AI agent benchmark leaderboard (November 2026)
As of 13 Nov 2026, BenchLeader tracks 66 AI agent benchmarks from 25 publishers, 17 of which feed the BenchLeader Index. Across agents & tools benchmarks, Claude Fable 5.2 (Anthropic) leads with an agents score of 82.9, ahead of Claude Opus 4.5 (70.6) and Claude Sonnet 5.5 (70.6). On the most-weighted live board, τ³-Banking is led by Qwen3.8 Max (max) at 51.3%. Of the 66 boards, 57 are active, 6 near-saturated, 2 saturated and 1 retired (retired: no new result in a year).
Last updated · rebuilt every six hours · independent publishers only, no launch-post numbers
- benchmarks
- 66
- active
- 57
- saturated or near
- 8
- publishers
- 25
Other categories: AI coding benchmarks · Reasoning and math benchmarks · Multimodal AI benchmarks · all LLM benchmarks
Which models lead agents & tools benchmarks
Ranked on the agents & tools category of the BenchLeader Index: every board below that feeds the index, standardised (50 is the average model) and averaged. One row per model, at its best reasoning effort; models need results on at least 3 of these boards.
| # | Model | Provider | Agents score | Index | Blended $/M |
|---|---|---|---|---|---|
| 1 | Claude Fable 5.2 | Anthropic | 82.9 | 65.9 | $16.00 |
| 2 | Claude Opus 4.5 | Anthropic | 70.6 | 57.7 | $10.00 |
| 3 | Claude Sonnet 5.5 (max) | Anthropic | 70.6 | 64.7 | $4.00 |
| 4 | GPT-5.5 (high) | OpenAI | 70.1 | 63.5 | $11.25 |
| 5 | GPT-5.6 Sol (xhigh) | OpenAI | 69.6 | 63.9 | $8.00 |
| 6 | Gemini 4 Argon | 68.8 | 64.3 | $4.00 | |
| 7 | Claude Opus 4.6 | Anthropic | 66.5 | 62.1 | $10.00 |
| 8 | Claude Opus 5 (xhigh) | Anthropic | 66.2 | 60.1 | $10.00 |
| 9 | Claude Fable 5 (max) | Anthropic | 66.0 | 65.0 | $20.00 |
| 10 | Claude Fable 5.1 | Anthropic | 65.7 | 66.4 | $20.00 |
Every agents & tool use benchmark, what it measures and who leads (66)
Agent benchmarks put the model in an environment and score the outcome: a shell (Terminal-Bench), a customer-service conversation with tools and an uncooperative user (τ²-bench), function calling (BFCL), the Model Context Protocol (MCP Atlas) or a real desktop (OSWorld).
Use them for: tool-calling products, automation and long multi-step tasks.
Caveat: Harness effects are large: the same model can move 10–20 percentage points with a different scaffold, prompt or tool set, so scores from different harnesses are not comparable.
