The new GPT-5.6 family: Luna, Terra, Sol
Simon Willison · Simon Willison · 2026-07-09
Simon Willison reviews the GPT-5.6 model family launch, noting competitive pricing and new API features while flagging OpenAI's suspicious timing in publishing a SWE-Bench Pro audit that undermines the benchmark on which Claude Fable 5 most clearly outperforms GPT-5.6.
Appears in
Extraction
Topics: gpt-5.6llm-benchmarksllm-pricingmodel-releasebenchmark-integrity
Claims
- GPT-5.6 is priced at Luna $1/$6, Terra $2.50/$15, and Sol $5/$30 per million tokens, substantially undercutting Claude Fable 5 at $10/$50.
- All three GPT-5.6 models share a February 16, 2026 knowledge cutoff and a 1-million-token context window with 128,000 maximum output tokens.
- OpenAI published a SWE-Bench Pro audit estimating ~30% of tasks are broken on the same day as the GPT-5.6 launch, in which Fable 5 scored 80% versus Sol's 64.6%.
- New API features include Programmatic Tool Calling (in-context JavaScript orchestration), multi-agent support, and explicit prompt cache breakpoints.
- Willison's own early access testing found GPT-5.6 Sol competent but not clearly superior to Fable 5 for complex coding tasks.
Key quotes
Amusingly, one self-reported benchmark that Fable 5 crushed the GPT-5.6 family on was SWE-Bench Pro, where Fable 5 got 80% compared to GPT-5.6 Sol getting 64.6%. This may help explain why OpenAI chose to publish this article yesterday specifically calling out SWE-Bench Pro for problems they found while auditing that benchmark.
I've had some early access to GPT-5.6 Sol - it's definitely very competent, though so far it hasn't struck me as better than Fable at the kind of complex coding tasks I've been using with Anthropic's model.
price-per-million tokens doesn't tell us much now that the number of reasoning tokens can differ so much between models for the same task.