GPT-5.6 (Sol, Terra, Luna)
GPT-5.6 splits into Sol, Terra and Luna tiers, with Sol scoring 53.6 on Agents' Last Exam and 92.2% on BrowseComp at lower token cost.
Tiered family after a limited preview from 2026-06-26. Adds Programmatic Tool Calling, explicit prompt-cache breakpoints, max reasoning, an "ultra" mode of four parallel agents and Multi-agent beta. Cyber safeguards block about ten times more harmful activity; ExploitBench 73.5% vs 47.9% for GPT-5.5.
- Date
- Thursday, 9 July 2026
- Lab
- OpenAI
- Kind
- model
- Access
- closed API
- Price
- Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per 1M tokens at launch
Figures
| Measure | Value | Measured by |
|---|---|---|
| Agents' Last Exam (Sol) | 53.6 13.1 points above Claude Fable 5 per OpenAI; company-run | company |
| BrowseComp (Sol) | 92.2% | company |
| ExploitBench (Sol, no production safeguards) | 73.5% GPT-5.5 47.9% | company |
| API price Sol / Terra / Luna (input per 1M tokens) | $5 / $2.50 / $1 output $30 / $15 / $6; Sol cut to $4/$20 on 2026-08-21 | company |
Per datanorth.ai (secondary) a voluntary US government review preceded general availability; OpenAI's own page says only that it worked with expert organizations during the preview. On OpenAI's own table Claude Fable 5 led on SWE-Bench Pro (80% vs 64.6%). Luna fell 80% in price on 2026-07-30; Terra 20%.
Sources
- openai.com/index/gpt-5-6/
- deploymentsafety.openai.com/gpt-5-6/
- deploymentsafety.openai.com/gpt-5-6-preview/
- simonwillison.net/2026/Jul/9/gpt-5-6/
- developers.openai.com/api/docs/changelog
This record was checked against its sources on 6 October 2026. How we check