GPT-5.6 Just Optimised Itself. AI Costs Now Fall On Their Own.

On 29 July 2026, OpenAI published something more consequential than another price cut. It revealed that GPT-5.6 Sol optimised its own serving stack, autonomously rewriting production code to cut end-to-end serving costs by 20% and improve token-generation efficiency by more than 15%. Those gains fed straight into the 80% price drop on its cheapest model, Luna.

For growth leaders, this is the story behind falling AI prices. Costs are no longer dropping because of one-off hardware wins. They are dropping because the models are now improving the systems that run them, and that loop compounds.

What actually happened

OpenAI ran GPT-5.6 Sol inside Codex to analyse its own production traffic, then let it rewrite the code that serves every request. The detail is in the official engineering write-up on GPT-5.6 efficiency.

The model tackled load balancing, kernel optimisation, and its own speculative decoding draft model. It ran hundreds of experiments, monitored training runs, and intervened when hardware failed. In plain terms, the AI did the unglamorous infrastructure work that used to eat engineering months.

Where the savings came from

Two levers did most of the heavy lifting, and both were driven by the model itself.

Kernel rewrites: 20% lower serving cost

GPT-5.6 Sol autonomously rewrote OpenAI's production kernels, the core code executing the model's maths on GPUs. It found work that could be precomputed, avoided, or parallelised, then validated correctness with an open-source floating-point sanitiser. The result was a 20% cut in end-to-end serving cost.

Speculative decoding: 15%+ token efficiency

The model redesigned its own smaller draft model, which proposes tokens for the primary model to verify in parallel. Better proposals mean more output tokens per expensive pass. That lifted token-generation efficiency by more than 15%.

Why this changes AI unit economics

Here is the shift I think matters most. Previous efficiency gains needed scarce human engineering time, so they arrived in slow, expensive steps. Now the model closes the loop: measure, rewrite, verify, repeat. That makes the cost curve bend faster and more predictably.

For anyone running GEO monitoring or marketing agents at scale, this is the difference between AI being a line item you ration and one you deploy freely.

What it means for growth teams

Table: How self-improving inference efficiency reshapes common marketing AI workflows. Source: interpretation of OpenAI's GPT-5.6 efficiency disclosure.

From my observation, the businesses that win here are not the ones chasing the cheapest token today. They are the ones redesigning workflows around the assumption that AI compute keeps getting cheaper. That directly enabled the 80% Luna price cut that pushes AI agents to the front of the funnel.

The concrete action to take

Rerun your AI cost model with 20% lower serving cost and 15% better token efficiency baked in, then extend that trend forward. Identify one workflow you currently ration on cost, whether that is AI search visibility tracking or content variant testing, and pilot it at full volume this quarter.

In my opinion, the era of treating AI spend as a fixed constraint is ending. When models optimise themselves, cheaper compute stops being an event and becomes the default direction of travel. Build for that.