OpenAI says its newest large language model, GPT-5.6 Sol, autonomously rewrote and optimized the GPU kernels that power OpenAI’s production inference stack, cutting end-to-end serving costs by 20%. The company detailed the work in a pair of blog posts this week, tying the efficiency gains directly to price cuts across the GPT-5.6 model family.
How Sol optimized its own infrastructure
According to OpenAI, Sol worked inside Codex on the kernels that execute the mathematical operations underlying the model — code written in Triton and Gluon, two open-source GPU programming languages OpenAI maintains. The model identified computation that could be skipped, parallelized, or precomputed, and rewrote the kernels accordingly.
Sol also worked on speculative decoding, the technique that lets a smaller “draft” model propose tokens for a larger model to verify. OpenAI says Sol designed and ran hundreds of experiments on the draft model’s architecture, then launched and monitored its training, autonomously intervening when problems came up. That effort lifted token-generation efficiency by more than 15%, directly speeding up inference. Separately, Sol analyzed production traffic patterns to improve how requests are routed and load-balanced across data centers.
Because model-written code carries correctness risk, OpenAI says it validated the kernels with FpSan, an open-source floating-point sanitizer the company built to catch numerical errors before they reach production.
Price cuts across the lineup
The savings underwrote cheaper ChatGPT API pricing. GPT-5.6 Luna, the fastest and cheapest tier, dropped 80%, to $0.20 per million input tokens and $1.20 per million output tokens. GPT-5.6 Terra fell 20%, to $2 per million input tokens and $12 per million output tokens — a price OpenAI says now matches GPT-5.5-level performance at half the previous cost. Sol’s own pricing is unchanged, though OpenAI added a “Fast mode” that runs up to 2.5 times faster at twice the standard price.
The cheaper API pricing follows OpenAI’s move this week to open free frontier-model access to about 100,000 researchers through 2027, part of a broader push to widen who can afford to build on its most capable models.