Every major OpenAI price cut gets covered as a gift. GPT-4o gets cheaper, the write-ups celebrate accessibility, and the implication is that the company is generously passing along efficiency gains. That framing isn’t wrong, exactly - the prices do fall - but it keeps missing what’s structurally happening underneath.
OpenAI has cut API prices multiple times since GPT-4 launched. The pattern is consistent: a new model arrives, it’s priced high, a cheaper variant follows, then the previous generation drops. This looks like normal tech commoditization. But the sequencing matters more than the absolute numbers. Every price cut happens after a meaningful cohort of developers has already built on the current tier, integrated the SDK, trained their support docs around a specific model’s behavior, and shipped to production.
Switching costs in AI APIs are underappreciated. It isn’t just swapping an endpoint. Prompts that work well on one model often degrade on another, even a supposedly improved successor. Output formatting, reasoning behavior, refusal patterns - these differ enough that a production app can break in subtle ways after a model swap. Developers who’ve tuned against GPT-4o aren’t eagerly jumping to a competitor’s API the moment prices diverge by 20 percent.
The Margin Play Nobody Wants to Say Out Loud
OpenAI’s inference costs have dropped substantially as hardware and batching efficiency improved. When the company cuts prices by 50 percent but its own costs dropped by 70 percent, that’s a margin expansion event dressed as a discount. The developer gets a lower bill; OpenAI gets a higher margin on a stickier customer base. This isn’t predatory - it’s just how platform economics work - but it deserves to be called what it is.
Anthropic and Google are running similar playbooks. Claude’s API pricing has shifted multiple times. Gemini has tiered aggressively. None of these companies is losing money to be kind to startups.
What This Means for Anyone Building on These APIs
The practical implication is less interesting than it sounds: don’t build as if the current price is permanent, and don’t build as if it will keep dropping in your favor. The floor on inference pricing will settle somewhere that makes the hyperscaler comfortable, not somewhere that makes your per-query margin comfortable.
Developers optimizing purely for today’s token cost are solving the wrong problem. The more durable question is how much of your application logic you want sitting inside a model you don’t control, priced by a company whose cost structure you can’t audit.