OpenAI has unveiled GPT-5.6, a fundamentally restructured model family that abandons the single-model approach in favor of three distinct tiers designed for different workloads and budgets. The new lineup includes Sol (the flagship), Terra (for everyday production), and Luna (the fast, low-cost option), marking a strategic shift in how the company packages artificial intelligence capability.

The preview launched with roughly 20 trusted partners through OpenAI's API and Codex platform, with broader access to ChatGPT, Codex, and the API planned in the coming weeks. OpenAI also shared the models and plans with the U.S. government first, signaling the geopolitical importance of advanced AI systems.

What Makes GPT-5.6 Different From Previous Releases?

The shift from a single model to a tiered family represents a structural change in how OpenAI thinks about AI deployment. Rather than forcing all users into one model, the company now lets developers choose based on their specific needs: intelligence level, processing speed, or cost. Each tier can advance on its own schedule, giving developers clearer trade-offs without waiting for a monolithic upgrade cycle.

GPT-5.6 also introduces two new reasoning controls that fundamentally change how the model tackles complex problems. The first, called "max reasoning effort," gives Sol extended time to think through difficult tasks step by step. The second, "ultra mode," coordinates multiple subagents to split complex work in parallel, accelerating solutions for long-horizon problems. Think of max as deepening a single chain of thought, while ultra distributes work across several workers on one task.

How Do the Three Tiers Compare in Performance and Cost?

Sol, the flagship model, sets a new performance benchmark on Terminal-Bench 2.1, a test that measures command-line workflows requiring planning, iteration, and tool coordination. In ultra mode, Sol achieved 91.91% accuracy on this benchmark, while in max mode it reached 88.76%. On Agent's Last Exam, Sol was the only model to surpass the 50% mark, reaching 50.9% in code mode. On genomics analysis tasks (GeneBench v1), Sol outperformed the previous generation while using fewer tokens to do so.

Terra matches the performance of GPT-5.5, the previous generation, while costing roughly half as much, making it ideal for high-volume production work like summarizing thousands of support tickets daily. Luna brings strong capability at OpenAI's lowest price point, suited for latency-sensitive applications like email classification and autocomplete.

Pricing is structured per one million tokens processed. Sol costs $5 per million input tokens and $30 per million output tokens, matching GPT-5.5's pricing. Terra is approximately 2x cheaper than GPT-5.5, while Luna offers the fastest, most economical option for routine tasks. Prompt caching, a feature that stores frequently used text to reduce processing costs, now supports explicit cache breakpoints and a 30-minute minimum cache life, with cache writes costing 1.25x the uncached input rate and cache reads retaining a 90% discount.

Steps to Choose the Right GPT-5.6 Tier for Your Use Case

• Long-Horizon Coding and Security: Use Sol for multi-step automation tasks like planning, editing files, running tests, and iterating. Sol's Terminal-Bench performance makes it ideal for agents that need to coordinate multiple steps autonomously.
• High-Volume Production Work: Deploy Terra for chat features, document processing, and customer support at scale. Its cost efficiency and performance parity with GPT-5.5 make it suitable for applications processing thousands of requests daily.
• Latency-Sensitive Applications: Choose Luna for autocomplete, email routing, and simple data extraction where speed and cost matter more than deep reasoning capability.
• Defensive Security Research: Leverage Sol for vulnerability research and code patching, where the model's gains in cybersecurity benchmarks provide measurable advantage in finding and fixing memory bugs and security flaws.

On ExploitBench, a security-focused evaluation, Sol proved competitive with Claude Mythos 5 Preview while using approximately one-third of the output tokens, demonstrating significant efficiency gains for security-critical applications.

What Are the Practical Implications of This Tiered Approach?

The tiered model structure addresses a long-standing tension in AI deployment: developers often pay for capability they don't need. A company running routine customer service chatbots doesn't require Sol's reasoning power, just as a security researcher doesn't need Luna's speed constraints. By splitting the family, OpenAI lets teams right-size their AI spend while maintaining a clear upgrade path when needs change.

OpenAI also plans to run Sol on Cerebras hardware, targeting up to 750 tokens per second in July, which would represent a significant speed improvement for latency-critical applications. This hardware optimization suggests the company is preparing for enterprise-scale deployment where throughput directly impacts cost and user experience.

The new naming convention marks a departure from OpenAI's previous approach. The number (5.6) now indicates the generation, while the names (Sol, Terra, Luna) mark durable capability tiers that can evolve independently. This separation allows OpenAI to update Terra or Luna without necessarily updating Sol, reducing the pressure to bundle all improvements into a single release cycle.

However, access remains limited during the preview phase, and some details about real-world latency for max and ultra modes have not yet been disclosed. The company has also documented a layered safety stack, though some safeguards may block certain legitimate dual-use security research applications. Additionally, pricing sits above some open-weight competitors like GLM-5.2, which may influence adoption among cost-sensitive developers.