The agent leaderboard.
Which AI agents perform best on credible, maintainer-verified benchmarks — SWE-bench Verified, Terminal-Bench and Princeton’s HAL — with sources, dates and a transparent method.
Data retrieved 2026-11-01. Only official, maintained leaderboards with maintainer-checked or independently run results are used. Vendor self-reported scores are excluded.
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.
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.
Sources: [1] Pandaily · DeepSeek Harness Super-agent [2] Pandaily · HarnessPro (Tsinghua, Zhipu et al.) [3] OpenAI · GPT-6.1 Terra [4] Terminal-Bench official leaderboard [5] SWE-bench official leaderboard [6] HAL — Holistic Agent Leaderboard, Princeton
Composite rank
Not published today: only 4 models appear on three or more of the verified boards, below our threshold of 5. The boards cover different model generations, so a composite would mostly rank older models. See each benchmark below instead.
Composite method
For each benchmark we take the top 15 verified results and convert each model’s best position to a percentile (1 = top, 0 = last in that list). Model names are normalised (dates and effort settings removed). The composite is the mean percentile across benchmarks, shown only for models present on at least three benchmarks and only when at least 5 models qualify. Missing benchmarks are left blank, never imputed. Scaffolds differ between submissions, so a composite compares agent-plus-model configurations, not models alone.
SWE-bench Verified
% resolved (500 human-validated GitHub issues). Maintainer-checked submissions only (logs and trajectories reviewed by the SWE-bench team). Retrieved 2026-11-01. Official leaderboard
| # | Agent / scaffold | Model | Score | Submitted |
|---|---|---|---|---|
| 1 | live-SWE-agent + Gemini 3 Pro Preview (2025-11-18) · UIUC | Gemini 3 Pro Preview | 77.4% | 2025-11-20 |
| 2 | Claude 4.5 Opus (20251101) (medium) · SWE-agent | Claude 4.5 Opus | 74.4% | 2025-11-24 |
| 3 | Gemini 3 Pro Preview (2025-11-18) · SWE-agent | Gemini 3 Pro Preview | 74.2% | 2025-11-18 |
| 4 | OpenHands + GPT-5 · OpenHands | GPT-5 | 71.8% | 2025-08-07 |
| 5 | Moatless Tools + Claude 4 Sonnet · Moatless AI | Claude 4 Sonnet | 70.8% | 2025-06-11 |
| 6 | Claude 4.5 Sonnet (20250929) · SWE-agent | Claude 4.5 Sonnet | 70.6% | 2025-09-29 |
| 7 | OpenHands + Claude 4 Sonnet · OpenHands | Claude 4 Sonnet | 70.4% | 2025-05-24 |
| 8 | Claude 4 Opus (20250514) · SWE-agent | Claude 4 Opus | 67.6% | 2025-08-02 |
| 9 | SWE-agent + Claude 4 Sonnet · SWE-agent | Claude 4 Sonnet | 66.6% | 2025-05-22 |
| 10 | OpenHands · OpenHands | Undisclosed | 65.8% | 2025-04-15 |
| 11 | OpenHands + Kimi K2 · OpenHands | Kimi K2 | 65.4% | 2025-07-16 |
| 12 | GPT 5 (2025-08-07) (medium) · SWE-agent | GPT 5 | 65% | 2025-08-07 |
| 13 | Claude 4 Sonnet (20250514) · SWE-agent | Claude 4 Sonnet | 64.93% | 2025-07-26 |
| 14 | Kimi K2 Thinking · SWE-agent | Kimi K2 Thinking | 63.4% | 2025-12-10 |
| 15 | SWE-agent + Claude 3.7 Sonnet w/ Review Heavy · SWE-agent | Claude 3.7 Sonnet w/ Review Heavy | 62.4% | 2025-02-25 |
Terminal-Bench 4.0
% tasks solved in a real terminal (mean over trials, ±95% CI). Official maintainer leaderboard; results shown by maintainers with trial counts and confidence intervals. Retrieved 2026-11-01. Official leaderboard
| # | Agent / scaffold | Model | Score | Submitted |
|---|---|---|---|---|
