TL;DR

OpenAI has reduced developer pricing on its frontier model by over a fifth, for a three-month window, citing pressure from Anthropic and from Chinese labs. Output tokens take the deeper cut, falling by a third. Consumer subscriptions are untouched.

What actually changed

The reduction applies to API usage and to credits on the company’s agentic product and its coding tool, not to Pro, Plus or Business subscriptions. On standard short-context work, input pricing drops from $5 to $4 per million tokens, while output falls from $30 to $20 — a third off the side of the ledger that dominates most real workloads.

That asymmetry matters. Applications that summarise, draft or generate code consume output tokens far faster than input, so the effective saving for a typical deployment sits closer to the headline third than to the advertised fifth.

It follows cuts last month further down the range, where the mid-tier model came down by 20% and the cheapest by 80%. The pattern is a price ladder being compressed from below.

The competitive context

Anthropic’s comparable pricing gives the reason. Per million tokens, Opus 5 is published at $5 in and $25 out, with the frontier Fable 5 at double that again — $10 and $50. Chinese open-weight models continue to close the capability gap at a fraction of either.

Looking forward

For UK firms budgeting AI deployments, the useful detail is the expiry date. A three-month promotional window is not a basis for a business case, and any model of unit costs built on it should assume reversion.

The more interesting reading comes from pairing this with the other pricing story today: Nvidia has told large customers that server prices will rise above 15% as memory costs climb. Inference is getting cheaper to buy while the hardware producing it gets more expensive to own. Those two lines cannot diverge indefinitely — either the labs are absorbing margin to hold market share, or falling token prices are being subsidised by capital that eventually wants a return.

Either way, UK buyers should treat current frontier pricing as a competitive moment rather than a structural trend, and avoid designing anything whose economics only work at today’s rate card.