What changed
OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026. API list prices are $2/M input and $10/M output for Sol and $0.10/M input and $0.50/M output for Luna, which OpenAI describes as 50% below GPT-5.6 promotional pricing. The models inherit methods from GPT-6 Astra and target cheaper professional, coding, agent and computer-use workloads. OpenAI also improved GPT-6 prompt caching: cached input reads receive a 90% discount, reasoning-effort and tool-availability changes can preserve earlier cached context, and developers get cache diagnostics and explicit prefix breakpoints.
Why it matters
For builders running agents continuously, model economics are moving from token sticker price toward cost per completed task and cost of repeatedly reprocessing context. Sol and Luna lower both sides of that equation: list prices fall sharply while caching makes long-lived conversations and agent prefixes cheaper to reuse. This can change routing decisions even where the flagship Astra remains more capable, because many recurring coding, extraction, workflow and computer-use jobs do not need the most expensive tier.
OpenAI cut the recurring price floor, not just a launch promotion
GPT-6 Sol costs $2/M input and $10/M output; GPT-6 Luna costs $0.10/M input and $0.50/M output. OpenAI says those are 50% below the GPT-5.6 promotional rates. Independent specialist coverage characterizes the models as broadly preserving predecessor-class performance while making price the larger generational change.
The coding economics are aimed directly at sustained agent use
OpenAI reports GPT-6 Sol at 68.8% on DeepSWE v1.1, within 1.1 percentage points of the Claude Fable 5 result it cites while costing about 80% less per task under its assumptions. Luna reaches 66.6% at max effort and is claimed to cost 93% less per task than Opus 5 and 96% less than Fable 5 in the cited comparison. These are OpenAI-presented comparisons and should be reproduced on real repositories.
Caching becomes part of the agent cost model
GPT-6 cached input reads are discounted by 90%. OpenAI says changing reasoning effort or enabled tools can now preserve earlier context for cache reuse, and explicit cache breakpoints plus diagnostics make missed caching visible. GitHub reports that related improvements reduced the share of prompt tokens requiring fresh processing by more than 50% across billions of OpenAI-model requests in Copilot.
Astra stays the escalation tier
OpenAI still positions GPT-6 Astra as its best model. The practical architecture is therefore increasingly tiered: Luna for high-volume inexpensive work, Sol for stronger recurring coding and agent tasks, and Astra for the hardest cases. Applications that already route by task difficulty can potentially capture the new economics without changing their overall orchestration design.