| Benchmark | What it measures | Leader today | Status | Models | Newest result |
|---|---|---|---|---|---|
| τ³-Banking (AA) | Artificial Analysis' τ²-Bench banking-domain run: a support agent using tools with a simulated customer. | Qwen3.8 Max (max) 51.3% | ActiveHarness-dependent | 208 | 5 Nov 2026 |
| Terminal-Bench | Command-line agent tasks; best result per model across harnesses. | GPT-5.5 84.7% | ActiveHarness-dependent | 106 | 12 Nov 2026 |
| BFCL Overall | Berkeley Function Calling Leaderboard overall accuracy. | Claude Opus 4.5 77.5% | ActiveHarness-dependent | 83 | – |
| MCP Atlas | Real-world tool use through the Model Context Protocol. Scale AI. | Gemini 4 Argon 90.7% | Near-saturatedHarness-dependent | 37 | 6 Nov 2026 |
| τ²-bench | Mean pass^1 across the airline, retail and telecom domains. | Qwen3.5 397B A17B (thinking) 87.9% | ActiveHarness-dependent | 8 | 2 Mar 2026 |
| Terminal-Bench Hard | Hard subset of Terminal-Bench in a fixed harness. Run by Artificial Analysis. | GPT-5.6 Sol (max) 65.9% | ActiveHarness-dependent | 411 | 9 Jul 2026 |
| Terminal-Bench 2.1 (Vals) | Terminal-Bench 2.1 in Vals AI's fixed harness. | Claude Opus 5.5 87.6% | Near-saturatedHarness-dependent | 76 | 1 Oct 2026 |
| APEX-Agents | Long-horizon professional agent tasks (Mercor). | Gemini 4 Carbon 84.2% | ActiveHarness-dependent | 65 | 10 Nov 2026 |
| Arena Agent | Task-outcome score (0–100) from real agentic sessions judged by their users. | Claude Fable 5.2 (max) 30.4 | ActiveJudge-based | 58 | 11 Nov 2026 |
| HiL-Bench | Whether agents notice information gaps and ask clarifying questions. Scale AI. | Claude Fable 5.2 80.0% | ActiveHarness-dependent | 35 | 13 Nov 2026 |
| DeepSec-Bench (AA) | Deeper security engineering tasks than a single patch: reason about a system and defend it. Run… | GPT-6 Sol (max) 46.3% | ActiveHarness-dependent | 24 | 3 Nov 2026 |
| CWE-Bench (AA) | Finding and fixing the vulnerability classes CWE catalogues, in real repositories. Run by Artificial Analysis. | Hy4 75.8% | ActiveHarness-dependent | 23 | 3 Nov 2026 |
| CyberGym end-to-end (AA) | A full exploit-and-patch cycle in one run, scored end to end. Run by Artificial Analysis. | Kimi K3.1 Pro 94.7% | Near-saturatedHarness-dependent | 23 | 3 Nov 2026 |
| Cybench | Unguided capture-the-flag cybersecurity tasks. | Claude Opus 4.6 93.0% | Near-saturatedHarness-dependent | 21 | 5 Feb 2026 |
| Remote Labor Index | Real freelance projects completed end to end (Scale AI / CAIS). | Claude Fable 5.2 41.3% | ActiveHarness-dependent | 18 | 9 Nov 2026 |
| OSWorld-Verified 2.0 | Computer-use tasks in a real desktop environment. | Claude Opus 5 (max) 31.4% | ActiveHarness-dependent | 14 | 24 Jul 2026 |
| GDPval | Win rate against industry professionals on economically valuable tasks (OpenAI). | GPT-5.2 (no reasoning) 49.7% | ActiveJudge-based | 11 | 11 Dec 2025 |
| AutomationBenchnot in index | Office automation tasks run end to end. Scored on partial credit, so a model that gets… | Gemini 4 Carbon (high) 77.8% | ActiveHarness-dependent | 261 | 12 Nov 2026 |
| GDP.pdf (AA)not in index | Document work on real PDFs, scored all-or-nothing: every required element has to be right for the… | Gemini 4 Hydrogen (medium) 32.8% | ActiveHarness-dependent | 259 | 12 Nov 2026 |
| AA-Briefcase v1.1not in index | Knowledge work delivered as real files — spreadsheets, decks, documents — graded head to head, so… | Claude Sonnet 5.5 (max) 1816 | ActiveHarness-dependent | 173 | 12 Nov 2026 |
