Skip to main content
Radar

Keep up with the models you depend on. Follow price changes, retirements, and API updates.Follow the models you depend on.

Follow model changes
Agentic benchmark report

Best LLMs for AgenticSeptember 2026 Leaderboard

Data refreshed:

Tool use, browser research, and computer-use workflows

As of September 2026, the top agentic model on the BenchLM leaderboard is Claude Fable 5.1 with a BenchAlign agentic score of 80.2.

Decision lens: the score determines position; Supported and Estimated labels describe the evidence behind that position without removing sparsely reported models.

Data refreshed
September 10, 2026
Ranked
152 of 489 models
Supported / Estimated
48 / 104
Weighted evidence
4 of 30 benchmarks
30 tracked benchmarks

Terminal-Bench 2.0, BrowseComp, OSWorld-Verified, OSWorld 2.0, CyberGym, CWE-Bench, Cybench, ExploitGym, JobBench, BrowseComp-VL, OSWorld, AndroidWorld, WebVoyager, MCP Atlas, Toolathlon, Finance Agent v2, GDPval-AA, ZClawBench, Tau2-Telecom, DeepSearchQA, Tau2-Airline, PinchBench, OpenHands Index, SWE-Atlas Refactoring, SWE Refactor Bench, AI4AI-Bench, BFCL v4, MLE-Bench Lite, MM-ClawBench, Gert Labs

Scope: Terminal/tool use, Browser research, Computer use

Best Agentic picks

BenchLM summaries for agentic plus the practical tradeoffs users check next: open weights, price, speed, latency, and context.

How BenchLM scores these

Agentic AI Leaderboard

Primary score: BenchAlign agentic score. Higher values rank first. Use the Show metric control to change the value shown in each row.

Updated Embed leaderboard

Supported positions have diverse direct evidence. Estimated positions remain ranked but carry wider uncertainty.

Filters
Supported positions have diverse direct evidence. Estimated positions remain ranked with wider uncertainty.
Rank / modelWeighted Agentic
1
Claude Fable 5.1Anthropic · ClosedSupported
80.2%
2
Claude Opus 5Anthropic · ClosedSupported
78.1%
3
Claude Fable 5Anthropic · ClosedSupported
74.6%
4
Kimi K3Moonshot AI · ClosedSupported
71.9%
5
GPT-6 AstraOpenAI · ClosedSupported
70.4%
6
GPT-5.6 SolOpenAI · ClosedSupported
70.0%
7
Grok 4.6xAI · ClosedSupported
69.5%
8
GLM-5.3Z.AI · Open weightSupported
68.4%
9
Qwen3.8 MaxAlibaba · Open weightSupported
67.2%
10
Gemini 3.8 FlashGoogle · ClosedSupported
66.3%
11
Claude Sonnet 5Anthropic · ClosedSupported
65.8%
12
Gemini 3.7 FlashGoogle · ClosedSupported
64.0%
13
Qwen3.8-27BAlibaba · Open weightSupported
63.4%
14
Claude Opus 4.8Anthropic · ClosedSupported
62.5%
15
GPT-5.3 CodexOpenAI · ClosedEstimated
61.6%
16
Muse Spark 1.2Meta · ClosedSupported
61.1%
17
GPT-5.5OpenAI · ClosedSupported
61.0%
18
Ornith-1.5-397BOrnith AI · Open weightEstimated
60.8%
19
Grok 4.5xAI · ClosedSupported
60.5%
20
GPT-5.6 TerraOpenAI · ClosedSupported
60.2%
21
GLM-5.3-FlashZ.AI · Open weightSupported
60.1%
22
Qwen 3.6 Max (preview)Alibaba · ClosedEstimated
59.7%
23
Muse Spark 1.1Meta · ClosedSupported
59.4%
24
GPT-5.1OpenAI · ClosedEstimated
59.4%
25
dots3-note PreviewDots Studio · Open weightEstimated
59.3%

Top AI Models for AgenticSeptember 2026

As of September 2026, Claude Fable 5.1 leads the BenchAlign agentic leaderboard with a score of 80.2, followed by Claude Opus 5 (78.1) and Claude Fable 5 (74.6). BenchLM is currently showing 48 Supported and 104 Estimated models in this category.

What changed

Claude Fable 5.1 ranks #1 at 80.2 with a Supported evidence label.

Claude Opus 5 ranks #2 at 78.1 with a Supported evidence label.

