Summary

On June 26, 2026 OpenAI previewed GPT-5.6 as three tiered models: Sol (flagship), Terra (balanced), and Luna (fastest and cheapest), with Sol setting a new state of the art on Terminal-Bench 2.1 and a new max-reasoning 'ultra' mode. Broader availability follows a limited partner preview.

What changed

OpenAI previewed GPT-5.6 in three tiers: Sol ($5/$30 per 1M tokens), Terra ($2.50/$15, positioned as ~2x cheaper than GPT-5.5-competitive), and Luna ($1/$6). Sol set a new SOTA on Terminal-Bench 2.1; the release added max reasoning effort, an 'ultra' subagent mode, improved prompt caching with a 30-minute minimum, and is slated to run on Cerebras at up to 750 tokens/sec in July.

Why it matters

Tiering the frontier line by price and latency and topping a command-line coding benchmark targets agentic coding workloads directly, while the Cerebras deal chases the low-latency inference that long agent loops need. It sharpens the head-to-head with Claude Sonnet 5 and Gemini on both cost and coding capability.

Evidence excerpt

GPT-5.6 Sol sets a new SOTA on Terminal-Bench 2.1, which tests command-line workflows requiring planning, iteration, and tool coordination.

Sources