GDPval benchmarks

Sourcing and comparability

Source types: 🏢 vendor-reported scores come directly from the model developer and may use proprietary evaluation setups. 📊 leaderboard scores are from public, independently-maintained evaluation runs. 🔬 independent scores are from credible third-party evaluations. 📄 paper scores are from peer-reviewed or arXiv publications. Every score on this page is 🔬 independent — see below.

About the GDPval-AA v2.1 benchmark

GDPval-AA v2.1 builds on GDPval (OpenAI, 2025), a set of ~220 real-world knowledge-work tasks developed with industry professionals across finance, healthcare, legal, and other professional domains. Models work through each task in an agentic loop (Artificial Analysis's Stirrup harness, with shell and web access) and hand in actual deliverables — documents, spreadsheets, slides, diagrams. Those deliverables are compared blind, two at a time, and the results become an Elo rating, not a pass-rate.

How to read a v2.1 score

An Elo number only means something next to other numbers on the same scale. Artificial Analysis fixes the v2.1 scale by pinning one model, DeepSeek V4.1 Flash (max), at 1,600, and fits the ratings with a Crowd-BT model. The scale is frozen, so a model's rating shouldn't drift as new models join the pool. That's a change from v2, where published figures drifted as the pool grew. Read the gaps between models, not the absolute figure: a higher rating means judges preferred that model's deliverables to the other models' more consistently. There is no human-expert reference point on v2.1, so a score doesn't say whether a model beats a professional (see Artificial Analysis's methodology ↗).

Why the version changed: until mid-September 2026 this page used GDPval-AA v2, anchored to a human-expert baseline of 1,000. Artificial Analysis then replaced v2 with v2.1 at the same address. Its methodology says v2.1 "changes only how the Elo scale is fixed" (same tasks, harness and judging) and that rank ordering is "largely preserved", but every number moved. We switched on 2026-09-23 and re-checked every model against the live v2.1 leaderboard. The old v2 figures are kept in the source registry for history but aren't shown, because they sit on a different scale.

Why this is sourced differently from Coding/Science/Agentic: Every other vertical on this site anchors on scores labs self-report in their own system cards. GDPval-AA is run independently by Artificial Analysis instead. Some labs now quote its figures in their own launch materials (Anthropic's Claude Opus 5.5 system card, 2026-09-22, cites GDPval-AA v2.1), but every score here comes from Artificial Analysis directly. Every score here is cited as an independent evaluation, not a vendor claim, and each model is cited at its best-confirmed reasoning-effort configuration (disclosed under the score date) — scores at different configurations are not directly comparable.

What a GDPval task actually looks like

The tasks aren't quiz questions — they're the kind of assignment a professional gets handed on a Monday morning. GDPval's authors, industry practitioners averaging 14 years of experience, needed an average of about 7 hours per task to complete them themselves (median 5 hours; the longest run to multiple weeks), and each one ends in a real file: a spreadsheet, a slide deck, a legal memo, a diagram. Three examples from the openly published 220-task subset:

What the score means in practice: a higher GDPval-AA v2.1 rating means blind judges preferred that model's finished file over other models' on the same task more consistently. That's a statement about relative first-draft quality, not a promise of hands-off automation. Preference isn't accuracy either: judges picked the deliverable they preferred, not necessarily the one a specialist would find error-free. GDPval's own authors note that once you add the time a professional spends reviewing and fixing the output, the real-world speed-up is far smaller than the raw generation speed (roughly 1.0–1.2×, per arXiv 2510.04374). Reasoning-effort configuration matters too: the same model scores materially differently at medium, high or max effort, so each row is cited at its best-confirmed configuration.

Data is growing: This table is sourced from modelglass-gdpval, a registry of GDPval-AA knowledge-work benchmark results. Scores are append-only — once recorded, a result is never edited or deleted.