at distinct price points and reasoning-effort settings, with general availability announced on 9th July 2026. Luna is the smallest, Terra is medium, and Sol is the largest; OpenAI published example costs ranging from 0.71 cents (gpt-5.6-luna at effort none) up to 48.55 cents (gpt-5.6-sol at max reasoning level) for specific pelican-reasoning examples.

The company framed the family around getting "more useful work from every token," and packaged the release with a page of example outputs showing six reasoning effort levels across the three models and calculated token costs.

How does GPT-5.6 compare on benchmarks?

OpenAI highlighted long-running agent performance where GPT-5.6 Sol scored a new high of 53.6 on Agents' Last Exam, which OpenAI said eclipses Claude Fable 5 (adaptive reasoning) by 13.1 points. The company also noted that, at medium reasoning, GPT-5.6 beats Fable 5 at roughly one-quarter the estimated cost, and that Terra and Luna outperform Fable 5 at around one-sixteenth the cost.

Not every benchmark favours GPT-5.6. On SWE-Bench Pro the source shows Claude Fable 5 scored 80% while GPT-5.6 Sol scored 64.6% (the writeup reproduces the latter as "GUT-5.6 Sol getting 64.6%"). OpenAI circulated an audit of SWE-Bench Pro alongside the release, estimating that "~30% of SWE-bench Pro tasks are broken," and advised model developers to examine results carefully.

Simon Willison, who had early access to GPT-5.6 Sol, called the model "definitely very competent," while noting it has not yet struck him as better than Claude Fable 5 on the complex coding tasks he uses with Anthropic's model.

What new API features are included?

GPT-5.6 ships with several API features intended for tool-enabled and multi-agent workflows: Programmatic Tool Calling, Multi-agent, and prompt cache breakpoints, plus image handling tweaks. Programmatic Tool Calling lets the model compose and run JavaScript that orchestrates tool calls, effectively allowing code execution to control tool usage inside a single model turn. Multi-agent allows the model to "spin up subagents for parallel, focused work," embedding the sub-agent pattern directly in the API. Prompt cache breakpoints let developers explicitly mark where cached prompt state should break, an approach OpenAI added alongside continued support for automatic detection. The release also added a detail: original option for image requests to avoid resizing before processing.

OpenAI demonstrated the models with playful examples — a livestream included a demo at 17:50 showing 3D pelicans riding a tricycle, a bicycle, a pony, and another pelican — and an examples page lists per-task token and cost calculations across reasoning-effort settings.

Why it matters

The package combines cost tiers and explicit tooling primitives, signalling a push to make longer agentic workflows cheaper and more composable. The Agents' Last Exam claim (Sol at 53.6) focuses attention on multi-step professional workflows rather than single-turn benchmarks. At the same time, the SWE-Bench Pro discrepancy and OpenAI's audit note about broken tasks highlight that benchmark choice and hygiene still shape perceived model rankings.

What to watch

Monitor independent reproductions of Agents' Last Exam and SWE-Bench Pro runs, plus early developer uptake of Programmatic Tool Calling and the Multi-agent primitives. The next confirmatory signal will be community evaluations showing whether cost-per-work unit improvements OpenAI claims hold up across diverse, audited tasks.