Summary
OpenAI released GPT-5.6 in three tiers on July 9, 2026: Sol at $5/$30 per million tokens (91.9% on Terminal-Bench 2.1), Terra at $2.50/$15 for everyday use, and Luna at $1/$6 as the low-cost option. The models are available in paid ChatGPT tiers, the OpenAI API, and Codex, add prompt caching with a 90% discount on cache reads, and ship extended reasoning modes including an Ultra mode that runs multiple parallel subagents. GPT-5.6 Sol also runs on Cerebras wafer-scale hardware at up to 750 tokens per second.
What changed
OpenAI shipped the three-tier GPT-5.6 family (Sol, Terra, Luna) to ChatGPT, API, and Codex with published per-token pricing, 90% prompt-cache read discounts, Ultra and Max reasoning modes, and a Cerebras deployment reaching about 750 tokens per second.
Why it matters
A tiered frontier lineup with explicit price-performance segmentation lets teams match model cost to task, while Codex availability and an Ultra mode that spawns parallel subagents push agentic coding forward. Wafer-scale inference at ~15x typical GPU speed signals that serving throughput, not just model quality, is becoming a competitive axis.
Evidence excerpt
OpenAI launched GPT-5.6 Sol, Terra, and Luna for all ChatGPT users and API developers; GPT-5.6 Sol runs on Cerebras wafer-scale hardware at up to 750 tokens per second.