OpenAI introduced two new models for everyday work on September 22, 2026: GPT-6 Sol for complex programming tasks and GPT-6 Luna for quick routine tasks. Both replace the previous GPT-5.6 variants in ChatGPT Work, Codex, and the API, costing about half as much. Just 90 minutes earlier, Anthropic had released its model Claude Opus 5.5 with its own price reduction.
OpenAI halves the token prices for Sol and Luna
For GPT-6 Sol, the price per million input tokens drops from four to two dollars, and for output from 20 to ten dollars. GPT-6 Luna now costs only ten instead of 20 cents per million input tokens and 50 cents instead of 1.20 dollars for output. OpenAI attributes the discount to advancements in caching and inference, as TechCrunch reports.
One million tokens corresponds to about 750,000 words of continuous text, equivalent to a stack of documents several hundred pages thick – for two dollars, a complete contract folder can be processed company-wide. Both models consume fewer tokens per task than their predecessors. The actual savings in practice are likely to be higher than the mere list price. Sol makes only half as many errors as its predecessor. It thus achieves nearly the reliability of the larger model GPT-6 Astra – at a fraction of its cost. Luna, on the other hand, approaches the quality of GPT-5.6 Sol under high computational load, but costs only a fraction. It is particularly suitable for high-throughput tasks such as document review or customer inquiries.
Rollout reaches ChatGPT Work and Codex today
Both models are now available in ChatGPT Work, in Codex, and via the API. Initially, access is available to Plus, Pro, Business, Enterprise, and Edu accounts. Enterprise administrators must manually enable the new models in their settings before employees can select them in the web client or desktop app. Users with a free account or the entry-level Go plan will initially receive only GPT-6 Luna – and that exclusively in the desktop app. The web version remains closed to these accounts.
The rollout is spread throughout September 22, so not all accounts are switched at the same time. Individual teams will therefore see the new models in their accounts only later in the day. Whether the models are available simultaneously in Germany and the rest of the EU is left open by OpenAI. The company does not specify a separate regional restriction. OpenAI has not yet communicated a fixed end date for access to the predecessor models GPT-5.6 Sol and Luna. Existing Codex credits remain usable until a separate announcement.
Anthropic announces 90 minutes earlier
Just a little earlier on the same day, Anthropic released its new flagship model Claude Opus 5.5, reports TechCrunch. The price drops by 20 percent to four dollars per million input tokens and 20 dollars per million output tokens. Access to cached content costs even 60 percent less. The output speed increases by over 30 percent. Opus 5.5 runs on AWS, Google Cloud, and Microsoft Azure. According to Anthropic, it surpasses the previous model Fable 5.1 in agent-based programming, knowledge work, and image understanding.
OpenAI has so far only compared Sol in its own tests with the older model Opus 5. Under high computational load, Sol reportedly achieves a similar value as Opus 5 under medium load – 60.5 to 60.3 percent on an in-house agent benchmark. The costs are about 80 percent lower. A comparison with the newly released Opus 5.5 is still missing. The figure is independently unverified.
It will be crucial whether Anthropic renews its price comparison after its own Opus-5.5 launch or whether OpenAI follows up with a direct comparison between Sol and Opus 5.5. The smaller models Claude Sonnet 5.5 and Haiku 5.5 have also been announced, which according to Anthropic are expected to follow within a few weeks. It is the next step in a price race that provides corporate customers with cheaper AI usage from two major providers within a single day.
