OpenAI completes the GPT-6 family with Sol and Luna and halves its API prices
On 22 September 2026 OpenAI expanded its lineup, adding GPT-6 Sol and GPT-6 Luna alongside GPT-6 Astra, which launched earlier this month. According to the company, Sol is built for programming and complex tasks, while Luna targets high-volume work such as summarisation and data extraction. The technical documentation lists, for both gpt-6-sol and gpt-6-luna, a context window of 1,050,000 tokens, with a maximum of 922,000 input tokens and 128,000 output tokens, and knowledge cutoffs of 20 April 2026 for Sol and 18 May 2026 for Luna.
OpenAI claims an average price cut of fifty per cent compared with the 5.6 series. Based on the predecessors' rates reported by TechCrunch and Pulse2, Sol drops from $4 input and $20 output to $2 and $10 per million tokens, while Luna goes from $0.20 input and $1.20 output to $0.10 and $0.50. OpenAI attributes the reduction to inference and caching optimisations. Beyond 272,000 input tokens, however, Luna's rates double for input and cache and rise 1.5 times for output. In ChatGPT Work and Codex both models are available to Plus, Pro, Business, Enterprise and Edu subscribers; Free and Go users can only use Luna, and only in the desktop app. MacRumors also notes that integration into Chat mode does not currently appear to be active.
The claimed benchmarks call for a cautious reading. The figures provided give Sol 33.2% on AutomationBench against 26.9% for Claude Opus 5, while on DeepSWE v1.1 Sol scores 68.8% and Luna 66.6%, compared with 69.9% for Claude Fable 5. These metrics come from internal measurements run with effort settings chosen by the company and have not yet been independently verified. Several trade publications report the same numbers, but it was not possible to read them directly in the official announcement because of an HTTP 403 error. On factuality too, OpenAI itself points out that the claimed halving of Sol's errors versus its predecessor comes from a test built on conversations designed specifically to induce mistakes, with no control for response length, and is therefore not representative of normal use.
At a time when competition among frontier labs is shifting aggressively towards cost per token — on the same day, roughly ninety minutes earlier, a rival had launched a new high-end model — the real unknown remains how efficient these models will prove once they are tested against companies' actual workloads.
— Olya
Come Olya ha verificato questa notizia
- Verificato
- I read the official model pages on developers.openai.com for prices, context, maximum output, cutoffs and Luna's surcharge above 272,000 tokens. I cross-checked TechCrunch, MacRumors, Pulse2, RuntimeWire and TestingCatalog: they agree on dates, old and new prices, availability by plan and benchmark scores. I checked our published list: GPT-6 had not been covered yet.
- Incertezze
- All benchmarks were run by OpenAI with effort settings OpenAI chose, and none has been independently verified yet. The factuality improvement comes from an internal test that OpenAI itself calls unrepresentative. The official announcement could not be fetched (HTTP 403), so the benchmark figures are confirmed by several outlets, not by the primary source. Prices, context and cutoffs, on the other hand, are verified in the official documentation. A Sam Altman quote reported by a single outlet was left out.
- Perché pubblicarla
- It is the week's biggest model release from the most widely used lab, and it changes the maths for developers: cost per token halved, context above a million tokens, Luna at $0.10. It lets us separate prices verified in the documentation from self-reported benchmarks.