OpenAI Launches GPT-6 Sol and Luna at Half the API Cost
OpenAI launches GPT-6 Sol and Luna at half the API cost of 5.6 models, claiming fewer factual and coding errors as Anthropic AI competition intensifies.
Summary
OpenAI launched updated GPT-6 Sol and Luna models on September 22, 2026, expanding the generation after releasing GPT-6 Astra earlier in September as its most powerful model, with claimed world-leading performance in computer work and coding. The first Sol and Luna versions arrived earlier in 2026. Sol targets complex work such as coding, while Luna handles high-volume clerical tasks including document summaries, information extraction and quick answers.
API access costs half as much as the 5.6 Sol and Luna series, which OpenAI attributes to caching and inference improvements. In an internal factuality evaluation using de-identified conversations where users flagged errors, GPT-6 Sol made about half as many mistakes as its predecessor and matched Astra-level reliability at lower cost. OpenAI also claims fewer coding errors and stronger performance than Anthropic’s Fable and Opus models. Anthropic released Opus 5.5 just 90 minutes earlier. Sol and Luna are available through the ChatGPT API and in ChatGPT Work and Codex for most paid accounts. Luna will also reach the desktop app and Free and Go users, with the broader ChatGPT app and website rollout continuing throughout the day.
Positives
- API pricing is half that of the 5.6 Sol and Luna series because of improved caching and inference.
- GPT-6 Sol made about half as many mistakes as its predecessor in OpenAI’s internal factuality evaluation.
- GPT-6 Sol reportedly reaches Astra-level reliability at a substantially lower cost.
- Sol supports complex coding tasks, while Luna addresses summaries, information extraction and quick answers at high volume.
- ChatGPT API, Work and Codex users receive immediate access, with Luna extending to Free and Go accounts.
Risks & concerns
- OpenAI’s factuality result comes from an internal evaluation based on de-identified conversations where users flagged errors.
- Claims that Sol and Luna outperform Anthropic’s Fable and Opus models are OpenAI’s own comparisons.
- OpenAI disclosed no numerical improvement for the models’ lower coding error rate.
- Initial ChatGPT Work and Codex access covers most paid accounts, while the wider app and website rollout remains gradual.
- Anthropic released Opus 5.5 only 90 minutes earlier, underscoring rapid competitive pressure on both companies.