OpenAI has released two new AI models, GPT-6 Soul and GPT-6 Luna, that cut API costs roughly in half compared to earlier versions while significantly reducing errors and improving accuracy. The launch expands OpenAI's GPT-6 family just weeks after introducing GPT-6 Astra, its most powerful model to date, and comes amid escalating competition with Anthropic over AI efficiency and performance.

What Are Soul and Luna Designed to Do?

Soul and Luna serve different purposes within OpenAI's ecosystem. Soul is built for complex tasks requiring advanced reasoning, such as coding and professional work that demands deep analysis. Luna, by contrast, targets high-volume office work with clear objectives, like summarizing documents, extracting information, and answering straightforward questions. Both models were trained using methods similar to those behind GPT-6 Astra, allowing OpenAI to bring improvements in professional work, factual accuracy, coding, computer use, and alignment to faster and more affordable models.

How Much Cheaper Are These Models?

The pricing reduction is dramatic. GPT-6 Soul costs $2 per million input tokens and $10 per million output tokens, while GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens. This represents roughly a 50% reduction in API costs compared with GPT-5.6 Soul and Luna. OpenAI attributed the lower prices to improvements in caching and inference technologies, allowing the models to operate more efficiently at a lower cost.

Where Can You Access These Models?

The new versions of Soul and Luna are available across multiple platforms and user tiers. Here's how to find them:

• ChatGPT Work and Codex: Available for users on Plus, Pro, Enterprise, Business, and Edu plans
• ChatGPT API: Accessible through OpenAI's API for developers and businesses
• Desktop App: Luna is available through the ChatGPT desktop app for Free and Go plan users
• Gradual Rollout: The models were not yet available in ChatGPT's main chat interface at announcement time, with OpenAI planning to roll them out throughout the day via the app and website

How Much Better Are These Models at Accuracy?

OpenAI conducted internal evaluations showing substantial improvements in reliability and error reduction. In accuracy testing based on anonymized real-world conversations where users had reported model errors, GPT-6 Soul made about half as many errors as its predecessor. The company stated that Soul now offers a level of reliability close to Astra, but at a significantly lower cost.

Luna showed equally impressive gains. According to OpenAI's data, GPT-6 Luna outperformed version 5.6 by 5.4% on an automation-task performance test, while the cost of completing each task fell by 58%. These improvements suggest that OpenAI has managed to increase both capability and efficiency simultaneously, a rare achievement in AI development.

How Does This Fit Into OpenAI's Competitive Strategy?

The launch comes amid intensifying competition between OpenAI and Anthropic. Anthropic announced a new version of its Opus 5.5 model approximately 90 minutes before OpenAI announced Soul and Luna, signaling the rapid pace of model releases in the AI industry. OpenAI claimed that its two new models outperform Anthropic's advanced models in several tasks, including its Fable and Opus models, based on performance tests published by the company.

The timing and competitive positioning highlight how the AI landscape is shifting from raw capability races toward efficiency and cost optimization. Both companies are now competing on delivering powerful models at lower operational costs, making advanced AI more accessible to businesses and developers. This shift could reshape how enterprises choose between AI providers, favoring those that offer the best balance of performance and affordability.