---
title: "Fable 5.2 Costs Less Per Token, More Per Job"
url: https://nativefirst.ai/blog/fable-52-cheaper-per-token-more-per-job/
description: "Anthropic cut the token price 20%. Zapier measures $4.64 a finished task against Opus 5.5's $1.44. The lever that reconciles them is an effort setting."
author: John Tan
category: "AI Economics"
published: 2026-11-09
updated: 2026-11-09
publisher: nativefirst.ai
format: Markdown copy of the web page at the url above, same text.
---

# Fable 5.2 Got 20% Cheaper Per Token.

A finished job costs 3.2x what it costs on Opus 5.5.

John Tan · November 9, 2026 · 7 min read

Updated November 14, 2026

**Zapier reran the refused tasks, and the comparison below no longer holds.**

Opus 5.5 Max · 40.0% → 42.47% Fable 5.2 Max · $4.64, unchanged 3.6x → 3.2x per job

The live AutomationBench board now states, in a footnote, that for Claude Opus 5.5 and Claude Sonnet 5.5 "refused tasks were rerun with them; other task results are unchanged." Opus 5.5 at Max moved from 40.0% at $1.28 to **42.47% at $1.44**, and Sonnet 5.5 at Max from 44.6% at $1.00 to 44.75% at $1.14. Fable 5.2 did not move: 51.45% at $4.64 at Max, 39.73% at $1.30 at Low. So Figure 1 is now wrong in the reader's favour of Opus: Fable 5.2 at its laziest no longer matches Opus 5.5 at its hardest, it sits 2.7 points below it and 14 cents cheaper. The cost multiple moves with it: $4.64 against Opus 5.5 was 3.6x on launch day and is 3.2x now. The argument the post is making, that a cheaper token price is swamped by a model that does nearly twice the work per job, is unchanged. Figures read from [zapier.com/benchmarks](https://zapier.com/benchmarks) on 14 November 2026.

Anthropic cut the price of Claude Fable 5.2 by 20% on 9 November. $8 per million input tokens, down from $10. $40 output, down from $50.

Twenty-eight minutes later, Zapier published what one finished task actually costs on the new model. **$4.64.** The same work on Opus 5.5 costs $1.28.

Both numbers are correct. They are in different units, and that gap is the whole story of what this release does to a budget.

## Why a cheaper model costs more

Wade Foster, co-founder and chief executive of Zapier, ran Fable 5.2 through AutomationBench, Zapier's internal measure of whether a model can complete a real business workflow end to end, and published the tool-call counts alongside the scores. That is the part that explains everything else.

**Fable 5.2 averages about 97 tool calls per task. Fable 5.1 averaged 54.** It plans, it checks its own work, and it verifies before it acts. Thoroughness is billed by the token, so a 20% cut in the token price is easily swamped by a model that does nearly twice as much work per job.

This is the part that breaks forecasting. A per-token price is only a budget if the number of tokens is roughly fixed. Once the model decides for itself how hard to work, the price per token stops predicting the bill, and the only unit worth forecasting is the finished task.

## The setting almost nobody mentioned

Buried in the same post is the most useful finding of the launch, and it turns the cost story upside down.

**Figure 1. Zapier AutomationBench, live board · Fable 5.2's cheapest setting now trails Opus 5.5's hardest by 2.7 points.**

Claude Fable 5.2

Low effort

39.73%

AutomationBench

$1.30 per task

<

Claude Opus 5.5

Max effort

42.47%

AutomationBench

$1.44 per task

Opus 5.5 is ahead again, by 2.7 points for 14 cents more. Fable 5.2's cheapest setting still lands within reach of what maximum effort used to buy.

Zapier AutomationBench, live board read 14 November 2026. Launch-day figures in the update note. [zapier.com/benchmarks](https://zapier.com/benchmarks)

Read it as a floor rather than a headline. The interesting thing is not that Fable 5.2 can go higher. It is that the result you previously had to pay maximum effort for now arrives at the cheapest setting, which means your maximum-effort budget is free for work that genuinely needs it.

