Skip to content
Primores AI wiki

Cost per Completed Task: Why AI Token Prices Mislead (2026)

Cost per Completed Task

TL;DR: A price list tells you what a million tokens cost. It doesn’t tell you what a finished job costs, because models need different numbers of tokens, steps and retries, and one model can cost ten times more at its highest effort setting than at its lowest. On Artificial Analysis’s index in November 2026, Claude Fable 5.2 at high effort ($8 / $40 per million tokens) cost $9.08 per task and scored 62, while Claude Opus 5.5 at max effort ($4 / $20) cost $10.04 and scored 59. Anthropic’s own guide says to “compare models on cost per completed task, not per token, because you pay for completed tasks.” Before you switch models to save money, price a sample of your own tasks, counting the failures.

⚠️ Read 2026-11-11. Two kinds of evidence, both limited. Anthropic’s figures come from its own benchmarks on its own models, which it labels “Anthropic-internal” and “directional, not guarantees”. Artificial Analysis’s figures are cost per attempted index task at list prices, shown next to a score; its method page was only partly read. Nobody has published cost-per-task data for ordinary small-business work such as support replies, invoices or content.

What it means

Cost per completed task is the total you spent on a type of job, divided by the number of jobs that passed your check:

cost per completed task = (tokens for every attempt, including failed ones and retries, each priced at its rate) ÷ (tasks that passed)

It differs from three numbers that look similar:

NumberWhat it leaves out
Price per million tokensHow many tokens the model needs for the job
Cost per requestThe other requests in the same task (an agent makes many)
Cost per attemptThe attempts that failed and had to be redone
Cost per completed taskNothing on the model’s bill. It still leaves out your people’s time checking the output

Why the price per token misleads

1. Effort level moves the cost of one model by ten times

Current models take an effort setting that controls how much they think, call tools and check their own work. The same model at the same token price lands in very different places. Artificial Analysis’s index, read 11 November 2026:

Claude Opus 5.5 at effortIndex scoreCost per index task
low35$0.98
medium (the API default)48$2.45
high52$3.46
xhigh57$6.05
max59$10.04

The defaults also differ by model. On Anthropic’s API, Claude Fable 5.2 and Sonnet 5.5 default to high, and Opus 5.5 and Haiku 5.5 default to medium. A test that runs two models “as they come” compares two different effort levels.

2. A pricier model can finish with less work

Three comparisons from the same index, same date:

ComparisonCheaper per tokenCheaper per task
Fable 5.2 at high (62, $9.08) vs Opus 5.5 at max (59, $10.04)Opus 5.5, at half the token priceFable 5.2, and it scored higher
Sonnet 5.5 at max (55, $10.02) vs Opus 5.5 at xhigh (57, $6.05)Sonnet 5.5, at half the token priceOpus 5.5, and it scored higher
Three settings that all score 55: GPT-6.1 Sol at max ($1.27), GPT-6 Astra at high ($3.24), Claude Sonnet 5.5 at max ($10.02)Sol and Sonnet list at the same $2 / $10Sol, by a factor of eight over Sonnet

Anthropic’s guide gives the reason: “A more capable model can finish a task with less work, which can offset much of the per-token premium.”

3. Lower effort is not always cheaper

On the same index, Gemini 4 Carbon cost $7.19 per task at low effort (score 54) and $6.59 at medium (score 58). The lower setting cost more and scored less. A run that fails or wanders still bills every token.

4. Failures and retries are part of the price

Anthropic: “The bill is decided by the tasks the cheaper model fails, because a failed task still bills its tokens, then the retry, then whatever the failure costs downstream.” Gartner’s analysts make the same point about open models, which “can be more verbose and retry more often”, so a saving of up to twenty times on paper becomes “closer to five or six times” per task (see glossary/open-weight-models).

Spend is also uneven across tasks. On a 20-problem research run in Anthropic’s guide, two problems carried 43% of the spend. Its advice: “Price the tail of your workload, not the median.”

5. Caching changes what a token costs

An agent resends its whole conversation on every step. With prompt caching, the repeated part is billed at 2.5% to 10% of the input price on Claude models. Anthropic re-priced its own benchmark runs without the discount: 3.1 to 3.3 times the cost on a coding benchmark, and 7.3 times on a small bug-triage agent. Menlo Ventures’ 2026 report works a similar example: at a mix of 3% fresh input and 2% output, an Opus 5.5 agent session costs about $0.70 per million tokens, against list prices of $4 and $20. Details on glossary/prompt-caching.

6. A token is not a fixed unit

The same text is about 30% more tokens on Claude 4.7 and later models than on earlier ones, because the tokenizer changed. Anthropic’s own note: a per-token comparison “makes the newer models look more expensive by construction.” See glossary/tokens.

What Anthropic measured on its own models

From the cost guide, on a subset of the SWE-bench Pro coding benchmark (463 problems), priced as a customer is billed:

Model and effortTasks solvedCost per solved task
Opus 5.5, medium (default)93.1%$0.22
Opus 5.5, high95.5%$0.30
Fable 5.2, low94.6%$0.39
Opus 5.5, xhigh96.5%$0.75
Fable 5.2, high (default)96.3%$1.39

On this benchmark the cheaper-per-token model was also cheaper per solved task at the same pass rate: Opus 5.5 at xhigh was “level with Fable 5.2 at high for a little over half the cost.” It is the opposite of the Artificial Analysis comparison above, on a different set of tasks. Which model is cheaper per task depends on the work.

On a research benchmark (DeepResearch Bench II) the frontier model bought nothing: “Claude Fable 5.2 cost 4.3 times as much as Claude Sonnet 5.5 for no measurable gain in score, and 2.9 times as much as Claude Opus 5.5 for a score 1.9 points lower.”

