GPT-6 moves downmarket with Sol and Luna as API prices are cut in half

GPT-6 Sol and GPT-6 Luna extend the Astra family with lower-cost models designed for professional work, coding, and agents. OpenAI is cutting API prices by 50% compared with GPT-5.6 while also improving caching for long-running workflows.

Astra stays at the top, Sol and Luna lower the cost

Following GPT-6 Astra, the family now expands with two models designed for less expensive workloads. GPT-6 Sol and GPT-6 Luna use training methods similar to Astra, with claimed improvements across professional work, factuality, coding, computer use, and alignment. Astra remains the top model in the lineup for the most demanding tasks.

The first major difference is price. GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20 for GPT-5.6 Sol at its promotional reference pricing. Luna drops to $0.10 for input and $0.50 for output, compared with $0.20 and $1.20 previously. OpenAI describes the change as a 50% API price reduction across both models. Sol for heavier workloads, Luna for lower costs

Sol sits between Astra and Luna. On AutomationBench, which evaluates professional workflows across 47 tools, GPT-6 Sol reaches 33.2% at xhigh effort for a reported average cost of $0.27 per task. Claude Opus 5 scores 26.9% in OpenAI's published evaluation, with an estimated cost 11.1 times higher. These comparisons remain specific to OpenAI's evaluation setup and cost methodology.

For software engineering, Sol reaches 68.8% on DeepSWE 1.1, compared with Claude Fable 5's highest measured score of 69.9%, at a per-task cost OpenAI estimates to be roughly 80% lower. Luna reaches 66.6% on the same benchmark and approaches intermediate configurations of Anthropic models in the published evaluation at a substantially lower reported cost.

A similar pattern appears in computer use. GPT-6 Sol reaches 60.5% on the offline set of OSWorld 2.0 at xhigh effort, compared with 60.3% for Claude Opus 5 at medium effort in the results reported by OpenAI. Fewer factual errors in OpenAI's internal evaluation

OpenAI also reports that GPT-6 Sol makes roughly half as many factual mistakes as GPT-5.6 Sol on its internal factuality evaluation. The test uses de-identified conversations where users had previously flagged factual errors.

The dataset is deliberately difficult, and OpenAI notes that it is not representative of typical conversations. Luna also improves and, at higher reasoning settings, approaches GPT-5.6 Sol's reliability at a fraction of the cost, according to the company. Caching becomes another part of the cost equation

Lower token pricing comes alongside changes to prompt caching. GPT-6 can reuse more context between requests, with a 90% discount applied to cached input-token reads.

Developers can also change reasoning effort or tool availability without breaking the previously cached context. Explicit breakpoints provide more control over which prompt prefixes are retained. GitHub says these changes have reduced the share of prompt tokens requiring fresh processing by more than 50% across billions of OpenAI model requests in Copilot over the past several months. High capability in cyber and bio/chem

The GPT-6 System Card adds another layer beyond product positioning. OpenAI classifies both Sol and Luna as High capability in cybersecurity and biological and chemical domains, while keeping them below the Critical threshold. Neither reaches the High threshold for AI Self-Improvement.

The gap with Astra remains substantial on some cyber evaluations. On ExploitBench, GPT-6 Sol reaches a maximum success rate of 5.5%, while Luna completes none of the challenges, compared with 31.5% for Astra. OpenAI concluded that both models remain below its Critical cybersecurity threshold.

GPT-6 Sol and Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, as well as through the API as `gpt-6-sol` and `gpt-6-luna`. Luna is also available to Free and Go users through the desktop app.