Two years ago, a GPT-4 API call cost roughly 10x what a comparable Claude or GPT-4o call costs today. Prices dropped fast, and most of us assumed they would keep dropping. That assumption might be wrong.
A growing body of evidence suggests that current AI pricing is artificially low - subsidized by venture capital to grab market share - and the correction is already starting.
The Subsidy Math
OpenAI, Anthropic, and Google have collectively burned through tens of billions building and running large language models. The standard playbook: price below cost, capture users, then raise prices once switching costs are high enough. It is the same strategy Uber used with rides and Amazon used with Prime.
The difference is timing. OpenAI is reportedly preparing for an IPO. Anthropic has raised over $13 billion. Investors who wrote those checks expect returns, and returns require margins. You do not build margins by selling API calls at or below the cost of the GPU time to run them.
Where Prices Could Move
Consumer subscription tiers are the most visible target. ChatGPT Plus is $20/month. Claude Pro is $20/month. Both give you access to models that cost real money to run per query, especially on long conversations or complex reasoning tasks. These prices have held steady for over a year while model capabilities (and compute costs) have increased.
API pricing tells a different story. Per-token costs have dropped for standard inference (running a prompt and getting a response). But the newer reasoning models - o1, o3, Claude with extended thinking - cost significantly more per query because they run multiple inference passes internally. As these become the default, average costs per useful output go up, not down.
There is also the infrastructure side. NVIDIA GPUs, data center power, and cooling are not getting cheaper. The efficiency gains from better model architectures have so far been eaten by making models bigger and more capable rather than by reducing the cost of existing capability levels.
What This Means for Your Budget
If you are building workflows around AI APIs, the smart move is to track your actual usage costs now and build in headroom. A 2-3x price increase on your most-used models within the next 12-18 months is not a wild scenario - it is what the business models require.
For individual users on $20/month plans, a jump to $30-40 would not be surprising. Google already tested this with its Gemini Advanced tier. OpenAI's Plus plan has stayed at $20 since launch in early 2023, which in inflation-adjusted terms means it has actually gotten cheaper.
The practical takeaway: do not architect your business around the assumption that AI will always cost what it costs right now. The current pricing reflects a land grab, not a steady state.