AI DevelopmentMethod6 min readPublished September 20, 2026

3 terms · 3 designs · 3 failure rates · published rates only · the bill is mostly the checking, sometimes

When Checking AI Output Costs More Than Generating It

A cost formula for AI jobs at volume: generation, verification and rework. Worked at three failure rates and three checking designs on published 2026 rates.

DA
Digital Applied Team
Research and practical guidance
Editorial dateSeptember 20, 2026
Rates readSeptember 22, 2026

Most estimates for an AI job at volume are one multiplication: a token price times a token count. The bill has two more terms. Every unit has to be checked, by a model or a person, and every unit that fails the check has to be made again and checked again. Whether those two terms are a rounding error or most of the cost depends on one thing you control, the checking design, and one thing you have to measure, the failure rate.

This post defines the three-term model and works it through for a single unit of writing at three failure rates and three checking designs, using Anthropic's published per-million-token rates as read on September 22, 2026. It is arithmetic over a price list. We did not survey anyone, test any model, or measure any failure rate; the rates are quoted with their date and the failure rates are chosen to show the shape of the curve.

Key takeaways
  1. 01
    With a cheaper model doing the checking, generation dominates until failures pass 62.5%.In the worked example, a Sonnet 5 check on Opus 5 output costs under a quarter of the generation. Checking plus rework only overtakes generation at a failure rate no one would ship at.
  2. 02
    With the same model checking its own class of output, the crossover is 26.8%.A same-model verification pass costs more than half of generation on its own. Add rework and checking is the larger term at any failure rate above about a quarter.
  3. 03
    With a person checking, checking is the bill at any failure rate.Three minutes of review at $60 an hour is $3.00 against $0.065 of generation. Verification and rework are 98% of the total whether 5% or 30% of units fail.
  4. 04
    The crossover moves with the checking design, not the model price.Halving the generation price changes the totals but barely moves the crossover for a human design, because the human term does not scale with tokens. Decide how you check before you price the job.

01The modelThe formula, every term defined

For one unit of work, the expected cost is generation plus verification plus rework. Generation is the input tokens times the input rate plus the output tokens times the output rate, for the model that makes the unit. Verification is whatever it costs to decide whether the unit is acceptable: a second model reading the brief and the output and writing a verdict, a person spending a stated number of minutes at a stated rate, or both. Rework is the failure rate times the cost of generating and verifying again. The model assumes one retry; a unit that fails twice costs more, so the numbers below are a floor.

GenerationInput tokens × input rate + output tokens × output rate, for the generating model
G
VerificationA second model pass priced the same way, or minutes × an hourly rate, or both
V
ReworkFailure rate f × (G + V), assuming one retry that is generated and checked again
f × (G + V)
Expected cost per unitThe three terms summed; checking share is (V + rework) ÷ total
(G + V) × (1 + f)

The last line is the useful one. Because rework multiplies both earlier terms, the total is simply generation plus verification, scaled by one plus the failure rate. Checking plus rework overtakes generation exactly when the failure rate exceeds generation minus verification, divided by generation plus verification. That expression is the whole sensitivity analysis, and it says the crossover is fixed by the ratio of the two costs, not by the level of either.

02The numbersThe worked example

One unit is a piece of writing generated from a 3,000-token brief into 2,000 output tokens. Generation uses Claude Opus 5 at $5 per million input tokens and $25 per million output tokens, which is $0.065 per unit. Design A checks with Claude Sonnet 5 at $2 and $10, reading the brief and the output (5,000 tokens) and writing a 500-token verdict: $0.015. Design B does the same check with Opus 5: $0.0375. Design C is a person spending three minutes at $60 an hour: $3.00, an assumption stated here and not a measurement of anyone's review speed. All model rates are from Anthropic's pricing page as read on September 22, 2026, without caching or batch discounts.

Computed from Anthropic's published rates read September 22, 2026and the stated assumptions. Failure rates are illustrative, not measured. Model-only totals are rounded to the nearest tenth of a cent, totals including human review to the nearest cent.
Checking designFailure rateCost per unitChecking + rework share
A: cheaper model checks5%$0.08422.6%
A: cheaper model checks15%$0.09229.3%
A: cheaper model checks30%$0.10437.5%
B: same model checks5%$0.10839.6%
B: same model checks15%$0.11844.9%
B: same model checks30%$0.13351.2%
C: a person checks5%$3.2298.0%
C: a person checks15%$3.5298.2%
C: a person checks30%$3.9898.4%

Checking plus rework as a share of cost per unit, at a 15% failure rate

Computed from Anthropic's published rates read September 22, 2026; illustrative failure rate
A: cheaper model checksSonnet 5 verifies Opus 5
29.3%
B: same model checksOpus 5 verifies Opus 5
44.9%
C: a person checks3 minutes at $60/hour
98.2%

03The lineThe sensitivity line

Apply the crossover expression to each design and you get one number per design: the failure rate above which checking and rework cost more than generation.

Design A
Cheaper model checks
62.5%failure rate

Below this, generation is the larger term. No production job runs here, so with a cheap checker the generation price is what to optimise.

