Everyone Thinks Token Prices Will Rise. I Think They Will Fall.

“Token prices are going to rise.”

You hear that sentence a lot right now, as LLM providers move from generous flat-rate subscriptions toward token-level billing.

The obvious thesis is this:

If models become better, create more value, and are used by more people, token prices should rise.

But you can also build the opposite narrative.

Many models are now surprisingly close to each other. That creates commodity effects. When the differences become smaller, price competition begins.

This could become especially interesting at the frontier. For example, if Anthropic’s Fable 5 clearly pulls ahead of future OpenAI models like GPT-5.6 or GPT-6, OpenAI would have a strong reason to defend market share aggressively through pricing. There have already been early hints in that direction on X and in the Wall Street Journal.

Then there is the pressure from China. Models like MiniMax M3 do not quite reach the top performance of the best frontier models, but they cost only a fraction: $1.20 per million tokens compared with $25 for Opus 4.8.

At the same time, massive amounts of capital are flowing into data centers for AI inference. More supply puts pressure on prices.

The counterargument remains strong, though.

So far, only very few people use AI at least once a day. Once the late majority arrives, demand could rise faster than new data centers can be built.

Modern agent harnesses are also much more token-hungry than chatbots. An agentic workflow reads, plans, calls tools, checks, corrects, and repeats. That consumes context.

I am not sure. But for Q3 2026, I would currently bet more on falling token prices and a price war between OpenAI and Anthropic.