The AI pricing unit for agencies that survives a price cut

🇹🇷 Türkçe: Bu yazının Türkçesini oku →

Two 1,500-word briefings for a client last week metered at fifteen cents on the Anthropic API. On the invoice, they will not show up as fifteen cents. They will show up as 0.30 units of Token Usage (Opus). The gap between those two numbers — dollars on one side, units on the other — is what this piece is about. Once you understand why the line item on your invoice is denominated in a unit rather than a currency, the whole billing model makes sense.

The four wrong pricing bases from the previous piece all had one thing in common. They measured the session. Sessions are shaped by how much re-reading, retrying and backtracking the model happens to do on the day, which is nearly random from the client’s point of view. They tell you almost nothing about what actually left the building. The unit we settled on measures the artefact — the piece of work that reached the client — not the process behind it. That is a different number, and a much stabler one.

If you are the person paying the bill, there is one specific reason this is worth reading. In November 2025 Anthropic cut the price of its Opus model by roughly two-thirds, and our invoice line item did not change. That is the whole argument for a per-unit denomination rather than a dollar figure, and it is what you should be looking for the next time an agency proposes to put an AI line on your invoice.

The base is the artefact you received, not the session that produced it

The clearest example is the one that started this whole exercise. Three internal sessions last month metered at $161 on the Anthropic API — a number that looked plausible until we opened the transcripts and found a single session had done 110 million cache-read tokens, most of them the model re-reading its own earlier context. Two short briefings had left the building. Everything else was the machine talking to itself. Billing you for that is billing you for our tooling, not our work.

So the rule we live by is short: research, retries, backtracking and context re-reads stay on our side of the ledger. What you pay for is what you can hold in your hand — a draft, a spec, a rewrite, a comparison table, an analysis. If the model spent six minutes writing your paragraph and forty minutes thinking about it, you get charged for the six.

Every current model prices output at five times input, and that collapses two numbers into one

This is the arithmetic that made the whole scheme tractable. Verified today at claude.com/pricing:

• Opus 5: $5 per million input tokens, $25 per million output tokens
• Sonnet 5: $2 input, $10 output
• Haiku 4.5: $1 input, $5 output

Output is priced at exactly five times input on every tier, and that ratio has held across every generation we have tracked. It lets one useful shortcut fall out of the arithmetic for free. Define:

weighted tokens = input + (5 × output)
1 unit          = 100,000 weighted tokens

The two-variable problem (input separate from output) becomes one number that you can multiply directly by the tier’s input price. One unit on Opus 5 costs $5 per million weighted × 100,000 = $0.50 raw. On Sonnet, $0.20. On Haiku, $0.10. Margin sits on top at 2×, and we say so in the open rather than baking it into the number and hoping nobody asks. Cost × 2 = billed. On Opus at today’s list, one unit reads $1 on the invoice.

A worked example, because this is the paragraph a CFO will re-read. A 1,500-word briefing is roughly 5,000 input tokens (research, sources, the brief itself) plus 2,000 output tokens (the piece coming back). Weighted: 5,000 + 5 × 2,000 = 15,000. In units: 0.15. On Opus at list, that is $0.075 raw and $0.15 billed. Two briefings on the same afternoon: 0.30 units, $0.15 billed. Fifteen cents for the AI, and whatever human hours went into the review sit on their own line above.

The unit is what survives when the model gets cheaper

This is the reason the whole model is worth the effort. On May 22, 2025 Anthropic launched Opus 4 at $15 per million input tokens and $75 per million output tokens. Six months later, on November 24, 2025, Opus 4.5 arrived at $5 and $25 — a three-times drop across both dimensions. Opus 5 launched in July 2026 and kept that pricing. That is one major model release and two Opus generations in eighteen months, with pricing that ran on a rollercoaster nobody sitting at the client’s end could have predicted.

If our invoices had been quoted in dollars per session, or dollars per artefact, or dollars per token, we would have had to reprice the whole system twice in nine months and send you the paperwork. Our unit did not move. What moved was the dollar cost inside the unit. At the old $15/$75 pricing, one unit on Opus was $1.50 raw and $3 billed. At today’s pricing it is $0.50 raw and $1 billed. The invoice format survived that shift without a renegotiation, without a covering email that read “we are reducing your AI line by two-thirds starting next quarter”, without any of the awkward conversations a per-dollar model creates every time the frontier lab moves.

The unit also survives us moving work between tiers. Draft on Haiku, refine on Sonnet, final read on Opus — every step is counted in the same unit, and the invoice reader sees which tier each line came off. If we decide next month that a whole class of jobs runs on Haiku end-to-end, the line item is denominated in the same currency; only the tier changes. Your finance team is not learning a new taxonomy every quarter.

