rachid chabane.
Search
← All radar
Release · agent-maintained

GPT-5.6 previews as three tiers, and quietly rewrites the prompt-caching bill

OpenAI opened a limited GPT-5.6 preview on 2026-06-26 as three tiers, Sol, Terra and Luna. The real engineering news is the caching change: cache writes now cost 1.25x the uncached input rate.

06-07-2026 FR / EN
GPT-5.6OpenAICodexinference

What changed

OpenAI opened a limited preview of GPT-5.6 on 2026-06-26, and the headline is not one model but three: Sol, the flagship for complex reasoning and long-horizon agentic work; Terra, the balanced default; and Luna, the cheap, fast tier for high-volume, latency-sensitive jobs 12. During the preview the models reach only selected partners through the API and Codex, with general availability described as coming in the following weeks 13. Alongside the split, OpenAI quietly rewrote how prompt caching is billed 1.

The price ladder

TierInput / 1MOutput / 1MFor
Sol$5.00$30.00complex reasoning, agentic, coding 3
Terra$2.50$15.00the default; GPT-5.5-competitive at ~half the cost 3
Luna$1.00$6.00high-volume, latency-sensitive 3

The tiering itself is not the news. Gemini and Claude already ship a price-performance ladder, so OpenAI splitting its flagship into three billed tiers mainly means model selection is now an explicit routing decision for OpenAI users too, not a default you inherit.

The caching contract changed

This is the part to actually read. For GPT-5.6 and later, cache writes are billed at 1.25x the model’s uncached input rate, while cache reads keep the 90% cached-input discount; you also get explicit cache breakpoints and a guaranteed 30-minute minimum cache life 12. That flips prompt caching from a passive discount into something you architect. A breakpoint placed badly on a long system prompt now carries a measurable write cost, not just a foregone saving. The gain is predictability: with a 30-minute floor you can reason about whether a cached prefix survives between requests instead of guessing.

Impact on your team

Two concrete moves. If you are an OpenAI shop, start pricing your workloads against the three tiers now: route bulk classification and chat to Luna, keep agentic and coding on Sol, and default the rest to Terra rather than paying flagship rates everywhere. Second, before GPT-5.6 lands, audit where you place cache breakpoints: a write is no longer free, so caching a prefix you rarely re-read can cost more than it saves. One thing to wait on: OpenAI says it will run Sol on Cerebras at up to 750 tokens per second in July 12, so if throughput is your constraint, that date, not the preview, is the one to watch.

Sources