The Price per Task Sheet – Mathew's Library

01 Price a task, not a token

AI vendors bill per million tokens. A token is a chunk of text, roughly three quarters of a word. So the price page tells you what a token costs, not what your work costs.

On September 22, 2026, Anthropic released Opus 5.5, 20% cheaper per token than Opus 5. Artificial Analysis, a benchmark firm, priced the same test tasks on both: $5.98 a task on the new model, $5.86 on the old one. At maximum effort the new model used more tokens per task, so the cut per token left the cost per task about level.

It's one benchmark, and not your tasks. Others argue caching makes real agent bills drop by more. The point stands: a price cut per token isn't automatically a price cut per task. If you pay a flat seat you never see a token, so this matters most if you pay through the API or run agents.

02 The price-per-task table

Take ten real tasks, say a landing page draft or a ticket reply. Run each on two models. Read the cost off the usage page of your account with the vendor. Divide the cost by the number of tasks. That's your price per task. Write the result on one line, and do it again whenever a model changes: the same task can use more tokens on the new one.

One line per task

Task: [WHAT IT IS, e.g. ticket reply]
Model: [MODEL NAME]
Cost per task: [$ FROM THE USAGE PAGE / NUMBER OF TASKS]
Passed my check: [YES / NO, and why]
Tested on: [DATE]

03 A worked example

A made-up support reply reads 20,000 tokens of ticket history and writes 2,000 back. At Anthropic's list prices:

Model Price per million tokens (in / out) Cost of one reply

Haiku 5.5 $0.10 / $0.50 About a third of a cent

Sonnet 5.5 $2 / $10 6 cents (4 in, 2 out)

Opus 5.5 $4 / $20 12 cents

Same price page, up to forty times apart. If Haiku passes your check on that task, the tier you pick matters far more than a 20% cut. If it doesn't pass, the cheap price is the expensive one, because somebody redoes the work.

04 Pick the tier, then commit carefully

Routine work on the cheapest tier that passes your own check. Your one-line records tell you which tier that is.

Work that can wait goes through the Batch API. Anthropic's Batch API takes 50% off for it.

Careful with long commitments at today's list price. Inside 2026 prices moved down at the middle and the bottom and up at the top, so a one-year lock is a bet on one tier.

The model name goes in one setting, not forty scripts. Ramp says businesses are switching providers more often than before; make switching cheap.

05 The vendor questions

An adjusted number is one where the company chooses which costs to leave out. That isn't fake. It's a choice. Picture a restaurant that quotes its margin on the plate and leaves out the kitchen: the plate margin can be true, it just isn't the whole cost of dinner. When a vendor says "profitable on an adjusted basis", ask:

What's left out? In September the Financial Times reported Anthropic's margin was before the cut partners like Amazon take, and before the cost of training models.

Is stock pay in it? One analysis says Anthropic's adjusted operating margin leaves out stock pay.

Is it audited and published? Anthropic's third-quarter numbers were neither; the unadjusted number comes when the filing is public.

What does it cost per finished task? Ask for your own tasks, not the price per million tokens.

Sources

Anthropic: Claude pricing Artificial Analysis: Opus 5.5 cost per task Ramp: Token pricing power falls Quartz: Anthropic told investors Q3 revenue reached $16.8 billion Investing.com: Anthropic Q3, harder questions

Mathew's LibraryLibrary card

Free access

Keep reading free with your library card

One email opens this template and all 86 resources in Mathew's library. No account, no password.