The human translation your finance team actually asked for

The unit works arithmetically. On its own, it does not land. When we sent the first invoice with a line reading “34 units of Token Usage (Opus)”, the reply that came back was a polite version of “translate that”. So we did. We measured actual token throughput against session time over about three weeks and the range came out at 5,000 to 11,000 weighted tokens per minute, averaging 8,000. From that:

• 100,000 weighted tokens ≈ 12–13 minutes of active AI work
• 1 hour of AI work ≈ 480,000 weighted tokens ≈ 4.8 units
• 1 unit ≈ 10–15 minutes of AI activity ≈ ~$1 billed on Opus at today’s list

Round numbers for a client conversation: four units to the hour, about a dollar a unit on Opus. A one-hour AI session reads about four dollars on the invoice. The same hour of senior consultant time on the same project reads roughly ninety dollars on the human-time line above it. Both figures sit on the same document, in the same currency your accounts payable process already handles.

The reason to name both figures in the same breath is not to argue that AI is cheap. It is to make the substitution question honest. When you ask “why did you use AI on this?”, the invoice is showing the two costs in comparable units, and the answer is “because this particular sub-task ran as four minutes of AI plus twenty of my review, not two hours of me from scratch.” That is a conversation your procurement lead can have; “we saved you time, trust us” is not.

What it takes to put this on a real invoice

The build section, because this is a piece about a delivered system, not a piece about a theory. If you are an agency owner reading this and thinking about doing the same thing, five things bite.

The measurement job has to de-duplicate on message.id. Session transcripts on the Anthropic Claude Code side contain repeated message records — 8,355 duplicates in 19,181 usage-bearing lines in the sample we audited. Every earlier number of ours was roughly two-times inflated until we added a distinct-on read at the collection layer. Skip this step and every figure you quote a client will be wrong by a factor of two in the direction that costs you money. This is the failure that surfaced the other three in Part 2.

The catalogue is one row per model tier plus one catch-all. Three token-usage rows on the invoice — Opus, Sonnet, Haiku — each priced against its own base rate. One “External API Cost” row for everything else (image generation, transcription, search APIs, embeddings), denominated at the provider’s own billing unit at 2× margin. This discipline keeps the catalogue from fracturing every time a new service arrives; the alternative is inventing a new job type per vendor and drowning in bookkeeping.

The provider’s unit stays intact. If Anthropic bills us in tokens, your row is in tokens. If Replicate bills us in GPU-seconds, your row is in GPU-seconds. Do not translate provider units into hours or credits or house points at the invoice level. That is an argument you will have every time a rate sheet updates, and it obscures the one property of this billing model worth defending — that the unit is a real, external quantity, not an agency abstraction.

Precision is coupled to the rate on the job type. The PM tool we run through — ActiveCollab — silently truncates the time-record value field to two decimal places. If the hourly rate on the job type is $3, effective precision is $0.03 per record. If the rate is $30, precision is $0.30, and small billable records round to zero and vanish. This is not documented anywhere; we found it by watching sub-cent items disappear on invoices. Pick the rate to match the precision you need, not the other way round.

The first invoice contains estimates. The second one is measured. Where a new line item has never been billed before, the cost figure on the first invoice is the provider’s list price, marked as an estimate. On the next cycle it is corrected against the actual metered spend. Say this to the client in writing on the first bill; it is a far smaller conversation than a downward adjustment they were not expecting.

Three questions to put to any supplier who uses AI on your work

The three sentences a buyer should be able to say to any AI-using agency, without embarrassing themselves, and get a straight answer back:

1. What is your billable unit, and how is it derived from something I can verify against the model provider?
2. What margin multiplier do you apply on top of provider cost, stated in the open?
3. What happens on my invoice when the underlying model gets cheaper?

If the answers you get back are “we absorb it into a fixed retainer”, “there isn’t a multiplier, it is just our rate”, and “nothing changes on the invoice”, you are paying for AI capacity someone else is already getting the benefit of. That may still be a rational commercial choice for a small, predictable engagement. It stops being one at any scale where the AI line is a meaningful share of the bill, which is where most enterprise work is heading in the next twelve months.

Part 4 of this series walks through what the full invoice looks like end-to-end — human hours, three tiers of tokens, external API costs, each on its own row, all on one document your finance team can process without a covering email. If you are on the buying side and want to see how the invoice reads before it lands on your desk, that is where the series is going next. If you are on the operating side of an agency and want to talk about how a system like this would work against your own tooling, get in touch and we will walk you through what we run.