| GDPval-AA v2.1not in index | Artificial Analysis' run of OpenAI's GDPval professional tasks with shell and web access in an agentic… | Claude Opus 5.5 (max) 1866 | ActiveHarness-dependent | 171 | 9 Nov 2026 |
| GDPval-AA v2.1, normalised scorenot in index | Artificial Analysis' normalised run of OpenAI's GDPval professional-task comparison, frozen at the last percentages it published… | Claude Opus 5.5 (max) 68.3% | ActiveJudge-based | 167 | 23 Oct 2026 |
| Terminal-Bench-Science 0.1 (AA)not in index | Scientific computing tasks at a terminal: set up the environment, run the analysis, get the answer… | Claude Fable 5.2 (max) 73.3% | ActiveHarness-dependent | 144 | 12 Nov 2026 |
| Legal Research Benchnot in index | Agentic US legal research questions. Run by Vals AI. | Gemini 4 Carbon (high) 55.8% | ActiveHarness-dependent | 92 | 11 Nov 2026 |
| Tax Agent Benchnot in index | Agentic US corporate tax research. Run by Vals AI. | Gemini 4 Hydrogen (xhigh) 52.3% | ActiveHarness-dependent | 83 | 11 Nov 2026 |
| Finance Agent v2not in index | Core financial-analyst tasks with tools. Run by Vals AI. | Gemini 4 Argon (high) 65.4% | ActiveHarness-dependent | 80 | 29 Oct 2026 |
| Harvey's Legal Agent Benchmark v1not in index | The first version of the Harvey legal agent benchmark, which Vals AI keeps online. Run by… | Muse Spark 1.2 (xhigh) 25.4% | ActiveHarness-dependent | 80 | 18 Oct 2026 |
| Design Arena Agenticnot in index | Models working as agents build web apps, slide decks, games and 3D scenes; visitors vote between… | GPT-6.1 Sol (max) 1420 | ActiveJudge-based | 79 | 13 Nov 2026 |
| Vending-Bench 2not in index | Running a simulated vending-machine business for a year. | Gemini 4 Carbon 16484.2 | ActiveHarness-dependent | 72 | 10 Nov 2026 |
| LiveBench Agentic Codingnot in index | Agentic coding category of LiveBench 1.0, frozen at its final release (25 June 2026). LiveBench 2.0… | DeepSeek V4.1 Flash (max) 77.3% | ActiveHarness-dependent | 68 | 25 Jun 2026 |
| Terminal-Bench 2.0 (Vals)not in index | Terminal-Bench 2.0 in Vals AI's fixed harness. Run by Vals AI. | GPT-5.5 (high) 73.2% | ActiveHarness-dependent | 66 | 4 Jun 2026 |
| Terminal-Bench 4.0 (Vals)not in index | Terminal-Bench 4.0, the frontier-difficulty set, in Vals AI's harness. Run by Vals AI. | Claude Fable 5.2 65.2% | ActiveHarness-dependent | 62 | 11 Nov 2026 |
| Analyst Agent (AA)not in index | Multi-step financial-analyst tasks with tools, run by Artificial Analysis. | Gemini 3.7 Flash (high) 60.0% | ActiveHarness-dependent | 52 | 5 Nov 2026 |
| EnterpriseOps-Gymnot in index | Enterprise operations tasks in a simulated company: tickets, approvals and systems a model has to drive… | Claude Fable 5 (max) 51.1% | ActiveHarness-dependent | 51 | 5 Nov 2026 |
| CyberBenchnot in index | Finding and patching security bugs in real code. Run by Vals AI. | Muse Wonder (max) 80.4% | ActiveHarness-dependent | 48 | 5 Nov 2026 |
| ITBench SRE (AA)not in index | Site-reliability incident tasks in a live IT environment, run by Artificial Analysis. | GPT-5.6 Sol (max) 56.2% | ActiveHarness-dependent | 45 | 22 Sep 2026 |
| Terminal-Bench Sciencenot in index | Expert-authored scientific research workflows in a terminal. Run by Vals AI. | GPT-6 Astra (max) 62.9% | ActiveHarness-dependent | 45 | 9 Nov 2026 |
| Arena Searchnot in index | Pairwise votes on answers that used web search. | GPT-5.6 Sol (xhigh) 1255 | ActiveJudge-based | 42 | 9 Nov 2026 |
| BrowseCompnot in index | Hard questions whose answers are buried deep on the open web: the model has to browse… | Gemini 3.1 Pro 31.2% | ActiveHarness-dependent | 41 | 25 Mar 2026 |