Anthropic found the same thing on a different benchmark, in a sentence almost nobody quoted.

> "In early testing on the FrontierCode 2.0 evaluation, **Fable 5.2 at High effort scored about level with Opus 5.5 at Max effort, for roughly half the cost per task.** At Xhigh effort, it also scored the highest of any released model tested."
>
> **Claude** · Anthropic · [Introducing Claude Fable 5.2 and Claude Mythos 5.2, 9 November 2026](https://www.anthropic.com/claude-fable-and-mythos-5-2)

Two organisations, two different benchmarks, one conclusion. On coding the parity shows up at high effort for half the money. On business workflows the cheapest setting now lands close to what maximum effort used to buy, for under a third of Fable 5.2's own top-effort price. Neither of them is selling you the ceiling. Both of them are telling you the floor moved.

## There is no such thing as the cost per task

Two minutes after Zapier posted, Perplexity published its own measurement, and the number is nothing like Zapier's.

> "On our WANDR benchmark, it achieved Pareto frontier of performance and efficiency, with a **record score of 0.851 at $14.51 per task.**"
>
> **Perplexity** · [19:29 UTC, November 9, 2026](https://x.com/perplexity_ai/status/2119875089974104489)

Same model, same day, and a cost per task eleven times higher than the one Zapier reported. Neither is wrong. Zapier is running short business workflows, Perplexity is running wide-and-deep research that reads the open web for as long as it takes, and Perplexity still calls $14.51 the efficient frontier because the alternative is worse.

So the useful figure is a range rather than a number, and the range is set by your harness and your job rather than by the model.

**Figure 2. Three measurements of the same model, 9 November 2026 · $1.30 to $14.51. All of them correct.**

Who measured Cost per task, to scale Per task

ZapierAutomationBench, low effort $1.30

ZapierAutomationBench, max effort $4.64

PerplexityWANDR, wide and deep research $14.51

Zapier figures from Wade Foster, 19:27 UTC. Perplexity figure from its own post at 19:29 UTC, where Fable 5.2 set a record WANDR score of 0.851. Different benchmarks measuring different work, which is the point.

Anyone quoting a single cost per task for Fable 5.2 has not said which job they mean. Before you budget, measure your own.

## What the ceiling actually buys

Maximum effort is not a waste. It is a purchase, and Foster published the price list.

**Figure 3. AutomationBench at max effort, live board · Best score on the board. Most expensive job on the board.**

Model Score Score Per task

Fable 5.2 51.45% $4.64

Sonnet 5.5Zapier's previous best 44.75% $1.14

Opus 5.5 42.47% $1.44

Fable 5.1the model it replaces 22.37% $2.45

Zapier AutomationBench, live board read 14 November 2026. Foster's 2.3x improvement over Fable 5.1 still holds: 51.45 divided by 22.37. [zapier.com/benchmarks](https://zapier.com/benchmarks)

Seven points over the previous best for about four times the price. Whether that is worth paying is not a judgement about the model, it is a judgement about the task, and the CEO of the automation company said so himself on launch day.

> "So figure out which workflows need Fable 5.2, and **run the rest on Opus 5.5 or Sonnet 5.5.**"
>
> **Wade Foster** · CEO of Zapier · [19:27 UTC, November 9, 2026](https://x.com/wadefoster/status/2119874501747814409)

## Where the ceiling is worth it

Foster published the per-department breakdown too, and it is the part to act on, because the gains are not spread evenly.

- **Finance: 16% to 53%.** Also the widest gap over Opus 5.5, which manages 36%.
- **Sales: 21% to 54%.**
- **HR: 9% to 37%.**

Those are three functions crossing from not-worth-attempting to worth-piloting in a single release. If you have been waiting for a reason to point an agent at reconciliation, quote-to-cash or onboarding paperwork, this is the first number that supports it, and it is also the place where paying $4.64 a task is easiest to justify.

