Nobody will tell you how much you get

I wanted to know whether I was on the right plan. Simple question, I thought: measure what I actually use, then price it across the market.

The measuring part worked. Every agent harness writes session files with full token accounting, so a week of my own traffic came out of ~/.omp/agent/sessions and ~/.claude/projects in about three seconds.

The shopping part did not. Four of the vendors I checked will not tell you how much usage their plan includes. Not "it depends" — no number exists on the pricing page, in the comparison table, in the FAQ, or in the docs the FAQ links to.

One week, counted#

5.87 billion tokens across 16,442 requests, averaging 358,000 tokens each. Three of those seven days were near-idle, under 50 requests. One day was 8,611.

The number that reorganised everything: 97.2% of my traffic is cache reads. Not input, not output. Re-reading context the model has already seen, turn after turn, because that is what an agent loop does.

Once you see that, most published pricing becomes irrelevant. The headline input and output rates apply to 2.8% of what I send. The only price that decides my bill is the one nobody puts in the big font.

That is not a 2× spread. It is 167×, on the line item that is 97% of my volume.

What a subscription is actually worth#

My week at Anthropic's API list rate costs $4,028. A month of the same costs $17,262.

I pay $200.

At 25 billion tokens a month the subscription returns about 86× its price in list-rate value. I went looking for a cheaper plan and found that the one I already had was the cheapest thing on the market, by a distance, for this particular shape of work.

That was not the answer I expected and it is the main reason I wrote this down. The instinct to go shopping was right. The assumption that shopping would find something better was not.

The part where nobody quotes a number#

Here is what four vendors publish as their usage allowance.

In every case the base tier is unsized. xAI's Plus plan is described as "Shared AI usage pool". Meta publishes a five-hour prompt range for its cheapest tier only — a range five times wide — and no weekly figure at all. The multiples are multiples of nothing you can check.

I don't think this is an oversight. An allowance nobody can measure is an allowance you can quietly adjust. Several of the vendors who do publish hard numbers have revised them this year, and you can tell, because the old number was written down.

The honest ones are the ones competing on price. DeepSeek, OpenCode, CommandCode, Z.ai and Xiaomi all publish either a dollar allowance or a credit conversion table. That is enough to do arithmetic with.

Dollars beat credits#

Normalising everything to tokens per dollar, on a neutral workload with no cache reuse:

Every entry at the top is a gateway selling you dollars of API usage at a discount — four to eight times your money back, in a unit you can audit against a published rate card. Credit plans sit lower, because the vendor owns the exchange rate.

Two traps worth naming. The headline allowance on every pricing page assumes you use the cheap model: Z.ai's top tier drops from 2,091M tokens to 697M the moment you pick their flagship over their flash model. And per-model caps can swing the value of a single subscription by 40× — OpenCode's $10 plan gives DeepSeek an $80 allowance and Grok $15, which is why Grok appears at both ends of that table.

The number I couldn't verify, and then could#

Everything above rests on cache hit rates, and no vendor guarantees one. Anthropic uses explicit cache breakpoints; DeepSeek's is automatic with an undisclosed TTL. So my estimate for DeepSeek — $194 a month for my volume — carried a twenty-fold risk. If the cache missed, those 5.7 billion tokens reprice from $0.003 to $0.15 and the bill becomes $3,792.

Then people started posting their DeepSeek usage dashboards.

Four independent accounts, and the pattern is consistent: the heavier the use, the better the hit rate. Account C ran 3.7 billion tokens through 13,295 requests at 281,000 tokens each — the same shape as my own traffic, at two-thirds the volume — and paid $24.57 for the month.

Scaled to my 25 billion, C's observed rate works out to $165 a month. My off-peak estimate had been $194. The estimate was right and slightly pessimistic.

It also means the plans are marked up. Account C, buying direct, got 152M tokens per dollar. The best subscription in my table manages 41M per dollar on the same model.

What I took from it#

Measure first. The session files are already on your disk, the accounting is already in them, and three seconds of parsing told me more than a week of reading pricing pages.

Then find the line item that is 97% of your volume, because every workload has one and it is rarely the one the vendor leads with. For anything agentic it will be cached input, and the market's spread on that single number is three orders of magnitude.

And treat an unsized allowance as a red flag rather than a gap in the documentation. "15× Plus" is not a price. It's a number you can only discover after you've paid, and only by watching a bar move.