OpenAI unveils GPT-5.6: three models, Sol, Terra and Luna, for different tasks and budgets

OpenAI has officially unveiled the GPT-5.6 family of language models, structured across three capability tiers: the flagship Sol, the balanced Terra, and the accessible Luna. In the first phase, access is limited to a select circle of trusted partners through API and Codex. The company expects a broader launch, including through ChatGPT, in the coming weeks.

Three models for three tasks

GPT-5.6 launches as a family with a clear hierarchy — not a single universal flagship, but three separate models with distinct target scenarios. This is a direct response to demand fragmentation: a corporate client with thousands of requests per second needs one price-to-capability ratio, while a research group solving complex multi-step tasks needs something entirely different.

Sol — the top model in the family. Designed for scenarios where maximum reasoning accuracy matters: deep analysis of large documents, complex multi-step code, creation of agent chains, and decision-making under uncertainty. In positioning, Sol competes with Claude Opus 4 and Gemini 2.5 Ultra.

Terra occupies the central niche. It's a balanced model for the corporate mainstream: powerful enough for production tasks and fast enough for high-frequency use. Teams that built products on GPT-4 Turbo or GPT-4o will likely migrate to Terra — it's positioned as their direct successor.

Luna — the most accessible and fastest version. In price profile, it's close to GPT-4o mini or Claude Haiku: optimal for scenarios with high volumes of simple tasks — consumer chatbots, support automation, embedded assistants in mobile applications, where the cost of each token matters.

Closed launch through partners

Phased access rollout has become a consistent OpenAI tradition since GPT-4. The logic is straightforward: before the model reaches hundreds of millions of ChatGPT users, it must be tested by developers in real products under real loads. This allows identification of edge cases, security issues, and unexpected behaviors that don't emerge in laboratory tests.

In the first phase, GPT-5.6 Sol, Terra, and Luna are available through two channels:

• API — for development teams embedding models in their own products and services
• Codex — for software development tasks: code generation, refactoring, agentic programming
• ChatGPT — broad public access expected in several weeks
• Enterprise — corporate version to launch in parallel with the main public release

Response to market pressure

The appeal to a three-tier hierarchy reflects a fundamental shift in the LLM market. Today, no major provider launches a single "best" model. Google distributed Gemini 2.5 across several variants (Flash, Pro, Ultra), Anthropic built a Haiku–Sonnet–Opus lineup, Meta provides Llama 4 in multiple sizes under an open license. The principle is one: cover as many price segments as possible and retain customers within your ecosystem.

For developers, the Sol/Terra/Luna hierarchy simplifies decision-making — instead of comparing abstract benchmarks, the company itself has mapped the product matrix: task requires depth → Sol, need a combination of speed and quality → Terra, minimal request cost is critical → Luna.

What this means

GPT-5.6 is not a revolution, but a methodical step in OpenAI's transformation from an AI lab into a mature platform company with a classical product matrix. Developers wanting to build on cutting-edge models right now will either need to obtain partner access or wait for public launch. Regular ChatGPT users likely won't notice the transition — OpenAI traditionally updates the platform's base models without separate announcements.

*Meta is recognized as an extremist organization and is banned in the Russian Federation.