## Three things to do with this

**Set the default to low and let the work argue for more.** On business workflows, low effort buys you last week's best model at last week's price. Reserve high and max for the tasks where Figure 3's seven points change an outcome rather than a score.

**Forecast tasks, not tokens.** Take whatever you spent last month, divide by jobs completed, and track that number instead. A per-token price that falls while the cost per job rises is not a contradiction, it is what happens when the model controls its own token count.

**Set the budget before you walk away.** Mike Taylor, head of evals at Every, left Fable 5.2 running over a weekend on a product they are building, and it made 75 changes unattended. His rule is the one to copy, because the third item is the one teams forget.

> "I can't be in the loop because it's too slow. But I don't need to be in the loop so much because it's so deliberative." Before walking away, give it three things: **what counts as done, how much testing is enough, and the budget.**
>
> **Mike Taylor** · Head of Evals at Every · [23:01 UTC, November 9, 2026](https://x.com/every/status/2119928286465892357)

Deliberation is the thing you are buying and the thing you are billed for. They are the same property, which is why a budget is the only instruction that reliably stops it.

**Check what your subscription already includes.** Boris Cherny, who works on Claude at Anthropic, said on launch night that the longer you have been subscribed, the more Fable resets you get. Capacity is now priced partly by tenure, so older accounts may be sitting on headroom nobody has looked at.

The token price is the number Anthropic publishes. The cost per finished job is the number you pay. This release moved them in opposite directions, and the lever that reconciles them is a setting most teams have never changed.

**Turn the effort down.**

[John Tan ↗](https://www.linkedin.com/in/johntans/)

Founder of nativefirst.ai. Runs everything on agents and writes down what works.

Sources

1 Anthropic, Introducing Claude Fable 5.2 and Claude Mythos 5.2, 9 November 2026, read at source. Pricing: $8 per million input tokens and $40 per million output, down from $10 and $50. [anthropic.com](https://www.anthropic.com/news/claude-fable-and-mythos-5-2)

2 Same post, on effort and efficiency: in early testing on FrontierCode 2.0, Fable 5.2 at High effort scored about level with Opus 5.5 at Max effort, for roughly half the cost per task. At Xhigh effort it scored the highest of any released model tested. [anthropic.com](https://www.anthropic.com/news/claude-fable-and-mythos-5-2)

3 Wade Foster, CEO of Zapier, 19:27 UTC 9 November 2026. AutomationBench at max effort: Fable 5.2 51.4%, Sonnet 5.5 44.6%, Opus 5.5 40.0%, Fable 5.1 22.4%. Cost per task $4.64, $1.00 and $1.28. About 97 tool calls per task against 54 for Fable 5.1. At low effort Fable 5.2 scores 39.7% for about $1.30 against Opus 5.5 at max for $1.28. Figures confirmed on the live board 13 November 2026: Claude Fable 5.2 (Max) 51.45% at $4.64 per task. [zapier.com/benchmarks](https://zapier.com/benchmarks)

4 Same post, per-department gains over Fable 5.1: finance 16% to 53%, sales 21% to 54%, HR 9% to 37%. Finance is also the widest gap over Opus 5.5, 36% to 53%. [anthropic.com](https://www.anthropic.com/news/claude-fable-and-mythos-5-2)

5 Boris Cherny, 19:21 UTC 9 November 2026, on Fable resets scaling with subscription length. [x.com](https://x.com/bcherny/status/2119873095129550880)

6 Perplexity, 19:29 UTC 9 November 2026: Fable 5.2 on the WANDR benchmark scored a record 0.851 at $14.51 per task, which Perplexity calls the Pareto frontier of performance and efficiency. [x.com](https://x.com/perplexity_ai/status/2119875089974104489)

7 Every, 23:01 UTC 9 November 2026, on Mike Taylor leaving Fable 5.2 running over a weekend: it made 75 changes to Checks. Quoted: I can't be in the loop because it's too slow. But I don't need to be in the loop so much because it's so deliberative. Before walking away, give it what counts as done, how much testing is enough, and the budget. [x.com](https://x.com/every/status/2119928286465892357)
