fedi-index: Agents

Models as agents: tool use and multi-step tasks — Arena agent sessions and τ²-bench runs.

Live data for 2026-10 (recalculated daily) · Methodology v1 · Methodology · History · JSON

All Open weights Closed

Tier

Model

fedi-index

Percentile

People prefer

Solves tasks

Value

Sources

1,456 ±5

1,459

1,449

+44

Aτ

›

1,454 ±5

1,456

1,449

+45

Aτ

›

1,449 ±2

1,459

—

+92◆

›

1,448 ±3

1,458

—

›

1,442 ±3

1,451

—

+60

›

1,435 ±7

1,445

—

+33

›

1,430 ±2

1,439

—

+105◆

›

1,428 ±3

1,437

—

›

1,428 ±6

1,438

—

+98◆

›

1,424 ±3

1,433

—

+46

›

1,408 ±5

1,416

—

+68

›

1,402 ±4

1,410

—

+50

›

1,399 ±4

1,407

—

+80

›

1,388 ±8

—

1,411

+23

›

1,388 ±8

—

1,411

+27

›

1,386 ±6

1,394

—

+38

›

1,384 ±4

1,392

—

+27

›

1,377 ±6

1,384

—

-19

›

AArena (human votes)ττ²-bench (agents) ◆ Pareto frontier: nothing is both better and cheaper

fedi-index is an Elo-equivalent on a fixed scale: comparable between months within one methodology version. Tiers S–D: by percentile in the slice, and no better than the confidence interval allows (cut-offs in the methodology). “Provisional” — only one source family measured the model here. Methodology

Embed on your site

The table as a widget for your site: it updates by itself and keeps the attribution of the sources.

Code

<script src="https://fedi.software/embed.js" data-fig="fedi-index-agents" data-lang="en" async></script>