Token Volume Is Up 25x and Mid-Tier Models Do 90% of the Job. Your Pricing Has to Move.

Key takeaways

• Token volume is reported up 25-fold, so per-token cost decisions now compound across far more traffic than when most pricing was set.
• Mid-tier models are reported at about 90% of flagship capability for roughly one-sixth of the cost.
• That ratio makes flagship-by-default an explicit purchase of the last 10% of capability at six times the price.
• Volume growth and unit price decline pull margin in opposite directions. Which one wins depends on your routing, not on the market.
• Most SaaS pricing was set against a token cost that no longer exists.

What the numbers say

Two figures do the work here. Token consumption is reported to have grown 25-fold, and mid-tier models are reported to deliver around 90% of flagship capability at one-sixth the cost. Read together they describe a pricing reckoning: providers are selling far more units at falling effective prices, and the premium tier's value proposition has narrowed to a thin band of capability.

For anyone building on top of these APIs, that is not industry news, it is a direct input to gross margin.

Why does a 25x volume increase change your pricing?

Because cost per unit only matters in proportion to units consumed. When your product made a handful of calls per session, a difference of a few cents per thousand calls was noise, and picking the best model regardless of price was rational. At 25 times the volume, that same per-unit difference is a line item large enough to move gross margin by points.

The practical consequence is that the model routing decision has migrated out of engineering and into finance, whether or not anyone told finance.

Is 90% of capability good enough?

That is the wrong question, and the useful original point here is that 90% is not a property of the model, it is a property of the task.

A model at 90% of flagship capability is effectively 100% on classification, extraction, routing, summarization of clean input, and most structured transformations. It is meaningfully below 100% on long multi-step reasoning, ambiguous instructions, and tasks where a single wrong answer is expensive to recover from.

So the honest framing is not flagship versus mid-tier. It is: what share of your calls are actually hard? For most products the answer is a small minority, and those are the only calls that justify a six times cost multiple. Routing by task difficulty captures most of the savings while protecting the outcomes that matter.

What does that do to gross margin?

An illustrative example, since your real numbers depend on your token shape. Say a $30 per month plan currently spends $9 per user on inference, for 70% gross margin. If 80% of those calls are routine and move to a model at one-sixth the cost, the inference line drops to roughly $3, and gross margin moves toward 90%.

Nothing about the product changed. No price increase, no churn risk, no new feature. The same work is just being bought at the correct tier. That is the cheapest margin improvement available to most AI products right now, and it is available exactly once per company, so it is worth doing carefully.

Should you cut your price because inference got cheaper?

Not automatically, and this is where a lot of founders get it wrong in both directions.

Falling inference cost is not a reason to cut price if your pricing is anchored to customer value rather than to your COGS. It is a reason to either expand margin, expand usage limits at the same price, or move upmarket into workloads that were previously uneconomic. Cutting price because cost fell hands the entire benefit to customers who were not asking for it.

The one case where price should move is a usage-based model where your price is explicitly a markup on tokens. There, a falling floor with a fixed markup means your customers will eventually notice, and a competitor will tell them.

What to do this quarter

1. 1Classify your calls into hard and routine. Sample 200 real requests, do not estimate.
2. 2Run the routine set against a mid-tier model and score it on your own evaluations, not on a public benchmark.
3. 3Re-model gross margin per user with a routed mix rather than a single model.
4. 4Decide deliberately whether the recovered margin becomes profit, higher limits, or a new tier. Do not let it drift.

Token volume grew 25x while the cheap option got good enough for most of the work, so the question is no longer which model is best, it is which model each call deserves. You can compare capability against cost per model in the Calcaas model leaderboard before you route anything.