| METR Time Horizonsnot in index | METR's task suite behind the 50%-time-horizon measure. | Claude Mythos Preview 85.2% | ActiveHarness-dependent | 41 | 7 Apr 2026 |
| DeepResearch Benchnot in index | Long-form research reports judged for depth and accuracy. | Claude Opus 4.6 (high) 55.3% | ActiveJudge-based | 40 | 28 May 2026 |
| Finance Agent v2.1not in index | Finance Agent v2.1, the version Vals AI now runs new models on. Run by Vals AI. | Claude Fable 5.2 69.5% | ActiveHarness-dependent | 39 | 11 Nov 2026 |
| BALROGnot in index | Progress through classic text and grid games such as NetHack. | GPT-6 Astra (max) 68.3% | ActiveHarness-dependent | 37 | 3 Sep 2026 |
| Harvey's Legal Agent Benchmark v1.1not in index | Agentic legal work over files, written with Harvey. Run by Vals AI. | Gemini 4 Argon (high) 27.1% | ActiveHarness-dependent | 37 | 11 Nov 2026 |
| SREBenchnot in index | Site-reliability incidents to diagnose and fix. Run by Vals AI. | GPT-6 Astra (max) 56.9% | ActiveHarness-dependent | 37 | 5 Nov 2026 |
| LiveBench Browsernot in index | LiveBench 2.0 browser-use suite. | GPT-6.1 Sol 93.3% | Near-saturatedHarness-dependent | 34 | 6 Nov 2026 |
| LiveBench Long Agenticnot in index | LiveBench 2.0 long-horizon agentic suite. | Claude Fable 5.2 82.9% | ActiveHarness-dependent | 34 | 6 Nov 2026 |
| SkillsBenchnot in index | How much reusable skills help an agent complete tasks. Run by Vals AI. | DeepSeek V4.1 Flash (high) 69.8% | ActiveHarness-dependent | 34 | 29 Oct 2026 |
| APEX-Agents (AA)not in index | Artificial Analysis' run of Mercor's long-horizon professional agent tasks. | Gemini 3.5 Flash (high) 47.0% | ActiveHarness-dependent | 33 | 18 Sep 2026 |
| Harvey LAB-AA v1.1not in index | Legal work drawn from practice, built with Harvey. Run by Artificial Analysis. | Muse Wonder Full (max) 13.1% | ActiveHarness-dependent | 33 | 9 Nov 2026 |
| τ³-bench Bankingnot in index | pass^1 on the banking domain, which Sierra added after the original three. A support agent working… | Qwen3.8 Max (max) 55.1% | ActiveHarness-dependent | 27 | 4 Aug 2026 |
| GBAEvalnot in index | Playing Game Boy Advance games as an agent. | Claude Opus 5 79.6% | ActiveHarness-dependent | 23 | 24 Jul 2026 |
| BrowseComp-Plusnot in index | BrowseComp questions answered against a fixed 100k-document corpus instead of the live web, so the retriever… | GPT-5 57.6% | ActiveHarness-dependent | 21 | 31 Dec 2025 |
| GDPxnot in index | SemiAnalysis' agentic benchmark: 206 long tasks across 11 fields such as chip design, cybersecurity, biology and… | Claude Fable 5.2 (max) 16.4% | ActiveHarness-dependent | 21 | 9 Nov 2026 |
| PostTrainBenchnot in index | Post-training a small model to a target, end to end. | Claude Fable 5.1 (max) 44.6% | ActiveHarness-dependent | 20 | 13 Oct 2026 |
| Poker Agentnot in index | AI agents playing poker against each other. Run by Vals AI. | GPT-5.2 1131.8% | ActiveHarness-dependent | 17 | 23 Dec 2025 |
| Time Horizon Index: KSPnot in index | Running a Kerbal Space Program mission end to end. Run by Vals AI. | Claude Opus 5.5 91.3% | Near-saturatedHarness-dependent | 13 | 7 Oct 2026 |
| DeepSearchQAnot in index | Research questions with many separate facts to gather; scored on answers that are fully correct, so… | Gemini Deep Research Agent 66.1% | ActiveHarness-dependent | 12 | 13 Apr 2026 |
| CUA-benchnot in index | Playing six commercial video games with only a keyboard and a mouse; half are held out… | GPT-6 Astra (max) 19.2% | ActiveHarness-dependent | 9 | 13 Oct 2026 |
