GPT-6 Sol and GPT-6 Luna: Pricing, Availability and Benchmarks

Two cheaper tiers under Astra

OpenAI added two models to the GPT-6 family nineteen days after GPT-6 Astra. Its system card describes GPT-6 Sol as a highly capable, lower-cost alternative to Astra and GPT-6 Luna as its fastest and most cost-efficient model. OpenAI says both were trained with methods similar to Astra's. [1][4]

In the API, Sol costs $2 per million input tokens and $10 per million output tokens, and Luna $0.10 and $0.50. OpenAI presents that as a 50 percent cut from the promotional prices of GPT-5.6 Sol and Luna; Luna's output price actually fell from $1.20, a slightly larger cut. Both models accept a context window of about 1.05 million tokens and return up to 128,000 tokens, and prompts beyond 272,000 input tokens are billed at higher rates. [1][5]

Work and Codex first, not Chat

On launch day the models reached ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, rolling out gradually through the day. Free and Go users could use Luna in the desktop app. OpenAI stated that the models were not yet available in Chat, the ordinary ChatGPT conversation view. [1]

Atlas interpretation: Putting the new tiers where agents do long, token-heavy work before putting them in the chat window matches the pitch: this release is about the price of sustained use, which is where coding and business-automation agents spend money. It also landed twelve days after OpenAI stopped selling its most usage-heavy consumer plan, pausing new $200 Pro sign-ups under Astra demand. [1][6]

OpenAI's comparisons, and the cyber rating

OpenAI's own evaluations put Sol at 68.8 percent on DeepSWE v1.1 at maximum effort, within 1.1 points of Claude Fable 5 at about 80 percent lower cost per task, and at 60.5 percent on the offline OSWorld 2.0 set against 60.3 percent for Claude Opus 5 at medium effort. On OpenAI's internal factuality evaluation, built from conversations where users flagged mistakes, Sol made about half as many errors as its predecessor. [1]

An appendix added to the GPT-6 system card says OpenAI is treating both models as High capability in cybersecurity, below the Critical level that Astra crossed. [4]

Atlas interpretation: Every comparison above is OpenAI's measurement, and several pit Sol against competitors at chosen effort settings or against scores from an older Claude model when the newer one's were unavailable, as OpenAI's footnotes say. The same day, Anthropic shipped Claude Opus 5.5, so the comparison tables were out of date within hours. [1][2]