OpenAI Expands GPT-6 With Sol and Luna, and Cuts API Prices in Half

OpenAI is expanding the GPT-6 model family with GPT-6 Sol and GPT-6 Luna, two models designed to bring advances introduced with GPT-6 Astra to faster and more affordable tiers. The company says the models improve professional work, factual reliability, coding, computer use, collaboration style, and alignment while substantially lowering operating costs.

The biggest pricing change is a 50% reduction in API prices compared with GPT-5.6 promotional pricing. GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens, while GPT-6 Luna costs $0.10 for input and $0.50 for output. GPT-6 Astra remains OpenAI’s highest-performing model for users seeking maximum capability.

OpenAI is positioning Sol and Luna around the cost-versus-intelligence tradeoff. Across benchmarks cited in the article, Sol delivers stronger professional-work and coding performance than GPT-5.6 Sol while competing closely with, or outperforming, models from Anthropic at substantially lower cost. Luna is positioned as an even more economical option that still delivers major improvements over its predecessor.

Key Points

• API prices have been cut sharply. GPT-6 Sol’s input price falls from $4 to $2 per million tokens and output from $20 to $10. GPT-6 Luna falls from $0.20 to $0.10 for input and from $1.20 to $0.50 for output.
• GPT-6 Astra remains the flagship. OpenAI continues to describe Astra as its best overall model, while Sol and Luna target users who want strong intelligence with better economics.
• Sol shows significant gains in professional workflows. On AutomationBench, GPT-6 Sol at xhigh effort scored 33.2%, compared with 26.9% for Claude Opus 5 at maximum effort, while OpenAI says Sol operated at only 9% of Opus 5’s cost per task.
• Factual reliability improved. OpenAI says GPT-6 Sol produces roughly half the factual mistakes of GPT-5.6 Sol on its internal evaluation. Luna also improves and, at higher reasoning effort, can match GPT-5.6 Sol’s factuality at roughly one-hundredth of the cost.
• Coding performance is becoming cheaper to scale. On DeepSWE, GPT-6 Sol scored 68.8%, close to Claude Fable 5’s 69.9%, while OpenAI says Sol cost approximately 80% less per task. GPT-6 Luna reached 66.6% and was substantially cheaper than the compared Claude configurations.
• Computer-use capabilities also improved on cost efficiency. GPT-6 Sol scored 60.5% on the OSWorld 2.0 offline benchmark compared with Claude Opus 5’s 60.3% at medium effort, with OpenAI reporting roughly 80% lower cost per task.
• The communication style has been refined. OpenAI says Sol and Luna inherit Astra’s improvements in clarity, reduced jargon, fewer unnecessary details, and somewhat shorter responses without sacrificing substance.
• Prompt caching is becoming a bigger part of the economics. GPT-6 offers higher cache-hit rates and discounts of 90% on cached input-token reads. OpenAI also added tools for monitoring caching and changing reasoning effort or tool availability without invalidating earlier cached context.
• GitHub is already seeing an impact from caching improvements. According to the article, GitHub reports that the share of prompt tokens requiring fresh processing has fallen by more than 50% across billions of OpenAI-model requests over recent months.
• Alignment remains part of the upgrade. OpenAI says Sol and Luna improve on their GPT-5.6 counterparts in alignment evaluations, including lower rates of misleading claims concerning coding work, while stressing that these tests deliberately target difficult situations rather than typical use.
• Availability begins across OpenAI’s work-focused products and API. The article says GPT-6 Sol and Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, while Free and Go users can access GPT-6 Luna in the desktop app. API model names are gpt-6-sol and gpt-6-luna.

Key Quotes

“reducing API prices for Sol and Luna by 50%”

“about half as many mistakes as its predecessor”

“more clarity, less jargon”

Implications

The clearest implication in the article is that OpenAI wants advanced AI workloads to become economically practical at much larger scale. The combination of lower token prices, stronger benchmark performance, and improved caching means developers can potentially run longer conversations, coding agents, business workflows, and computer-use tasks while spending considerably less per task. OpenAI explicitly frames these changes as making advanced AI practical for more everyday applications at scale.

The announcement also creates a clearer division inside the GPT-6 family. Astra remains the option for maximum capability, while Sol is positioned for demanding professional work where performance and cost both matter. Luna pushes further toward high-volume, lower-cost workloads while still inheriting many of the improvements introduced with Astra.

For developers building agents, the caching changes may be nearly as important as the headline token-price cuts. A 90% discount on cached input reads, combined with the ability to change reasoning effort and available tools without losing cache reuse, is specifically designed to make long-running agents and context-heavy applications faster and less expensive to operate.

Finally, the article signals that OpenAI is competing not simply on absolute benchmark performance but increasingly on performance per dollar. Several of the comparisons emphasize cases where Sol or Luna approach or exceed competing model performance while costing substantially less per task.

Source: https://openai.com/index/introducing-gpt-6-sol-and-luna/