Generation dominates
Design B
Same model checks
26.8%failure rate

Whether a job sits above or below this is an empirical question about that job, not something this post answers. Above it, the checking pass is the cost to optimise, by making it cheaper or by lowering the failure rate it has to catch.

Depends on the task
Design C
A person checks
Noneat any rate

Verification costs 46 times generation before a single failure. The generation price is irrelevant to the bill; the review minutes are the bill.

Checking dominates

The premise this post set out to test was that checking, not generating, is the real cost of AI work. The honest result is narrower: it is true whenever a person is in the loop and false whenever a cheap model is, and the boundary in between is set by the ratio of the two costs. What the model price does is move the totals, not the boundary. A separate measurement of ours makes the same point about where the money is not, in our cost-per-task benchmark, where cache reads rather than fresh input held most of the tokens in seventeen of the eighteen runs; those measured figures are not reused here.

04The inputWhat makes the failure rate knowable

Every number above hangs on a failure rate, and a failure rate is not a feeling. It is the fraction of a sample that fails a written acceptance rule. To know yours, write the rule first, draw a random sample of finished units, have the checker you intend to use apply the rule, and have a person apply it to a subset of the same units so you also know how often the checker is wrong. Then you have three numbers: the failure rate, the checker's miss rate, and the checker's false-alarm rate. The first sets rework; the second sets what escapes to your customer; the third sets wasted rework.

The sample size does not need to be large to be useful. What it needs is to be random, drawn after the process is fixed, and repeated when anything upstream changes: the brief template, the model, the checker's prompt. A failure rate measured on last quarter's model is a historical fact, not an input. How we budget the same kind of job at fixed monthly spend is in our post on what a fixed AI budget buys; the harness overhead that sits on top of both terms is in our harness-cost post.

A design that changes the answer

Design A and design C are not the only choices. A cheap model check followed by human review of only the units the model flags, plus a small random sample of the ones it passed, keeps the human minutes proportional to the failure rate instead of the volume. In the worked example that turns a $3.00 term into about fifteen cents plus the sample at a 5% failure rate. The formula handles it: V becomes the model check plus the failure rate times the human minutes.

05The scopeWhat this is and is not

This is arithmetic over published prices on a stated date. It is not a benchmark of any model, not a measurement of any team's failure rate, and not a claim about which model fails more often. The unit sizes, the review minutes and the hourly rate are assumptions chosen to be typical, and every one is printed so you can replace it. The prices will change; the formula and the crossover expression will not. If you want the same model run on your own job with your own measured failure rate, our AI transformation service starts every volume engagement with exactly that sample.

06Next stepThe checking design sets the crossover; the failure rate decides which side you are on

Put it into practice

Price the check before the generation, then measure the failure rate on a real sample

Write the acceptance rule. Cost one unit of generation and one unit of each checking design you would actually use, at today's published rates. Compute the crossover for each. Then draw a random sample of real output and measure where you sit. If you are above the crossover, the money is in the checker; if below, it is in the generator.

Digital Applied

Cost your AI job on all three terms, not one.

We build the cost model for your job, run the sampled audit that gives you a real failure rate, and design the checking layer that keeps the human minutes proportional to the failures.

Three-term cost modelSampled failure auditChecking design
Your next project

Start with the sample

  • A written acceptance rule
  • A random sample of finished units
  • The failure rate and the checker's miss rate
Questions and answers

Applying this post

They are chosen to show the shape of the curve, not measured. This post attributes no failure rate to any model or task. Measure your own on a random sample against a written acceptance rule, and re-measure when the model or the brief changes.
Digital Applied newsletter

Deep dives on AI, marketing and development.

Practical guides and fresh insights by email. No recycled takes.

Related dispatches

Continue reading

AI Development

When Your Tool Picks the Model, Who Pays for Picking?

Claude Code 2.1.278 stopped charging for its auto-mode classifier where checks run server-side. What Claude Code, OpenRouter and Copilot bill for routing.

September 20, 2026 · 5 minRead
AI Development

Who Checks a Frontier AI Lab's Work? 12 Arrangements

A census of 12 external evaluation arrangements at Anthropic, OpenAI and Google DeepMind: who evaluates, who pays, what access they get, what gets published.

September 20, 2026 · 6 minRead
AI Development

Your CI Agent Can Publish Packages. Narrow That Token Now

npm added stage-only tokens on September 18, 2026 and targets January 2027 to end direct publishing by bypass-2FA tokens. Who moves to what, and what stays.

September 20, 2026 · 5 minRead
AI Development

Your Agent Pinned a Plugin to a Hash. It Got the Branch.

A September 17 advisory shows four coding agents resolving a pinned commit to a same-named branch. How git decides, which versions fix it and what a pin means.

September 20, 2026 · 6 minRead
AI Development

Cloudflare Blocks AI Agents on Ad Pages: Which Bots Are Hit

From September 15, 2026 new ad-supported Cloudflare domains block AI agents on ad pages and refuse AI training by default. A 20-bot census of who is affected.

September 15, 2026 · 10 minRead
AI Development

TabPFN 3.5 Beats Boosted Trees: When to Use It on Your Data

Prior Labs' TabPFN-3.5 report claims first place on seven tabular benchmarks. What a tabular foundation model is, when to use it, and what the licence allows.

September 15, 2026 · 8 minRead