Best Models for Automation Agents
Use cheaper classification/extraction models for frequent branches and stronger reasoning models only when the workflow reaches hard decisions.
Last verified: 2026-06-05
Current recommendation
Use cheaper classification/extraction models for frequent branches and stronger reasoning models only when the workflow reaches hard decisions.
- Use these as starting candidates, not universal winners.
- Run a short test in your own workflow before moving production volume.
- Use Usage Logs to compare request IDs, latency, usage, and billed amount.
Methodology
MetisRouter should publish measured benchmark values only when the same task set has been run across the compared models.
- Track task success rate, first-token latency, total latency, retry rate, and cost per successful task.
- Measure tool-calling reliability and JSON/schema adherence for agent workflows.
- For image and video, measure output quality, supported parameters, duration, resolution, and cost unit.
Measurement status
This page is a recommendation and methodology page until a reproducible benchmark run is published. It intentionally avoids fake scores, fake ranks, and Dataset schema because measured task results, prompts, timestamps, and Usage Logs are not yet published as a public benchmark dataset.
- Treat recommendations as candidate shortlists for your own workflow tests.
- Use exact model IDs, endpoint type, input size, output limits, and request timestamps when comparing models.
- A future measured release should publish task set, scoring rubric, p50/p95 latency, success rate, cost per successful task, and known failure modes.
Evaluation log
A useful benchmark needs enough operational detail to be reproducible and commercially meaningful.
- Model ID, endpoint, prompt or task class, input size, output limit, and timestamp.
- p50/p95 latency, error rate, billed usage, and retry count.
- Known failure modes and when the model should not be used.
FAQ
Why are there no fake leaderboard scores?
Publishing unmeasured scores would hurt trust. These pages publish recommendations and methodology until real benchmark runs exist.
How do I compare two models today?
Run the same bounded task with both exact model IDs, then compare Usage Logs for latency, tokens/assets, cost, and output quality.