| 1 | Claude Code · Anthropic | Opus 5.5 | 64.8% ±3.1 | 2026-09-22 |
| 2 | Claude Code · Anthropic | Sonnet 5.5 | 61.8% ±2.9 | 2026-09-28 |
| 3 | Codex · OpenAI | GPT-6 Astra | 58.2% ±2.8 | 2026-09-03 |
| 4 | Codex · OpenAI | GPT-6.1 Sol | 58.2% ±3.1 | 2026-09-29 |
| 5 | Claude Code · Anthropic | Fable 5.1 | 57.9% ±3.8 | 2026-09-01 |
| 6 | Codex · OpenAI | GPT-6 Astra | 57.9% ±2.7 | 2026-09-03 |
| 7 | Codex · OpenAI | GPT-6 Astra | 57.9% ±3.0 | 2026-09-03 |
| 8 | Claude Code · Anthropic | Fable 5.1 | 57.9% ±3.4 | 2026-09-01 |
| 9 | Claude Code · Anthropic | Fable 5.1 | 54.5% ±3.4 | 2026-09-01 |
| 10 | Codex · OpenAI | GPT-6 Astra | 54.2% ±2.7 | 2026-09-03 |
| 11 | Claude Code · Anthropic | Fable 5.1 | 53.9% ±3.4 | 2026-09-01 |
| 12 | Claude Code · Anthropic | Opus 5 | 53.9% ±3.2 | 2026-07-24 |
| 13 | Claude Code · Anthropic | Opus 5 | 51.8% ±3.4 | 2026-07-24 |
| 14 | Codex · OpenAI | GPT-6 Astra | 50.6% ±2.8 | 2026-09-03 |
| 15 | Claude Code · Anthropic | Opus 5 | 50.3% ±3.7 | 2026-07-24 |
GAIA (HAL)
% accuracy on GAIA validation tasks. Independent: run by Princeton HAL on standardised harnesses; only rows marked Verified are kept. Retrieved 2026-11-01. Official leaderboard
| # | Agent / scaffold | Model | Score | Submitted |
|---|---|---|---|---|
| 1 | HAL Generalist Agent Pareto optimal | Claude Sonnet 4.5 (September 2025) | 74.55% | — |
| 2 | HAL Generalist Agent | Claude Sonnet 4.5 High (September 2025) | 70.91% | — |
| 3 | HAL Generalist Agent | Claude Opus 4.1 High (August 2025) | 68.48% | — |
| 4 | HAL Generalist Agent | Claude Opus 4 High (May 2025) | 64.85% | — |
| 5 | HAL Generalist Agent | Claude-3.7 Sonnet High (February 2025) | 64.24% | — |
| 6 | HAL Generalist Agent | Claude Opus 4.1 (August 2025) | 64.24% | — |
| 7 | HF Open Deep Research | GPT-5 Medium (August 2025) | 62.8% | — |
| 8 | HAL Generalist Agent | GPT-5 Medium (August 2025) | 59.39% | — |
| 9 | HAL Generalist Agent Pareto optimal | o4-mini Low (April 2025) | 58.18% | — |
| 10 | HF Open Deep Research | Claude Opus 4 (May 2025) | 57.58% | — |
| 11 | HAL Generalist Agent | Claude-3.7 Sonnet (February 2025) | 56.36% | — |
| 12 | HAL Generalist Agent | Claude Haiku 4.5 (October 2025) | 56.36% | — |
| 13 | HF Open Deep Research | o4-mini High (April 2025) | 55.76% | — |
| 14 | HAL Generalist Agent Pareto optimal | o4-mini High (April 2025) | 54.55% | — |
| 15 | HF Open Deep Research | GPT-4.1 (April 2025) | 50.3% | — |
τ-bench Airline (HAL)
% tasks completed. Independent: run by Princeton HAL on standardised harnesses; only rows marked Verified are kept. Retrieved 2026-11-01. Official leaderboard
| # | Agent / scaffold | Model | Score | Submitted |
|---|---|---|---|---|
| 1 | TAU-bench Tool Calling Pareto optimal | o4-mini High (April 2025) | 56% | — |
| 2 | HAL Generalist Agent | Claude-3.7 Sonnet (February 2025) | 56% | — |
| 3 | TAU-bench Tool Calling | o3 Medium (April 2025) | 54% | — |
| 4 | HAL Generalist Agent | Claude Opus 4.1 (August 2025) | 54% | — |
| 5 | TAU-bench Tool Calling | Claude-3.7 Sonnet High (February 2025) | 52% | — |
| 6 | TAU-bench Tool Calling | Claude Opus 4.1 High (August 2025) | 52% | — |
| 7 | TAU-bench Tool Calling | Claude Opus 4.1 (August 2025) | 50% | — |
| 8 | TAU-bench Tool Calling | GPT-5 Medium (August 2025) | 48% | — |
| 9 | TAU-bench Tool Calling Pareto optimal | DeepSeek V3 (March 2025) | 44% | — |
| 10 | TAU-bench Tool Calling | Claude-3.7 Sonnet (February 2025) | 44% | — |
