Model B · low effort

Navigator Index78.995% CI 75.0–82.6

Rank7–9Ranks 7–9: we can't statistically separate it from the models in that range. Of 14.▼1−0.5

Cost / successful task$0.0158th cheapest of 14 priced models

Median latency5.1 sPass rate 76% · every time 64%

01Categories

Where it is strong, and where it is not

Category · weightModel B

Coding 15%73.357.5–83.6 · 5 tasks

Data analysis 15%85.881.2–89.6 · 5 tasks

Extraction 10%77.557.0–92.4 · 5 tasks

Reasoning 15%65.859.3–72.4 · 5 tasks

Research & agentic 15%For now these are single-call multi-step tasks: one model call returns plan → steps → deliverable, checked step by step. Tool-using, browsing and long agentic tasks are deferred.83.378.7–87.1 · 5 tasks

Summarization 10%77.562.6–90.5 · 5 tasks

Support 10%86.774.6–94.1 · 5 tasks

Writing 10%85.075.7–95.2 · 5 tasks

Run receipts120 demo runs · no receipts

02Cost

What a working answer costs

$0.015per successful task

$15.48per 1,000 successes

$1,548per 100,000 successes

76%runs that pass

120runs measured

1 of 13models it dominates

Cost per successful task divides everything measured on candidate calls by the tasks that passed, so failed runs are priced in. Judge cost is excluded. Definition

Dominates

Significantly better on the Index (paired test) and cheaper per successful task:

Model J

Dominated by

Each of these is significantly better on the Index and cheaper:

Model M · low effort

03Trend

Across published cycles

Navigator Index — higher is better · shaded: 95% interval

Published cycle

Navigator Bench · Index v2.0 · data through 14 Oct 2026Demo data

04Receipts

Every run, on the record

Demo data · not real results This is demo data: these runs do not exist, so there are no receipts. On a published cycle this lists every run with its prompt, output, checks and measured cost.

Rank card 16:9 Rank card 1:1 Compare with others Kill zone

Navigator Bench · Index v2.0 · data through 14 Oct 2026Methodology