Low effort clears Fable 5.1's best
At low effort, Fable 5.2 passes 261 of 657 tasks (39.7%), against 22.4% for Claude Fable 5.1 at max. High effort, the API default, reaches 44.3%, about double Fable 5.1's best.
Release report · November 2026Claude Fable 5.2 · With default fallbacks
More than double, in one release.
Claude Fable 5.2 completes 51.4% of AutomationBench's strict, end-to-end business tasks at max effort, more than double Claude Fable 5.1's 22.4%. It's the highest score from any Claude model; only Gemini 4 Carbon (53.9%) scores higher.
Best pass rate51.4%Max effort · 338 of 657 tasks
vs. Claude Fable 5.1+29.0pts22.4% at Max
Cost per task$4.64Max effort · launch-price estimate
On the frontier2/ 5effort settings no model in this comparison beats on both cost and accuracy
Key findings
Ahead of every Claude model
At max, Fable 5.2 is 6.8 points above Claude Sonnet 5.5 (44.6%) and 11.4 above Claude Opus 5.5 without fallbacks (40.0%). Gemini 4 Carbon is the only model higher, at 53.9%.
It does more on every task
At max, Fable 5.2 makes 96.9 tool calls per task, against about 54 for Fable 5.1. Even at low effort it makes 58.4, and no task at any effort hit the 50-step cap.
Max doubles the cost of xhigh
Going from xhigh to max adds 3.5 points (47.9% to 51.4%) and takes cost per task from $2.29 to $4.64. Counting failures, a pass costs $4.77 at xhigh and $9.01 at max.
01
Results by effort
Low and medium are a near tie. Each step after adds 3.5 to 4.1 points.
Scoring is strict. A task passes only if every assertion about the final state of the simulated company holds. All 657 tasks stay in the denominator at every effort, refusals included. The dashed line marks Claude Fable 5.1 at max effort.
Official pass rate · task_completed_correctly · 657 tasks per effort, all in the denominator
02
Against the field
Level with Gemini 4 Argon, behind only Gemini 4 Carbon.
Each bar is a model's best effort setting on AutomationBench 1.0.6. Fable 5.2 at max and Gemini 4 Argon at high (51.3%) are within run-to-run noise of each other. Claude Opus 5.5 and Sonnet 5.5 are the runs in this report's export; the other results are from zapier.com/benchmarks.
Gemini 4 CarbonHigh
Claude Fable 5.2Max
Gemini 4 ArgonHigh
Claude Sonnet 5.5Max
GPT-6 AstraMax
Claude Opus 5.5Max
Claude Fable 5.1Max
Best observed pass rate per model · AutomationBench 1.0.6
03
Cost vs. accuracy
Past 44.6%, Fable 5.2 is the only option.
Each line is one model's effort settings, pass rate against cost per task. Fable 5.2 holds the frontier at xhigh and max. Below that, Claude Sonnet 5.5 at max (44.6% for $1.00) costs less per task. Fable 5.2's estimates include cache writes at an assumed rate; the Opus 5.5 and Sonnet 5.5 estimates leave them out and run low.
Claude Fable 5.2 Frontier Frontier point DominatedHover a point for details
View the data
| Model | Low | Medium | High | XHigh | Max |
|---|---|---|---|---|---|
| Claude Fable 5.2 | 39.7%$1.30 | 40.2%$1.35 | 44.3%$1.69 | 47.9%$2.29 | 51.4%$4.64 |
| Claude Opus 5.5 | 23.3%$0.46 | 28.6%$0.60 | 32.0%$0.65 | 34.4%$0.80 | 40.0%$1.28 |
| Claude Sonnet 5.5 | — | — | — | — | 44.6%$1.00 |
API cost per task, launch-price estimate · pass rate at each effort, low to max · Claude Opus 5.5 (without fallbacks) and Claude Sonnet 5.5 are the comparison runs in this export, and their cost estimates leave out cache writes. Gemini 4 Carbon, Gemini 4 Argon, GPT-6 Astra and Claude Fable 5.1 are pass-rate results from the zapier.com/benchmarks leaderboard. Claude Fable 5.1 isn't plotted on cost because its cost accounting is incomplete.
04
By domain
HR more than doubles, 17.5% at low to 36.7% at max.
Each line runs from low to max. HR gains 19.2 points, the most of any domain, and is still the hardest. Operations is the strongest at 67.0%, and Marketing peaks at xhigh (62.0%), not max (57.0%).
