OpenAI introduces GPT-5.6 for more efficient AI

OpenAI introduces the GPT-5.6 family with a specific promise: deliver frontier capabilities at a lower cost, both when training the models and running each response. The company says its main model, GPT-5.6 Sol, outperforms Claude Fable 5 on the Artificial Analysis Coding Agent Index using maximum reasoning mode and at less than half the cost.

The family includes three models with different goals:

• Sol: the most capable model, designed for complex reasoning and programming tasks.
• Terra: delivers performance similar to GPT-5.5 on intelligence tests, but at half the price.
• Luna: the fastest and most affordable model, with a cost 80% lower than Sol.

The idea is not just to have a better model. It is to let you use it more often without making every query so expensive.

More work with fewer tokens

OpenAI says GPT-5.6 has been trained to do more work with each token, the basic unit models use to process and generate text. In practice, this means it tries to follow more direct paths to complete a task instead of spending capacity on unnecessary steps.

This optimization particularly affects long tasks. A programming agent may need to read files, review issues, modify code and run tests. If it needs fewer steps to reach the result, the time and resources used are reduced.

According to OpenAI, the GPT-5.6 family delivers its highest intelligence efficiency per token to date. The company also says it now serves 1,000 million active users and more than 2 million companies, figures it uses to explain why reducing the cost of each response has become a priority.

The model is not the whole story

A significant share of the improvements comes from the infrastructure that runs the models. OpenAI optimized how it distributes requests across servers, organizes computations on chips and reuses information instead of processing it again.

GPT-5.6 Sol, running inside Codex, OpenAI's environment for working with code, helped analyze real-world traffic and identify imbalances in server workloads. It also wrote and improved code for GPUs, the specialized chips that run much of these computations.

OpenAI says these improvements reduced the total cost of serving the model by 20%. In another optimization, called speculative decoding, a small model proposes several words or fragments and the main model checks them all at once. The result, according to the company, was an improvement of more than 15% in token generation efficiency.

Sol also designed experiments to improve that auxiliary model and supervised part of its training, stepping in when hardware failures or stability problems appeared. OpenAI says it used verification tools to check that automatically generated code worked correctly.

Less wasted context in agents

Agents such as Codex and ChatGPT Work do not respond with a single model call. To solve a task, they may make dozens of requests and use external tools. Each step adds instructions, previous results and data from the environment.

OpenAI developed an orchestration layer written in Rust, a programming language, to control this process. The company calls it an agentic harness, or agent framework. Its job is to coordinate the model, tools and work environment.

Its changes include:

• Showing tools, plugins and integrations only when they are needed.
• Limiting tool responses to 10,000 tokens by default.
• Reusing instructions and results that have already been processed.
• Keeping the history in a stable order to take advantage of the cache, which avoids repeating calculations.

This matters because one extra second on every request adds up quickly. In a task with 30 model calls, those small delays can become an additional half-minute, while also increasing the cost.

For you, the most visible change should be a combination of lower prices, greater speed and broader availability, especially for long tasks or tasks involving many tools. But the figures come from tests and internal measurements reported by OpenAI, so we will need to see how the models perform in daily use and against other alternatives.

The direction set by GPT-5.6 is clear: AI progress no longer depends only on making models more intelligent. It also depends on running them better, wasting less compute and turning each small improvement into a cheaper, more capable system.