GPT-6 Sol and Luna Launch, Cutting Token Prices and Expanding Agent Cache Controls

OpenAI launched GPT-6 Sol and Luna on September 22, with the API names `gpt-6-sol` and `gpt-6-luna`. Standard input/output prices per million tokens are $2/$10 for Sol and $0.10/$0.50 for Luna; both Sol rates are half its predecessor’s promotional prices. For coding agents that operate over multiple turns, the cache controls announced alongside the models are another key deployment consideration. [Model announcement](https://openai.com/index/introducing-gpt-6-sol-and-luna/)

Prompt caching stores the key-value states of shared prefixes, reducing repeated processing of instructions, tool definitions, and message history. The announcement introduces a cache dashboard, cache-miss diagnostics, and explicit breakpoints that let developers specify where reusable content ends. According to OpenAI, eligible prefixes qualify for caching discounts within a 30-minute window. These features make cache performance measurable and allow frequently changing content to be handled differently from stable prefixes. [Caching announcement](https://openai.com/index/better-prompt-caching-for-gpt-6/)

Using the API comes with important conditions: increasing reasoning effort in subsequent responses requires appending a `configuration_update` while keeping the request’s top-level reasoning settings unchanged; directly changing a top-level value may still alter the prefix. Tool definitions, ordering, and structure should also remain unchanged, with `allowed_tools` used to control which tools can be called. The documentation also specifies that cache writes cost 1.25 times the standard input rate, while reads cost 0.1 times that rate, so “90% cheaper reads” does not apply to the entire bill. A cacheable prefix must contain at least 1,024 visible input tokens. If explicit-breakpoint-only caching is selected but no breakpoints are placed, the request will neither create nor use a cache. [API documentation](https://developers.openai.com/api/docs/guides/prompt-caching)

Reasoning budgets still affect the cost of each task. Artificial Analysis’s Sol comparison page lists composite index scores of 43 and 48 for high and max, respectively, with weighted average costs per question of $0.37 and $1.06. These results reflect that benchmark suite and cannot be treated as direct measures of success rates for fixes in enterprise codebases. Engineering teams still need to evaluate the models using the same agent framework, test set, and retry rules. [Benchmark comparison](https://artificialanalysis.ai/models/comparisons/gpt-6-sol-high-vs-gpt-6-sol)

GitHub announced the same day that both models were coming to Copilot through a gradual rollout with usage-based billing. In practice, developers should measure cache hit rates, latency to the first answer, total token usage, and task pass rates separately before deciding when to increase reasoning effort. Lower model prices provide more room for experimentation, but actual savings still depend on whether context can be reused and how many additional turns are needed after failures. Workflows involving Traditional Chinese documents should also separately check retrieval citations, tool arguments, and final answers, rather than relying solely on English-language coding benchmarks to determine model routing. [GitHub announcement](https://github.blog/changelog/2026-09-22-openais-gpt-6-sol-and-gpt-6-luna-now-available/)