US inference prices fell nearly a quarter in a month. Your unit economics are stale.

Average prices for AI inference from leading US labs fell nearly 25% between mid-July and mid-August, according to Silicon Data analysis reported by the Financial Times ^[1]. If your cost model, your gross-margin deck, or your build-versus-buy memo was priced off mid-2026 rate cards, it is now overstating your inference bill by roughly a third ^[1].

The mechanics are specific rather than atmospheric. OpenAI moved on July 30, cutting two of the three GPT-5.6 tiers ^[2]: the mid-range Luna by 80% ^[3], and Terra by 20% ^[4]. Flagship Sol held its price, with OpenAI accelerating its performance options at the same price point ^[5]. Anthropic launched Claude Opus 5 at half the price of its predecessor, Fable 5 ^[6] - a deeper cut than Terra's but well short of Luna's ^[3].

The asymmetry is the tell. Luna now sells for one-fifth of its old price, while Terra keeps four-fifths of its own ^[2]. Cryptobriefing reads OpenAI's decision to hold Sol as a tiered strategy in which the best model defends margin and the cheaper tiers compete for volume ^[11]. That is the sane interpretation, and it also tells buyers where the negotiating room is: not at the frontier.

Crucially, this is not two US labs bidding each other down. The source attributes the pressure primarily to Chinese providers, with DeepSeek and Moonshot leading, offering models that match or closely approach US benchmarks at dramatically lower prices ^[7][8]. That matters for how long the discounting lasts. A domestic price war can end with a truce; a cost-structure competitor with different capital expectations does not negotiate.

Set the month against the decade. GPT-4-class capability cost more than $20 per million tokens in late 2022 and less than $1 by mid-2026, a decline of more than 95% in about three and a half years ^[9] - a trend rate of at least 57% a year ^[4]. One month at 25% off, if repeated, would compound to about a 97% annual decline ^[5]. Nobody should model that as the run rate, but it does say the past month was several times steeper than trend, and steepness is what breaks planning assumptions.

For buyers, this is straightforwardly good: cheaper inference removes a barrier to adoption and lets workloads that did not pencil out clear the bar ^[13]. For the sell side it is harder. US labs have spent billions on inference infrastructure justified by revenue projections that assumed particular price levels, and when prices fall faster than usage grows the return on that spending stretches ^[12]. Valuations built partly on the premise that inference stays a high-margin business need revising when revenue per query drops a quarter in a month ^[10].

What to watch: whether the next monthly read from Silicon Data shows the cut holding or partly clawed back; whether Sol's price survives the next DeepSeek or Moonshot release ^[5][7]; and whether token volumes at the labs grow faster than the price declines, which is the only way the infrastructure math closes ^[12]. On the buy side, check whether your vendor contracts pass cuts through automatically or require you to ask.