Model Bench

Raw scores live on different scales, so each benchmark gets two anchors: the raw score that maps to 0 and the one that maps to 100 (table above). Scores outside the anchors are clipped.

Example: 56 on Terminal-Bench becomes 78.5, because its anchors are 5 and 70.

Few configurations were measured on every benchmark, at every effort level, in every harness. For each benchmark the engine takes the measurement closest to what you asked for, then adjusts it. Each adjustment multiplies that value's reliability, so it carries less weight in the final score. The table marks every adjusted value in italics with a letter, and the record sheet lists the exact adjustments.

Harness first

It looks for a run in the harness you picked: the vendor's own harness in Native mode, the minimal common harness in Common mode. None of these models has an OpenCode run, so in Company mode non-Anthropic models start from common-harness runs.

Effort scaling ^e

Among the candidate runs it picks the effort level whose AA index is closest. If that is not the effort you are looking at, the score moves 1.5 benchmark points for each point of AA index between the two. Reliability × 0.8.

Previous-generation proxies ^p

Where a model has not been run yet, its predecessor's score stands in, with a reliability of 0.4 to 0.5 set per measurement. Every proxy is labelled in the sources table.

Harness conversion ^h

If the only run is in the other harness, the vendor's typical gap is added or subtracted (the conversion table). Only Terminal-Bench, DeepSWE and SWE-rebench are converted; the AA index and FrontierCode do not depend on the harness. Reliability × 0.6.

OpenCode offset ^o

In Company mode, OpenAI, xAI and Alibaba models run in OpenCode. Their harness-sensitive scores get the offset you set (default −3). This is the least certain number in the app: the only direct OpenCode measurement is of a model outside this list (GLM-5.3, about 2 points below the common harness on Terminal-Bench). Reliability × 0.7.

A benchmark with no measurement drops out and the others share its weight. Confidence is the denominator, Σ (weight × reliability). It is 1.00 only when all five benchmarks were measured directly for that model, effort and harness. Lower means more of the score rests on filled-in values. Qwen3.8 Flash-Next has only the AA index, so its confidence is 0.51: its capability is the AA index alone.

α says how many capability points a doubling of cost is worth. At 0, cost is ignored. At the default 6, a configuration that costs twice as much has to score 6 points higher to break even. At 16, cost dominates and cheap models win. Because cost is on a log axis in the chart, every line of equal value is straight; the dashed red line is the one through the current best, so nothing sits above it.

A configuration is on the Pareto front when no other one is at least as cheap and at least as capable, and strictly better on one of the two. The Sweet spot view shows front configurations within 15 value points of the best, the best-value configuration in each cost band (under $0.05, $0.05 to $0.25, $0.25 to $1.00, $1.00 and up), and anything free to you. Faint points in the chart are the ones the view hides.

Cost per task, output speed and time to first token are stored per configuration; values marked est. are estimates. With the Grok discount on, Grok's cost is multiplied by 0.1. Cost does not depend on the harness here, though in practice it does (see the open gaps).