News · 2026-09-23

OpenAI ships GPT-6 Sol and Luna, making cache economics part of the model product

OpenAI released GPT-6 Sol and GPT-6 Luna on 22 September as lower-cost siblings of GPT-6 Astra, turning cache reuse and task economics into central parts of its product pitch. Sol targets difficult coding and agentic work, while Luna targets focused, high-volume workloads; the practical importance is that a cheaper token rate no longer describes the whole cost of running an agent.

Key facts

• OpenAI announced Sol and Luna on 22 September 2026.
• Both API models support text and image input and have a 1.05-million-token context window.
• OpenAI says GitHub reduced the share of prompt tokens requiring fresh processing by more than 50% across billions of requests.
• Primary source: OpenAI's launch announcement.

The release is best understood as a pricing ladder. Astra remains OpenAI's highest-capability model; Sol is the serious working tier; Luna is the economical worker. The Sol and Luna pages document the same 1.05M context window and 128K output limit, so Luna's trade-off is not basic access to long context but reliability at the hard end of a task. OpenAI wrote that Sol and Luna were trained with methods similar to Astra and inherit improvements in professional work, coding, computer use and alignment.

The mechanism behind the headline is serving as much as model design. OpenAI says cached reads cost one tenth of uncached input and cache writes cost 1.25 times input; a developer can change reasoning effort or enabled tools without necessarily breaking an earlier cache prefix. Think of an agent that must reread a large codebase every turn. Conventional billing makes it pay the librarian to reread every shelf; a useful cache lets it keep the shelves it already catalogued and only inspect the changed files. OpenAI's claim that GitHub cut fresh processing by more than half is an important anchor, but it is OpenAI-reported rather than an independent audit.

OpenAI's own benchmarks cast Luna as surprisingly capable: it reports 66.6% on DeepSWE v1.1 at maximum effort. But the stronger caveat comes from Artificial Analysis, which found Sol's Coding Agent Index up two points from GPT-5.6 Sol while Luna fell two points below GPT-5.6 Luna. Luna also used about 51,000 output tokens per task versus about 41,000 for its predecessor. That is why a low per-token sticker price can still produce a disappointing task bill or latency.

The launch's most useful line is OpenAI's own: it calls Luna a model for “focused, high-volume work.” That is a more defensible deployment category than treating it as an Astra substitute. Use Luna for well-specified extraction, classification, transformations and bounded implementation loops; evaluate Sol or Astra for open-ended planning and engineering review. The honest caveat is that benchmark effort levels, output lengths, cache hits and retry rates all change the result. Teams should record completed-task success, latency and cost together, rather than using a model name or input-token price as a proxy for value.

Primary source, verified: read the paper →

Ground Truth. (2026, September 23). OpenAI ships GPT-6 Sol and Luna, making cache economics part of the model product. Ground Truth. https://groundtruth.day/news/openai-gpt-6-sol-luna-cost-curve.html

@misc{groundtruth:openai-gpt-6-sol-luna-cost-curve,
  title  = {OpenAI ships GPT-6 Sol and Luna, making cache economics part of the model product},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/openai-gpt-6-sol-luna-cost-curve.html}
}

Topics: openai · models · agents · inference · pricing · prompt-caching