0%10%20%30%40%50%60%70%80%
Low → MaxOperations
60.0%67.0%
+7.0ptsMarketing
51.0%57.0%
+6.0ptsSales
39.3%53.8%
+14.5ptsFinance
45.0%53.3%
+8.3ptsSupport
29.0%43.0%
+14.0ptsHR
17.5%36.7%
+19.2pts Low Medium, High, XHigh Max
Full matrix
| Domain | Low | Medium | High | XHigh | Max |
|---|---|---|---|---|---|
| Sales | 39.3%46/117 | 41.0%48/117 | 47.0%55/117 | 47.0%55/117 | 53.8%63/117 |
| Marketing | 51.0%51/100 | 50.0%50/100 | 51.0%51/100 | 62.0%62/100 | 57.0%57/100 |
| Operations | 60.0%60/100 | 60.0%60/100 | 64.0%64/100 | 60.0%60/100 | 67.0%67/100 |
| Support | 29.0%29/100 | 32.0%32/100 | 36.0%36/100 | 41.0%41/100 | 43.0%43/100 |
| Finance | 45.0%54/120 | 41.7%50/120 | 45.0%54/120 | 50.0%60/120 | 53.3%64/120 |
| HR | 17.5%21/120 | 20.0%24/120 | 25.8%31/120 | 30.8%37/120 | 36.7%44/120 |
Pass rate · passed / tasks per cell · darker = higher
05
Cost
$1.69 a task at high, $4.64 at max.
Spend per pass divides every task's cost by the number of passes, failures included: $3.28 at low, $9.01 at max. Estimates use Anthropic's launch rates of $8.00 input and $40.00 output per 1M tokens, with cache rates assumed until Anthropic publishes them. These are not billed costs.
| Effort | Cost / task | Passed task | Failed task | Spend per pass |
|---|---|---|---|---|
| Low | $1.30 | $1.48 | $1.19 | $3.28 |
| Medium | $1.35 | $1.54 | $1.22 | $3.36 |
| High | $1.69 | $1.86 | $1.55 | $3.81 |
| XHigh | $2.29 | $2.42 | $2.16 | $4.77 |
| Max | $4.64 | $4.65 | $4.63 | $9.01 |
USD, launch-price estimate · spend per pass = all task costs ÷ passes, failures included
06
Tokens & tools
Max makes 96.9 tool calls a task.
From low to max, tool calls per task rise 66% and steps from 10.4 to 15.7. Reasoning tokens grow 22.7× (2,055 to 46,657 per task). No task hit the 50-step cap, and 84.0 to 87.4% of input is served from cache.
Reasoning Visible output (text + tool calls)Mean output tokens per task
| Effort | Tool calls | Steps | Calls / step | Input tokens | Cached share | At step cap |
|---|---|---|---|---|---|---|
| Low | 58.4 | 10.4 | 5.6 | 352.4k | 84.0% | 0 |
| Medium | 59.4 | 10.6 | 5.6 | 375.7k | 84.5% | 0 |
| High | 65.6 | 12.0 | 5.5 | 514.5k | 86.5% | 0 |
| XHigh | 73.0 | 13.3 | 5.5 | 690.4k | 87.4% | 0 |
| Max | 96.9 | 15.7 | 6.2 | 1267.5k | 87.2% | 0 |
Mean per task, all tasks included · step cap 50
07
Refusals
Refusals stay at or under 1.1% at every effort.
A refused task counts as a failure. Across all five efforts, Fable 5.2 refused 18 tasks, every one labeled biology-related by Anthropic's classifier, and none of them passed. The most were in Operations at xhigh, 6 of 100 tasks.
0.5%
Low3/6570.8%
Medium5/6570.3%
High2/6571.1%
XHigh7/6570.2%
Max1/657| Domain | Low | Medium | High | XHigh | Max |
|---|---|---|---|---|---|
| Sales | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% |
| Marketing | 0.0% | 2.0% | 1.0% | 0.0% | 0.0% |
| Operations | 1.0% | 2.0% | 1.0% | 6.0% | 0.0% |
| Support | 0.0% | 1.0% | 0.0% | 1.0% | 1.0% |
| Finance | 0.8% | 0.0% | 0.0% | 0.0% | 0.0% |
| HR | 0.8% | 0.0% | 0.0% | 0.0% | 0.0% |
Recorded refusal flags · refused tasks stay in the denominator
08
Methodology
AutomationBench tests AI agents on end-to-end workflow execution using 47 simulated apps across six business functions: Sales, Marketing, Operations, Support, Finance and HR. Each task boots a small simulated company, hands the agent one request a real operations person would get, then grades the world it leaves behind with deterministic assertions.
- 01
Boot a tiny company
CRM records, inbox threads, spreadsheets and support cases, including traps like stale rows and near-duplicate names.
- 02
The agent works alone
The agent gets one trigger message and tools to search, read, update and send. It can't ask clarifying questions, and there's no human in the loop.
- 03
Grade the end state
Deterministic assertions check what changed. The task passes only if every one holds.
CostEstimates at Anthropic's announced launch rates for Claude Fable 5.2: $8.00 input and $40.00 output per 1M tokens, with reasoning billed as output. Cache reads ($0.20 per 1M) and cache writes ($10.00 per 1M) are assumptions until Anthropic publishes them. These are not billed costs.
ComparisonsClaude Opus 5.5 (without fallbacks) and Claude Sonnet 5.5 are the comparison runs in this export, and their cost estimates leave out cache writes. Gemini 4 Carbon, Gemini 4 Argon, GPT-6 Astra and Claude Fable 5.1 are pass-rate results from the zapier.com/benchmarks leaderboard. Claude Fable 5.1 isn't plotted on cost because its cost accounting is incomplete.