The price war reaches the flagship — and frontier tokens start to look like a commodity

When the lab that set the high end cuts the high end by 80%, the whole curve moves. The frontier is now fought on two axes at once, and per-token cost is falling fast enough to change what is worth automating.

The tell this week was not a new model. It was a price. OpenAI cut GPT-5.6 Luna by 80% to $0.20 input, and DeepSeek shipped V4-Flash the next day at $0.14. An 80% cut is not a promotion; it is a repricing of frontier-tier inference as a commodity to be won on cost rather than a premium to be defended on quality.

Two contests, not one

Capability and price have split into separate races. Claude Opus 5 holds the capability crown, topping the Intelligence Index at 61 and the Agentic Index at 55.3, while OpenAI cuts Luna 80% and DeepSeek undercuts everyone. A buyer no longer picks a single best model; they pick a point on a capability-versus-cost curve — and the curve moves under them weekly.

This is what a commoditising input looks like from the inside. The open-weight labs set a price floor; the closed labs, rather than sitting above it, are now cutting toward it. The gap between a premium API bill and a cheap-and-capable alternative is closing from both directions at once, which is precisely the dynamic that turns a differentiated product into an interchangeable one.

Falling prices expand the market

The counterintuitive consequence is that commoditisation grows the pie even as it compresses the margin. Every drop in per-token cost pulls a new class of workloads across the line from too-expensive-to-automate to worth-automating. The labs are trading margin per token for volume of tokens — a bet that the market expands faster than the price falls.

For anyone building on these models, the strategic move is to assume the token is cheap and getting cheaper, and to put the durable value in the layer above it — the eval harness, the gateway, the product. The model is becoming the commodity. The thing you build around it is not.

LLM-Stats — AI updates today (August 2026) — latest model releases → · Fello AI — Best AI models in August 2026 →