OpenAI Cut Its Cheapest Model 80% Three Weeks After Shipping It

OpenAI cut the API price of GPT-5.6 Luna by 80% on July 30. Input went from $1 to $0.20 per million tokens; output from $6 to $1.20. Combined, a million tokens through Luna fell from $7 to $1.40. GPT-5.6 Terra got a smaller trim — 20%, from $17.50 to $14 combined, at $2 in / $12 out.

Sam Altman announced it on X as "major price cuts today." OpenAI's stated reason was efficiency: gains realized during GPT-5.6's own development, including the model rewriting and optimizing production inference code and improving token generation throughput.

That explanation is probably true. It is also incomplete in a specific way.

The timing is the tell

The GPT-5.6 family became broadly available on July 9. The price cut landed three weeks later.

Three weeks is not an efficiency timeline. Inference optimization of the kind that justifies a five-fold price reduction is a quarters-long exercise — kernel work, batching changes, serving-stack rewrites, hardware allocation. Labs bank those gains as margin and pass them through on their own schedule, usually at the next model launch, because a price cut is the one lever you cannot un-pull.

Pulling it 21 days after launch means the launch price was already wrong by the time the market saw it. Something in the interval revised OpenAI's read of what the floor is.

What moved in the interval

The thing that moved was the share of inference not running on American models.

A CNBC analysis published July 7 put Chinese-origin models at roughly 46% of US enterprise token usage on OpenRouter, at points crossing above US-origin models on that platform. OpenRouter is not the whole market — it skews toward developers who route by price and benchmark, exactly the cohort most willing to switch — but it is the closest thing to a live ticker on where marginal tokens go, and it is the one number frontier labs cannot spin.

Then Moonshot AI shipped Kimi K3, a 2.8-trillion-parameter open-weight model, at $3 in / $15 out. Kimi K3's demand spike was severe enough that Moonshot paused new subscriptions. DeepSeek continues to sit below both on input pricing.

Against that field, Luna at $1/$6 was not a value tier. It was a mid-tier price wearing a value-tier name. At $0.20/$1.20 it undercuts DeepSeek on input — the first time OpenAI has priced beneath a Chinese frontier model on any axis that matters to a cost-sensitive buyer — while remaining more expensive on output.

Note which model got the deep cut. Terra, the higher-capability tier, moved 20%. Luna, the volume tier, moved 80%. Cuts are not distributed by efficiency; they are distributed by where the competition is. The pressure is entirely on the cheap end.

Pricing power was the thesis

For three years the working assumption underneath every frontier-lab valuation was that capability compounds into pricing power. You spend billions on training, you get a model nobody else has, and you charge accordingly. Inference costs fall over time, but they fall against a price you control, and the spread is the business.

That assumption survives only if the capability gap is wide enough that switching is irrational. What the last six months have demonstrated is that for the majority of production workloads — classification, extraction, summarization, routing, the unglamorous volume that actually generates token spend — the gap closed enough that switching became a spreadsheet exercise.

Once it is a spreadsheet exercise, the cheapest adequate model sets the price. Not the best model. The cheapest adequate one. And adequate is now being defined by open-weight releases from labs with lower capital costs, lower margin expectations, and a strategic interest in commoditizing exactly the layer OpenAI monetizes.

An 80% cut three weeks post-launch is what it looks like when a company recognizes that in public.

What this does to everyone else

Anthropic, Google, and every inference provider reselling open weights now face a repriced floor. Google has the most room — TPU economics and a search business that subsidizes the API. Anthropic has the least room on price and the most differentiated position on the workloads where output quality is worth paying for, which is the same barbell Microsoft described at Build: own the volume cheaply, pay up only at the edge.

For anyone building on these APIs, the practical read is narrower and more useful:

Your unit economics just improved and you should not assume they are stable. A workload that was marginal at $7 per million tokens is comfortably profitable at $1.40. But a price that can fall 80% in three weeks can be restructured in three weeks too — rate limits, tier gating, context-window pricing, priority queues. Cheap tokens are being used to buy position, and position, once held, gets monetized.

Model-portability is now a first-class engineering requirement. If your stack cannot swap providers behind a stable interface, you are paying a tax measured against a floor that is moving 80% at a time. The labs are telling you, through their own pricing behavior, that they expect you to be able to leave.

Volume tiers are where the war is. Capability tiers are still differentiated and still expensive; Terra only moved 20% for a reason. If your product's margin depends on frontier-tier calls, none of this helped you much.

The read

OpenAI framed the cut as passing along efficiency. Efficiency is real, and OpenAI has more of it than almost anyone.

But labs do not cut the volume tier by 80% and the capability tier by 20% because their serving stack got 80% better at one and 20% better at the other. They do it because the volume tier is the one being contested, and the contest is being run by models whose weights are public and whose price floor is set by whoever will host them cheapest.

The frontier still commands a premium. The floor underneath it no longer belongs to the labs that built it.