Breaking

Agent Arena Data Shows Hour-Long AI Agent Tasks Averaging 332 Tool Calls, With Model Choice Cutting Costs by Up to 66%

November 4, 2026 at 12:53 EST

AI Agents
Foundation Models

Nearly 30,000 AI agent tasks that ran for more than an hour were recorded on Agent Arena over the past 14 days, according to the platform's latest data, which also shows that choosing the right model can substantially reduce the cost of long-running work.

Past 14 days · AI agents

Long-running AI agents: choosing the right model can cut costs by up to 66%

Nearly 30,000 agent tasks running longer than an hour were recorded on one platform. In these long sessions, model selection emerges as a major lever on total spending.

~30,000

Agent tasks over one hour, in 14 days

332

Average tool calls in a long session

up to 66%

Cost reduction from the right model choice

$10.51

Median cost per hour of agent work

Relative cost of one long-running workload

Same task, same unit (relative cost). Column height is drawn to scale.

100%

Single default model

34%

Right model for long tasks

Takeaway: Up to 66% less spend on the same long-horizon work, when the model is matched to the task.

Why small per-step differences add up

Long run
Hour-plus agent session

→

~332 tool calls
Searches, code, API requests

→

Small cost per step
Compounds across every call

→

Model choice
Large gap in total bill

Read with caution

A snapshot from one platform, not a universal benchmark

Data covers a single platform over a 14-day window.
How the $10.51 median was calculated is not described.
The models compared are not named.
Task success rates alongside cost are not reported.
Savings may vary across other workloads and providers.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.