Artificial Intelligence
OpenAI launches cheaper GPT-6 Sol and Luna
OpenAI introduced GPT-6 Sol and Luna at lower token prices, said the models matched earlier versions on cost-adjusted benchmarks, and an independent analysis found mixed score changes.
- OpenAI released GPT-6 Sol and Luna and priced Sol at $2 per million input tokens and $10 per million output tokens, with Luna at $0.10 and $0.50.
- OpenAI said the lower prices came from caching and inference improvements, and that Terra was no longer available.
- The company said Sol was for recurring complex work, while Luna was for large volumes of well-defined tasks.
- OpenAI made the models available in ChatGPT Work and Codex, with Luna also available to free and Go users in the desktop app.
- Artificial Analysis said the models reduced per-task costs but did not improve intelligence scores overall, and it reported declines on GDPval-AA v2.1.
OpenAI released GPT-6 Sol and Luna and said Sol cost $2 per million input tokens and $10 per million output tokens, while Luna cost $0.10 and $0.50. It said the lower pricing came from caching and inference changes, and that Terra was no longer offered.
The company described Sol as a model for repeated complex work such as feature building, code review, debugging and data analysis. It said Luna was for large numbers of tightly defined tasks, including document summaries, information extraction and short answers.
OpenAI said GPT-6 caching now carried a 90% discount on cached input tokens. It also added a caching dashboard and diagnostics tool, and said developers could alter reasoning effort and tool access without invalidating the cache.
At launch, the models were available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu subscribers. Free and Go users could access Luna through the desktop app, and the API names were gpt-6-sol and gpt-6-luna.
OpenAI compared the models mainly with Anthropic’s Claude line and cited OSWorld 2.0, AutomationBench, FrontierCode 1.1 and DeepSWE v1.1. The article also said Artificial Analysis found lower per-task costs but mixed score changes, including declines on GDPval-AA v2.1.
Named in this story
Companies
- OpenAI
- launched GPT-6 Sol and Luna and set their prices
- Anthropic
- was the rival company whose models were used for comparison
Organisations
- Artificial Analysis
- published the independent analysis cited by the article
Products and systems
- GPT-6 Sol
- was the higher-tier new model aimed at recurring complex tasks
- Luna
- was the lower-cost new model aimed at routine high-volume tasks
- GPT-5.6 Sol and Luna
- were the previous models used for the price and performance comparison
- Terra
- was the cheapest earlier model and was no longer available
- ChatGPT Work
- was one of the places where the models were made available at launch
- Codex
- was another launch surface for the models
- Claude Opus 5
- was a rival model compared with GPT-6 Sol and Luna
- Claude Fable 5.1
- was a rival model compared on coding and software tasks
- Opus 5.5
- was cited as a newer rival model OpenAI appeared to have missed
- Astra
- was said to lead the OSWorld 2.0 computer-use benchmark
How the source tells it
The piece is openly sceptical, combining launch reporting with pointed criticism of OpenAI’s pricing, benchmark choices and presentation of results.
- scepticism the article repeatedly sets OpenAI’s claims against independent results and missing benchmarks
- critique selective test coverage and weaker scores are used to question the company’s framing
- dismissiveness rhetorical questions and pejorative asides are used to undermine the launch narrative