OpenAI Used GPT-5.6 Sol to Rewrite Its Own Kernels, Cutting Serving Costs by 20 Percent

OpenAI has revealed that its newest model, GPT-5.6 Sol, was used to optimize the infrastructure it runs on. Working through Codex, the model independently rewrote and improved the production kernels responsible for the math operations performed during inference, translating into a measurable reduction in the cost of running the service.

A model that improves itself

Kernels are the lowest-level code that runs the calculations on GPUs as a language model generates a response. Optimizing them is normally the job of specialized engineers, since small changes to this code can significantly affect the performance and cost of every query. OpenAI said that GPT-5.6 Sol, operating inside Codex, rewrote these code segments on its own, having been trained to work well with Triton and Gluon, two open-source GPU programming languages that the company develops and maintains itself.

The result wasn't a one-off experiment but a set of changes deployed to the production infrastructure that serves queries to OpenAI's models. The company said total serving costs fell by 20 percent, while token generation throughput rose by 15 percent thanks to improvements in speculative decoding, a technique in which a smaller draft model predicts upcoming tokens and a larger model merely verifies them.

What else got improved

Beyond the kernels themselves, OpenAI optimized handling of the KV cache and the systems that distribute workloads across GPUs, allowing more queries to be served on the same hardware. The company also changed how its agents operate: Model Context Protocol servers and additional skills are now invoked only when actually needed, the default limit on data returned by tools was capped at 10,000 tokens, and the order of tool calls was optimized for better cache utilization.

All these improvements, from routing to task planning to the implementation of the model itself, were built with significant input from GPT-5.6 Sol running inside Codex. According to accounts of the work, the model itself identified which pieces of code were worth fixing and then implemented the changes, which OpenAI engineers reviewed before deploying them to production.

Three variants of one model

GPT-5.6 rolled out in limited preview in the API and Codex in three variants that differ in price and capability. Sol, the flagship model, costs $5 per million input tokens and $30 per million output tokens, exactly the same as the previous GPT-5.5. Terra, the mid-tier variant, is priced at $2.5 per million input tokens and $15 per million output tokens, in line with the older GPT-5.4. Luna, the cheapest and fastest of the three, costs $1 per million input tokens and $6 per million output tokens, which OpenAI describes as an entirely new, cheaper price tier between GPT-5.4 and its mini version.

This setup lets companies using the API route harder queries to Sol while handing simpler tasks to Luna, giving them, in practice, flexibility to control inference spending depending on the scale and nature of their traffic.

Why it matters

Until now, the race among AI labs has focused mainly on which model scores better on benchmarks. The GPT-5.6 Sol example points to a different front in that rivalry: using AI to cut the cost of running the AI infrastructure itself, which matters at a time when tech companies' spending on data centers and compute is growing faster than revenue from AI services.

For companies using OpenAI's API, the practical upshot is lower real-world costs for processing large volumes of queries at the same nominal per-token price. For OpenAI itself, under pressure from increasingly fast-moving Chinese AI labs, it's an argument that efficiency has become just as important a competitive front as the raw quality of a model's answers.

For Polish companies and developers using GPT-5.6 through the API, it's mainly that second point that matters: at high query volumes, even a low double-digit percentage cut in inference costs has a real impact on the budgets of projects built on this model, even though the per-token prices themselves haven't changed for now.