OpenAI’s o-series models, Anthropic’s Claude with extended thinking, Google’s Gemini with deep research mode - every major AI lab now ships at least one ‘reasoning’ model that visibly thinks before it responds. The user-facing pitch is capability: harder problems, better logic, fewer confident wrong answers. The part that gets less attention is what it costs to run them.
Reasoning models generate large volumes of internal tokens before producing output. Those tokens aren’t shown to the user, but they’re computed and billed. On OpenAI’s API, o3 has been priced meaningfully higher per token than GPT-4o. The gap isn’t cosmetic - for applications that run many queries or process long inputs, the cost differential can be significant. Developers building on these APIs have noted this publicly in forums and release notes, though exact figures shift as pricing evolves.
The interesting problem isn’t the cost itself. It’s who absorbs it.
The Product Layer Is Quietly Squeezed
Most consumer AI products - writing assistants, coding tools, research apps - charge users a flat monthly fee. That fee was modelled on the inference costs of standard models. When those products quietly upgrade their underlying model to a reasoning variant (sometimes without announcing it), the cost structure changes but the subscription price doesn’t, at least not immediately.
Some companies have handled this by gating reasoning behind higher tiers. Anthropic offers extended thinking as a feature of Claude’s higher-usage plans. OpenAI separates o-series access from standard ChatGPT. But plenty of smaller SaaS products built on top of these APIs don’t have that flexibility - they’re absorbing cost increases without a clean way to pass them through.
This creates quiet pressure that won’t show up in press releases. It’ll show up in runway calculations.
The Capability Trap
There’s also a softer problem. Reasoning models genuinely perform better on certain tasks - multi-step logic, code debugging, complex document analysis. Once a product ships with that capability, rolling it back to a cheaper model registers as a regression even if users never consciously clocked the improvement. You can’t un-ship it.
So companies find themselves locked into a cost structure that compounds as usage grows, with limited ability to retreat without the product feeling noticeably worse.
The AI industry has spent two years talking about capability benchmarks. The next conversation - already starting to happen in smaller Slack channels and engineering blogs - is about what it actually costs to keep these models on by default, and whether the business models built around them can hold.