Best AI model for automation: cost per finished task

AutomationBench gives an AI agent one request, the kind an ops hire gets on Slack, inside a simulated company with a CRM, an inbox, spreadsheets and support cases across 47 apps. It then grades the data the agent leaves behind, not what the agent says it did. A task passes only if every check holds.

That is why it is the most useful benchmark for anyone choosing a model for an automation. Most benchmarks test whether a model can answer a question or write code. This one tests the boring, expensive part of ops work: picking the right "Meridian Corp" out of three near-identical CRM records, finding the routing rule buried in an old email, using the newest row in a spreadsheet instead of the stale one, and emailing the right team and nobody else. Zapier built the tasks from patterns across more than 2 billion monthly Zapier tasks, keeps the scored set private so models cannot be trained on it, and spans six functions: sales, marketing, operations, support, finance and HR.

Sources: zapier.com/benchmarks (leaderboard 1.0.6, full 134-row board read 11 November 2026) and Zapier's 28 October 2026 post on AutomationBench joining Fireworks' Specialized Intelligence Index. All scores and per-task costs below are Zapier's. The per-finished-task figures are our arithmetic on them; IV did not run the benchmark.

Cost per task is what you pay per attempt. What you care about is the cost of a task that actually finished. Divide the first by the success rate and Claude Sonnet 5.5 comes out at about $2.55 per finished task, Gemini 4 Carbon at $3.11 to $3.93, and Claude Fable 5.2 at $9.02. Same data, very different ranking.

How we calculated it: Zapier's published cost per task divided by its published score, both read 11 November 2026. "Per 1,000 runs" multiplies those by 1,000. These are model API costs inside Zapier's test harness on hard multi-app tasks; your own runs will cost less on simpler workflows and more on longer ones. They exclude any Zapier task charges, which we cover next.

The table hides the number that should worry you most: the failures. At 1,000 runs, even Gemini 4 Carbon leaves around 460 tasks unfinished, and Zapier's data says most of those will look finished. The cheap rows are cheap per finished task, but GPT-6 Sol leaves two in three tasks wrong. A cheap model with a high failure rate only works if something downstream catches the failures, and a person fixing a wrong CRM record costs far more than $2.

Claude Sonnet 5.5 at max effort for multi-app agent steps in Zapier. Gemini 4 Argon at medium effort for the same job in n8n while Google's promotional price holds. A cheap model for single steps. And no frontier model without a check step behind it. We take that side because the top of the board is separated by a few points, while cost per finished task varies by 3x to 6x, and every model fails often enough that the check step matters more than the model.

For a wider view of which model to use for writing, research and analysis rather than automation, see our which AI model for which task guide. If you want the model choice, the check steps and the cost mapped for your own workflows, that is what our automation practice does.