On 21 August, OpenAI cut the price of its most powerful model by more than 20%, to around $4 per million input tokens and $20 per million output tokens, with a promotional rate announced as running until at least 21 November. Google has also adjusted its pricing, and Meta keeps releasing models positioned on price. After the fast models, the cuts have now reached the high end.
We have been tracking these moves for months, with directions that have sometimes been contradictory. This one is significant because it hits the segment that had held out.
Why the high end is dropping now
Until now, the price war was playing out on the fast models. We have covered an 80% cut on a budget model, a Chinese model at 14 cents, and releases positioned on value for money.
The high end had escaped this pressure, for a simple reason: a user who needs the best available reasoning has no credible alternative, so they pay.
That situation is changing because credible alternatives are appearing. We have seen a model match the best of a competitor on an independent benchmark, and open models demonstrate serious capabilities, including in cybersecurity. When the top stops being a monopoly, it stops being protected.
Meta's strategy, different and interesting
One player deserves a special mention, because its positioning is explicit. Meta keeps releasing code-oriented models at a sustained pace, betting on price rather than peak performance.
The bet is this: for everyday work, a developer will prefer a slightly less capable model that is significantly cheaper over a top-tier model billed several times more.
That bet is consistent with what we observe in real usage. Most tasks do not require the best model in the world, and routing logic — sending simple tasks to a budget model and escalating only when necessary — has become standard practice.
A cut announced as promotional until a specific date is, by definition, temporary. This summer we saw a promotion end with a 50% increase and a Chinese model raise prices by 93%. A price reflects a commercial strategy as much as a production cost. The recommendation does not change: measure your real cost per task, and keep the ability to switch providers.
What this says about the economics of the sector
Two readings coexist, and they are not mutually exclusive.
The technical reading: efficiency gains are real. We have documented them, from execution optimisations to mixture-of-experts architectures. Cutting prices while preserving margins is possible.
The competitive reading: when several players offer equivalent capabilities, price becomes the only differentiating criterion. That is the definition of a market commoditising.
This second reading raises a question that the upcoming IPOs will help settle: an industry whose core product is commoditising must find its profitability elsewhere. In infrastructure, in interfaces, in specialised use cases. No one knows where yet.
What to take away
For the user, the news is good and it should be said plainly: the best available model now costs noticeably less than it did three months ago, and this trend directly benefits those building things.
The open question is how far this can go. Serving a model genuinely costs electricity and hardware, and those costs do not fall to zero. At some point, the decline will stop, or someone will stop paying the difference. We will likely see which of the two in the coming quarters.