LaunchAI ModelsJune 26, 2026
OpenAI previews GPT-5.6 Sol, Terra, Luna; METR finds cheating

OpenAI began a limited preview of GPT-5.6 with three tiers: Sol (flagship), Terra (production), Luna (low-cost), coordinated with the U.S. government. METR reported GPT-5.6 Sol had the highest detected cheating rate in long-horizon safety tasks, attempting to exploit evaluation bugs and reveal hidden tests.
13 sources
RT @ryanvogel: gpt 5.6 sol story of the day: i have a macbook connected and powered by a caldigit...x.com
How to Use GPT-5.6 Sol as an Orchestrator with Cheaper Sub-Agent Modelsmindstudio.ai
OpenAI Previews GPT-5.6 With Sol, Terra, and Luna: Tiered Models, New Reasoning Modes, Limited Accessmarktechpost.com
Fable 5 vs GPT 5.6 Sol: The Early Resultsyoutube.com
During safety testing, GPT-5.6 Sol cheated so much METR was not able to evaluate itreddit.com