Which AI agents do best on independently checked benchmarks. Each agent is listed once, with its best run.
AI-generated illustration
ANALYSIS
AI agents are getting better through smarter setups, not just smarter models
New work from DeepSeek and Tsinghua shows the software around an AI model can cut costs and raise scores as much as a new model.
Read the analysis
In Beijing, DeepSeek’s Harness team published on 31 October a design for an always-on “Super-agent” and an MIT-licensed prototype: a resident Router takes requests from GitHub and chat and dispatches them to isolated Worker sessions, while a separate Watchdog checks progress. The README calls it research code, not a product. On synthetic workloads the team reports half of tasks finished within 12.4 minutes, and lists prompt injection as an open risk. [1]
A Tsinghua–Zhipu-led paper, HarnessPro, lets agents analyse their own runs and rewrite their harness. With GPT-5.5 and Codex, the authors report Terminal-Bench 2 rising from 76.3% to 80.8% while cost per task fell from $1.09 to $0.72. These are the authors’ results; search costs are excluded. [2]
On the vendor side, OpenAI says GPT-6.1 Terra reaches 50.3% on Terminal-Bench 4.0 at max effort, at a fraction of rivals’ per-task cost. That is self-reported and not yet on the maintainers’ board. [3]
Verified boards did not move this week: Terminal-Bench 4.0 is still led by Claude Code with Opus 5.5 (64.8% ±3.1), and the maintainer-checked SWE-bench Verified and HAL entries are unchanged. [4][5][6]
Our interpretation: cost per solved task is becoming the competitive metric, and harness design moves it as much as model choice. Leaderboards that report only accuracy will increasingly miss that.
Uncertain: Paper and vendor figures are self-reported; only maintainer-run boards count toward our leaderboard. Retrieved 1 November 2026.
How we rankSources, verification and why there is no composite
We use only official leaderboards whose results are checked by the maintainers or run independently. Vendor-reported scores are excluded. Each agent appears once with its best run, so # is our rank after removing duplicates; the official rank is shown in small text.
No composite rank today: only 4 models appear on three or more boards, and the boards cover different model generations.
HAL — Holistic Agent Leaderboard (Princeton) Independent third-party runs on standardised harnesses, with cost; covers GAIA, τ-bench, Online Mind2Web, SWE-bench Verified Mini and more.
GAIA leaderboard (Hugging Face) Official GAIA leaderboard on a held-out test set; used for reference, HAL runs used for scoring.
τ²-bench (Sierra Research) Maintained tool-use and customer-service agent benchmark; scores currently taken from HAL’s independent τ-bench runs.
OSWorld / OSWorld-Verified Computer-use benchmark on real operating systems; leaderboard site os-world.github.io was unreachable on 1 Nov 2026 — not yet automated.
METR — time horizons Independent measure of the length of tasks agents can complete; reported by METR, not yet automated here.
BrowseComp (OpenAI) Web-browsing benchmark; most published scores are vendor-reported, so excluded from scoring unless independently reproduced.
METR blog Independent evaluations of frontier-model agent capabilities.
Anthropic Engineering First-party notes on agent design and evaluations (self-reported).
LangChain blog (RSS) Agent framework vendor; practitioner patterns (vendor perspective).
Latent Space (RSS) AI-engineering newsletter and podcast with frequent agent coverage.
xAI / Grok on independent boards
Grok appears on Terminal-Bench 4.0 (Grok Build + Grok 4.7: 37.6%, official rank 21; Grok 4.6: 20.3%; Grok 4.5: 12.4%). There are no Grok entries on SWE-bench Verified, OSWorld-Verified or Princeton HAL (GAIA, τ-bench airline, Online Mind2Web). xAI's own launch benchmarks are self-reported and not included.
METR measured Grok 4 (released July 2025) at a 50% time horizon of about 105 minutes (95% CI 48–208 min, METR-Horizon v1.0). METR lists no newer Grok model; recent leaders there are 2026 Claude, Gemini and GPT models.
