OpenAI Unveils GPT-6 Sol and Luna: A Paradigm Shift in Coding Intelligence and API Economics

Even in offline computer use via OSWorld 2.0, Sol (xhigh) achieved a 60.5% success rate, edging out Claude Opus 5 (medium) at 60.3%, but at a fraction of the operational cost.

• AutomationBench: Sol beats Claude Opus 5 at 9% cost.
• Agents' Last Exam: Sol scores 56.4% (surpassing Opus 5).
• DeepSWE v1.1: Sol (68.8%) nears Claude Fable 5 at 80% lower cost.
• OSWorld 2.0: Sol (60.5%) outperforms Opus 5 (60.3%) at 80% lower cost.

Revolutionary API Economics and Caching

OpenAI is aggressively targeting the developer market by slashing API prices by 50% compared to GPT-5.6 promotional rates. Sol is priced at $2/MTok input and $10/MTok output, making it competitive with high-end Sonnet-class models. Luna, however, is the true disruptor, priced at a mere $0.10/MTok input and $0.50/MTok output, enabling massive-scale extraction and summarization tasks at a tier previously thought impossible.

The real hero for production engineers is the new caching mechanism. OpenAI has implemented improved caching that delivers higher hit rates by default. Cached input-token reads are discounted by a staggering 90% ($0.20/MTok). This efficiency is already proven; GitHub reported seeing over 50% fewer fresh-processed prompt tokens across billions of requests using these optimized caching protocols.

• Sol Pricing: $2/MTok input, $10/MTok output.
• Luna Pricing: $0.10/MTok input, $0.50/MTok output.
• 90% Discount on Cached Input-Token Reads ($0.20/MTok).
• Massive efficiency gains for high-volume users like GitHub.

Strategic Use Cases

The dual-model approach allows for sophisticated pipeline architectures. Developers can now implement a 'tiered reasoning' strategy: use Luna for high-volume, low-complexity tasks like unit test generation, documentation scraping, and routine refactoring, then hand off complex architectural changes or debugging sessions to Sol.

For agentic workflows, Sol's high factuality—making roughly half as many errors as its predecessor—makes it ideal for autonomous coding agents. Meanwhile, Luna's performance at 'high effort' settings actually matches GPT-5.6 Sol at roughly 1/100th of the cost, making it the perfect candidate for massive RAG (Retrieval-Augmented Generation) pipelines and large-scale codebase indexing.

• Sol: Complex debugging, architectural design, and autonomous agents.
• Luna: Routine refactoring, documentation, and high-scale RAG.
• Hybrid Workflows: Tiered reasoning (Luna for speed, Sol for depth).

Getting Started

GPT-6 Sol and Luna are available immediately. For enterprise and professional users, they are integrated into ChatGPT Work and Codex (available for Plus, Pro, Business, Enterprise, and Edu tiers). Luna also extends its reach to Free and Go users via the ChatGPT desktop application.

Developers can access the models via the OpenAI API using the model IDs `gpt-6-sol` and `gpt-6-luna`. We recommend updating your SDKs to take advantage of the new caching headers to maximize your cost savings.

• API Model IDs: `gpt-6-sol`, `gpt-6-luna`.
• Availability: ChatGPT Work, Codex, and Desktop App.
• Tier Support: Plus, Pro, Business, Enterprise, Edu, Free, and Go.

API Pricing — Input: $2.00 per million tokens / Output: $10.00 per million tokens / Context: Sol Input: $2.00/MTok, Output: $10.00/MTok. Luna Input: $0.10/MTok, Output: $0.50/MTok. Cached input reads: $0.20/MTok.

Sources