minimax-m3

All workflows

averaged across the boards this model contested

Per board

4 measurements · each against that board's own field

Provisional

2 cards · measured, never ranked

Runs measured outside a contested field. An ability card is an absolute measurement — passed/total per axis, EFF against each workflow's own par — so it stays meaningful with no opponent. A rank does not: with no field, a first place measures nothing. Nothing here carries one, and nothing here reaches the leaderboard.

Run #129 — flash tier at each model's effort ceiling 2026-08-26 · report

A real seven-model field, but across all five golden workflows at once, so it is published as cards rather than a ranked board: a leaderboard section covers one workflow, and folding five into a single row per contestant would publish a cross-workflow average under one workflow's heading. Two trials per cell. This is the first flash-tier field measured after four harness defects were fixed — the Anthropic-protocol 4096-token output cap, the unpinned ZenMux upstream lottery, the unmeasurable truncated turn and malformed tool call, and the dead trap step — so it is NOT comparable to runs #8-#126 for these models. Every model ran at its OWN measured effort ceiling (max, xhigh or high depending on the route), which makes this a capability measurement, not a shippable configuration; run #130 is the same field pinned to `low`.

Run #130 — the same field at low effort 2026-08-26 · report

The paired arm of run #129: identical models, workflows, trials, output budget and commit, with every contestant pinned to `low` so effort is the only variable and each model is its own control. The ranking reorders — Gemini 3.7 Flash goes from third to first and leads on both axes — and four of seven models score HIGHER here than at their ceiling, because lower effort cuts tool calls and EFF is the axis with the widest spread. On the 25 of 35 paired cells where both arms have enough trial consistency to be readable, `low` costs a mean 2.4 objective points, and under one point for three of the four models with a clean sample. Four arms were swept to invalid by a network outage and re-run with `--resume 130`; all 35 are scored at the same effort and budget.