Defensible Benchmarks · verify before you sign up

Agents already work. Here's the receipts.

Three numbers from this year, each with a primary source you can open in a new tab. This is the public version of the credibility tile we show in the workshop, built so a sceptical operator can fact-check it before paying for a seat.

Source posture

Primary links first, methodology linked separately, and the as-of date travels with the tile so screenshots stay accountable.

Deep-research agents: GPT-5.6 Sol 92.2% on BrowseComp (2026-07) — a related but distinct job.

What the web-agent number does and doesn't cover. Read-heavy tasks (lookups, extraction) are largely solved; write-heavy tasks — logins, forms, file uploads — remain the hard part (~47% SOTA, Halluminate Web Bench). Online-Mind2Web leaderboard entries are submitted by the system authors and screened against the maintainers' published WebJudge auto-eval; Princeton HAL, which re-runs agents itself, reports 40–42% on the same suite for 2025-era research scaffolds.

Sources, in order: xlang.ai/blog/osworld-verified (OSWorld-Verified — maintainer verification process, XLANG Lab) · github.com/xlang-ai/OSWorld (OSWorld benchmark + NeurIPS 2024 paper — methodology of record; 85% carried identically by llm-stats, BenchLM and datalearner) · arxiv.org/abs/2504.01382 (“An Illusion of Progress?”, OSU NLP + UC Berkeley, COLM 2025 — Online-Mind2Web methodology of record, incl. WebJudge) · github.com/OSU-NLP-Group/Online-Mind2Web (Online-Mind2Web leaderboard + submission requirements) · webbench.ai (Halluminate Web Bench — read-heavy / write-heavy split) · leaderboard.steel.dev/registry/benchmarks/browsecomp (BrowseComp — deep-research callout) · averi.ai B2B SaaS citation benchmarks report (2026) (B2B SaaS citation benchmark). Benchmarks current as of 2026-08-31 — refreshed monthly under the Benchmark of the Month routine.

How to read this. Every number above links to its primary source — open them. The three logos are shown in greyscale on purpose: the numbers carry the slide, not the brands. We refresh these figures monthly; the "current as of" date moves even when the numbers don't, so the tile never goes stale-looking. Questions about methodology? The OSWorld NeurIPS 2024 paper is the method of record for the desktop number, and “An Illusion of Progress?” (COLM 2025) for the web-agent number — including the WebJudge auto-eval the leaderboard screens against.