| ExploitBenchnot in index | Developing working exploits for known vulnerabilities. | Claude Mythos Preview 73.8% | ActiveHarness-dependent | 9 | 23 Apr 2026 |
| Robo-benchnot in index | Controlling a robot arm and a humanoid in a real-time physics simulation from motor commands and… | GPT-6.1 Sol (max) 28.3% | ActiveHarness-dependent | 7 | 13 Oct 2026 |
| Web Search Indexnot in index | Answering questions that need live web search, with native and independent tools. Run by Vals AI. | Claude Opus 5.5 53.8% | ActiveHarness-dependent | 5 | 4 Nov 2026 |
| RoboDojo (LLM as robot policy)not in index | A general LLM controls a simulated robot arm directly, with no robot training, across 42 manipulation… | GPT-6 Astra 29.0 | ActiveHarness-dependent | 2 | 3 Nov 2026 |
| τ²-Bench Telecom (AA)not in index | Artificial Analysis' τ²-Bench telecom run. Shown for reference. | GLM 5.2 (max) 99.1% | SaturatedHarness-dependent | 418 | 9 Jul 2026 |
| FACTS Searchnot in index | Using web search as a tool and synthesising what comes back without inventing anything. Published by… | Gemini 4 Carbon 94.4% | SaturatedHarness-dependent | 26 | 12 Nov 2026 |
| TheAgentCompanynot in index | Office tasks in a simulated software company. | DeepSeek V3.2 42.9% | RetiredHarness-dependent | 14 | 29 Sep 2025 |
AI agent benchmarks: questions people ask
- What are the best AI agent benchmarks?
- The most informative AI agent benchmarks today are τ³-Banking (led by Qwen3.8 Max (max) at 51.3%), Terminal-Bench (led by GPT-5.5 at 84.7%) and BFCL Overall (led by Claude Opus 4.5 at 77.5%): each feeds the BenchLeader Index, covers many models and has not been saturated. Use them for: tool-calling products, automation and long multi-step tasks.
- Which AI model is best on AI agent benchmarks?
- As of 13 Nov 2026: Across agents & tools benchmarks, Claude Fable 5.2 (Anthropic) leads with an agents score of 82.9, ahead of Claude Opus 4.5 (70.6) and Claude Sonnet 5.5 (70.6). A category score standardises every board (50 is the average model, 15 points one standard deviation) and averages them, so one board cannot decide it.
- Are AI agent benchmarks comparable across boards?
- Only within one harness. An agent benchmark scores the model together with its scaffold: the prompt, the tools it is given, the number of turns allowed and how failures are retried. The same model can move 10 to 20 points between harnesses, so compare models on one board, run by one publisher, rather than across boards.
- Which AI agent benchmarks are saturated?
- As of 13 Nov 2026, MCP Atlas (near-saturated), Terminal-Bench 2.1 (near-saturated), CyberGym end-to-end (near-saturated), Cybench (near-saturated), LiveBench Browser (near-saturated), Time Horizon Index: KSP (near-saturated), τ²-Bench Telecom (saturated) and FACTS Search (saturated). A board is saturated when its fifth-best model scores 90% or more and near-saturated when the leader reaches 90% or the fifth-best 80%; both are decided from the data and change as scores do.
- How often is this leaderboard updated?
- BenchLeader re-reads every publisher's results every six hours (runs start at 04:03, 10:03, 16:03 and 22:03 UTC), and this page is rebuilt with each update. Each board's newest result date is in the table; publishers update on their own schedules.
Related
Cite as: BenchLeader, “AI agent benchmark leaderboard (November 2026)”, data as of 13 Nov 2026, https://www.benchleader.com/benchmarks/agents.