| 11 | HAL Generalist Agent | Claude-3.7 Sonnet High (February 2025) | 44% | — |
| 12 | HAL Generalist Agent | Claude Opus 4 (May 2025) | 44% | — |
| 13 | HAL Generalist Agent | Claude Opus 4 High (May 2025) | 44% | — |
| 14 | TAU-bench Tool Calling | o4-mini Low (April 2025) | 36% | — |
| 15 | TAU-bench Tool Calling | GPT-4.1 (April 2025) | 36% | — |
Online Mind2Web (HAL)
% web tasks completed. Independent: run by Princeton HAL on standardised harnesses; only rows marked Verified are kept. Retrieved 2026-11-01. Official leaderboard
| # | Agent / scaffold | Model | Score | Submitted |
|---|---|---|---|---|
| 1 | SeeAct Pareto optimal | GPT-5 Medium (August 2025) | 42.33% | — |
| 2 | Browser-Use | Claude Sonnet 4 (May 2025) | 40% | — |
| 3 | Browser-Use | Claude-3.7 Sonnet High (February 2025) | 39.33% | — |
| 4 | Browser-Use | Claude Sonnet 4 High (May 2025) | 39.33% | — |
| 5 | SeeAct | o3 Medium (April 2025) | 39% | — |
| 6 | Browser-Use | Claude-3.7 Sonnet (February 2025) | 38.33% | — |
| 7 | SeeAct | Claude Sonnet 4 (May 2025) | 36.67% | — |
| 8 | SeeAct | Claude Sonnet 4 High (May 2025) | 36.67% | — |
| 9 | Browser-Use | GPT-4.1 (April 2025) | 36.33% | — |
| 10 | Browser-Use | DeepSeek V3 (March 2025) | 32.33% | — |
| 11 | SeeAct | o4-mini High (April 2025) | 32% | — |
| 12 | Browser-Use | GPT-5 Medium (August 2025) | 32% | — |
| 13 | SeeAct | o4-mini Low (April 2025) | 31.67% | — |
| 14 | SeeAct | GPT-4.1 (April 2025) | 30.33% | — |
| 15 | SeeAct | Claude-3.7 Sonnet High (February 2025) | 30.33% | — |
Standards used and watched
SWE-bench Verified (official leaderboard)
benchmark · in leaderboardPrinceton/Stanford-maintained; 500 human-validated GitHub issues; maintainers check submitted logs.
Terminal-Bench (official leaderboard)
benchmark · in leaderboardStanford/Laude Institute; agents act in real terminals; multiple trials with 95% confidence intervals.
HAL — Holistic Agent Leaderboard (Princeton)
benchmark · in leaderboardIndependent third-party runs on standardised harnesses, with cost; covers GAIA, τ-bench, Online Mind2Web, SWE-bench Verified Mini and more.
GAIA leaderboard (Hugging Face)
benchmark · watched, not scoredOfficial GAIA leaderboard on a held-out test set; used for reference, HAL runs used for scoring.
τ²-bench (Sierra Research)
benchmark · watched, not scoredMaintained tool-use and customer-service agent benchmark; scores currently taken from HAL’s independent τ-bench runs.
OSWorld / OSWorld-Verified
benchmark · watched, not scoredComputer-use benchmark on real operating systems; leaderboard site os-world.github.io was unreachable on 1 Nov 2026 — not yet automated.
METR — time horizons
benchmark · watched, not scoredIndependent measure of the length of tasks agents can complete; reported by METR, not yet automated here.
BrowseComp (OpenAI)
benchmark · watched, not scoredWeb-browsing benchmark; most published scores are vendor-reported, so excluded from scoring unless independently reproduced.
METR blog
researchIndependent evaluations of frontier-model agent capabilities.
Anthropic Engineering
labFirst-party notes on agent design and evaluations (self-reported).
LangChain blog (RSS)
industryAgent framework vendor; practitioner patterns (vendor perspective).
Simon Willison’s Weblog (Atom)
independentIndependent, hands-on testing of coding and browsing agents.
Latent Space (RSS)
newsletterAI-engineering newsletter and podcast with frequent agent coverage.