Newer versions usually cost less per task than the ones they replace, by more than the change in token price. Anthropic reports Sonnet 5 to Sonnet 5.5 at 87% less per solved task on the coding subset at an unchanged token price, and Opus 5 to Opus 5.5 at 78% less with about the same score. Its launch posts give smaller figures for typical work (up to 30% and about 40%). The exception in the guide is the newest model: Fable 5.2 at high effort cost more per solved task than Fable 5.1 on the coding subset.

The cheapest setting is often a policy, not a model

If you can check the output automatically, the cheapest option in Anthropic’s data was neither model at a fixed setting. It was running everything at low effort and re-running only the failures at high:

Policy (Opus 5.5, coding subset)Tasks solvedCost per task
Low effort, failures re-run at highAbout 97%About $0.17
Medium effort, failures re-run at highAbout 97%About $0.24
Everything at high95.3%$0.29

Two conditions apply, both from the guide. You need a real failure signal, since “a checker that passes bad work lets those failures through.” And every first-pass failure takes two runs’ worth of time.

Caps that look like savings and aren’t

Output caps. A low max_tokens cuts a response off without the model knowing, so the work is thrown away and still billed. On Anthropic’s internal repository tasks, Fable 5.2 capped at 16,384 tokens had 85% of attempts cut off, solved 20.3%, and cost $27.96 per solved task. At 64,000 it solved 77.4% for $24.42. The guide’s rule: set the cap high and “save money with effort, which the model can see.”

Answer length. In a measurement on an earlier model, the same bug-triage agent, asked for a one-line answer, a two-line answer or a six-section memo, cost $0.49, $0.57 and $1.40 per run, with accuracy “within run-to-run noise” across all three. “Ask for the answer you will read, not the one that looks thorough.”

How to measure it on your own work

Anthropic’s four steps, in plain words:

  1. Pull real tasks from your logs, weighted like your real traffic, and write a pass or fail check for each: tests pass, ticket closed, row count correct.
  2. Run each candidate model at two or three effort levels, not only the default. Record cost per task next to the pass rate.
  3. Add a second model (an advisor or cheaper workers) only if one model’s results can’t reach the quality you need. A pairing has to beat the single model’s whole range. See glossary/advisor-strategy.
  4. Run the winner on a slice of real traffic before switching, and keep the test set.

To price one task, add up every request it made. Each response reports uncached input, cache writes, cache reads and output separately, and each has its own rate. Thinking counts as output.

Our rule of thumb for a small business with no test harness: 20 to 50 real examples, graded blind by the person who normally does the job, is enough to see a difference of several times in cost. It won’t resolve a difference of a few percent.

What it doesn’t tell you

  • Benchmark tasks aren’t your tasks. Every number above is a coding, research or index benchmark. The ranking flipped between two of them.
  • Vendor data covers the vendor’s models. Anthropic’s guide measures Claude only. Artificial Analysis covers many vendors but reports cost per attempted task with a score beside it, not cost per passed task.
  • Quality isn’t one number. Two outputs can both pass a check and differ in how much editing they need.
  • People’s time isn’t in it. The cost of reviewing and fixing output never shows on the model’s bill.

FAQ

Is the cheapest model per token the cheapest to run?

Not reliably. In November 2026, three model settings with the same Artificial Analysis index score of 55 cost $1.27, $3.24 and $10.02 per index task, and two of them list at the same token price.

What is cost per completed task?

Everything you spent on a type of job, including failed attempts and retries, divided by the number of jobs that passed your check.

Does a lower effort setting always cost less?

No. Gemini 4 Carbon cost more per index task at low effort ($7.19) than at medium ($6.59) on Artificial Analysis’s index. Lower effort usually costs less per attempt, but a failed attempt still bills.

Should I use a cheaper model and retry when it fails?

If you can check the output automatically, that can be the cheapest policy: in Anthropic’s coding data, low effort with failures re-run at high matched everything at high for a little over half the cost. Without a reliable check, bad work gets through.

Is Claude Fable 5.2 cheaper than Opus 5.5 per task?

It depends on the work. On Artificial Analysis’s index, Fable 5.2 at high effort beat Opus 5.5 at max effort for slightly less. On Anthropic’s coding subset, Opus 5.5 reached the same pass rate for a little over half the cost.

Do not cite

ClaimWhy not
”Fable 5.2 is cheaper than Opus 5.5”True at matched quality on one index only; Anthropic’s coding data shows the reverse
”Lower effort always saves money”Gemini 4 Carbon cost more at low than at medium on Artificial Analysis’s index
”Half the token price means half the cost”Three settings at index 55 ranged from $1.27 to $10.02 per task
Google’s “more than 2X lower cost per task” for Gemini 4 CarbonThe announcement names neither the comparison model nor the method
Anthropic’s dollar figures as typical costsInternal benchmarks at list prices on the measurement date, labelled directional
Artificial Analysis’s cost per task as cost per completed taskIt is cost per attempted index task, reported next to a score

Key Takeaways

  • Compare models on what a finished, checked job costs. The token price is one input to that.
  • Effort level is the biggest dial: one model spanned $0.98 to $10.04 per index task.
  • A pricier model was cheaper per task in some measurements and dearer in others. The answer depends on the work, so test on yours.
  • Count failures. A failed attempt bills its tokens, then the retry.
  • With an automatic check, “cheap first, redo the failures at a higher setting” was the cheapest policy Anthropic measured.
  • Low output caps and long answer formats raise cost without raising quality.

Articles that use this page: Claude Fable vs Opus, and when Fable is worth the cost and How much does AI cost a business?.

Sources

By Andrej Ruckij