OpenAI previews GPT-5.6: Sol, Terra, and Luna

Flagship Sol is trained for agentic coding and cybersecurity and supports ultra mode with subagents; priced at $5/$30 per 1M tokens, currently limited preview for select partners.

Author: Michael Kokin ·

OpenAI is rolling out its answer to Anthropic's Mythos — the GPT-5.6 series. Sol is the flagship, Terra and Luna are smaller and faster.

What the models do

Terra is promised at roughly GPT-5.5 level, but at half the price. Sol was trained with a focus on agentic coding, biology/genomics, and cybersecurity — specialized profiles that previous GPT models didn't have.

New modes

Two new modes are coming: max reasoning effort for deeper reasoning, and ultra mode, where the model uses subagents(!) for complex tasks. Effectively, one model becomes an orchestrator for a fleet of agents.

Safety and access

The model is significantly more capable, so OpenAI is doing a staged rollout with stricter safeguards. By their own assessment, Sol doesn't cross the Cyber Critical threshold — their internal risk classification for cybersecurity capabilities.

Access is currently a limited preview in API and Codex for select trusted partners.

Pricing per 1M tokens

Sol on Cerebras (maker of wafer-scale AI inference chips) is planned for July at up to 750 tokens/sec — limited access there too at first.

Source