New Models and Permanent Price Reductions

OpenAI has introduced two new models in its GPT-6 family, GPT-6 Sol and GPT-6 Luna, with API pricing reduced by approximately 50% compared with the promotional rates for their GPT-5.6 predecessors. The cuts took effect as permanent rates rather than temporary promotions.

GPT-6 Sol is priced at $2 per million input tokens and $10 per million output tokens, down from $4 and $20. GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens, reduced from $0.20 and $1.20. The output price for Luna represents a steeper decline of roughly 58%. GPT-6 Astra remains the top-tier model at $10 per million input tokens and $50 per million output tokens.

OpenAI attributes the lower rates to improvements in inference efficiency and prompt caching. Cached input tokens reused within a 30-minute window can receive discounts of up to 90%. New developer tools include a caching dashboard, diagnostics for cache misses, and the ability to adjust reasoning effort without breaking the cache.

Three-Tier Structure and Availability

The GPT-6 lineup now operates as a three-tier system sharing the same training approach. Astra handles the most demanding reasoning and multi-step work. Sol targets complex coding, feature development, code review and data analysis. Luna is designed for high-volume, narrower tasks such as summarization, information extraction and routine queries.

Both Sol and Luna are available immediately in the OpenAI API under the names gpt-6-sol and gpt-6-luna. They are also accessible in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Luna has reached the ChatGPT desktop app for Free and Go users. Neither model has yet rolled out to the standard ChatGPT chat interface, with OpenAI describing that process as gradual.

Performance Claims Focused on Cost Per Task

OpenAI is presenting the models primarily through the lens of cost per completed task rather than raw benchmark scores alone. On AutomationBench, which tests full business workflows across dozens of tools, GPT-6 Sol at its highest effort setting scored 33.2% at $0.27 per task according to OpenAI figures. Claude Opus 5 at maximum effort scored 26.9% while costing more than 11 times as much per task on the same metric.

On Agents’ Last Exam, Sol at maximum effort reached 56.4%, which OpenAI stated exceeded Claude Opus 5’s best recorded result at 60% lower cost per task. On the DeepSWE v1.1 coding benchmark, Sol achieved 68.8% at maximum effort, while Luna scored 66.6% at substantially lower cost. On the OSWorld 2.0 computer-use benchmark, Sol scored 60.5%, slightly above Opus 5’s medium-effort result of 60.3%, at roughly 80% lower cost per task.

OpenAI also reported gains in factual reliability. An internal evaluation based on previously flagged ChatGPT conversations found GPT-6 Sol made about half as many mistakes as GPT-5.6 Sol. Luna at higher effort settings matched GPT-5.6 Sol’s factuality at a fraction of the cost. The company noted these evaluations focus on error-prone cases and do not represent typical usage.

Same-Day Response from Anthropic

Anthropic released Claude Opus 5.5 on the same day, priced at $4 per million input tokens and $20 per million output tokens. That represents a 20% reduction from the prior Opus 5 rates. Anthropic stated the new model should reduce overall spending by about 40% on typical workloads through greater token efficiency depending on effort settings.

GPT-6 Sol’s input price of $2 sits at half the rate of Opus 5.5 and matches Claude Sonnet 5. Both GPT-6 Astra and Anthropic’s Claude Fable 5.1 remain at $10 input and $50 output per million tokens.

Direct head-to-head results between Sol and Opus 5.5 under identical conditions were not yet available at the time of the announcements. Some reported AutomationBench figures showed Opus 5.5 achieving a higher completion rate than Sol but at several times the cost per task.

Broader Competitive Pressure and Total Cost Trends

The releases occur against a backdrop of intensifying price competition. Open-weight models from Chinese firms, including offerings from Xiaomi, DeepSeek and others, continue to provide lower-cost alternatives, with some available for self-hosting under permissive licenses. Platform data has shown rising token share for non-U.S. models in recent months.

Analysts have noted that falling per-token prices do not automatically translate into lower overall spending. Gartner has projected that inference costs per agentic workflow could rise more than fivefold through 2028 as workflows grow more complex and models consume more tokens. A University of Oxford analysis found that while quality-adjusted inference prices have declined rapidly, costs measured per completed task have stopped falling because newer reasoning models require additional tokens.

OpenAI’s Sol and Luna models are positioned to compete on the combination of capability, reliability and task-level economics rather than on being the absolute lowest-priced options available.