Claude Fable 5 ranks #3 at 74.6 with a Supported evidence label.

Top models by benchmark

Agentic software engineering and terminal task completion benchmark(30% of category score)

RankModelReported score
2Kimi K388.3

Score in Context

What these scores mean

BenchAlign places direct benchmarks and independent external signals on a common calibrated scale. The score is relative to the current evidence universe; it is not a raw percentage from any single test.

Known limitations

Estimated rows have less diverse direct evidence and wider uncertainty. They remain ranked so a newly released model is not treated as weak merely because fewer benchmark publishers have evaluated it.

How we weight

This lens combines category-relevant external evidence with admitted benchmark protocols. Evidence sources are calibrated for difficulty before aggregation, and no generated benchmark row contributes to the score.

Leaderboards exclude benchmark rows that BenchLM generated from other scores or cloned from reference models. When a weighted benchmark is missing after that filter, the category falls back to the remaining trustworthy public rows instead of filling the gap with synthetic values.

The full scoring rules, freshness handling, and runtime/pricing caveats live on the BenchLM methodology page.

Scroll horizontally to read the full evidence ledger.

Agentic benchmark weights, ranking status, and descriptions
BenchmarkWeightStatusDescription
Terminal-Bench 2.030%WeightedAgentic software engineering and terminal task completion benchmark
BrowseComp25%WeightedWeb research benchmark for browsing agents
OSWorld-Verified25%WeightedComputer-use benchmark for GUI task completion
OSWorld 2.010%WeightedA long-horizon computer-use benchmark covering realistic workflows across everyday and professional desktop tasks.
CyberGymDisplay onlyCybersecurity task benchmark for evaluating defensive cyber workflows and vulnerability-oriented agent performance.
CWE-BenchDisplay onlyExternal benchmark for evaluating whether coding agents can produce correct patches for real-world software vulnerabilities.
CybenchDisplay onlyA cybersecurity benchmark of professional Capture the Flag tasks for measuring autonomous cyber agent capability and risk.
ExploitGymDisplay onlyA controlled benchmark for evaluating whether AI agents can extend vulnerability-triggering inputs into working exploits.
JobBenchDisplay onlyAn occupational agent benchmark for professional workflows that workers say they most want delegated to AI.
BrowseComp-VLDisplay onlyVision-language browsing benchmark for multimodal web research and tool-use tasks.
OSWorldDisplay onlyComputer-use benchmark for GUI task completion across the broader OSWorld task suite.
AndroidWorldDisplay onlyAndroid GUI agent benchmark for task completion across mobile app workflows.
WebVoyagerDisplay onlyBrowser agent benchmark for completing multi-step workflows on live websites.
MCP AtlasDisplay onlyTool-calling benchmark for Model Context Protocol integrations and multi-tool coordination
ToolathlonDisplay onlyGeneral tool-calling benchmark for multi-step API and tool usage
Finance Agent v2Display onlyFinancial analysis and decision-making benchmark for agentic expert tasks.
GDPval-AADisplay onlyReal-world agentic knowledge-work evaluation reported as an Elo score.
ZClawBenchDisplay onlyZ.AI's OpenClaw workflow benchmark for broad agent tasks across research, office work, data analysis, devops, automation, and security.
Tau2-TelecomDisplay onlyTelecom-focused tool-use benchmark for structured API workflows
DeepSearchQADisplay onlyAgentic browsing benchmark for list-style question answering with browser tools.
Tau2-AirlineDisplay onlyAirline-domain tool-use benchmark for structured workflow execution and API correctness.
PinchBenchDisplay onlyAn OpenClaw agent benchmark from Kilo that measures successful task completion across standardized real-world agent workflows.
OpenHands IndexDisplay onlyA holistic coding-agent benchmark that evaluates AI agents across issue resolution, frontend work, greenfield development, testing, and information gathering.
SWE-Atlas RefactoringDisplay onlyA Scale SWE-Atlas software-engineering agent benchmark focused on refactoring tasks.
SWE Refactor BenchDisplay onlyTests whether coding agents can complete long-horizon, whole-repository stack migrations while preserving the original program's behavior.
AI4AI-BenchDisplay onlyTests whether coding agents can improve the training algorithm inside an existing AI research codebase, then survive a sealed training run and held-out evaluation.
BFCL v4Display onlyFunction-calling benchmark for tool selection, schema adherence, and argument correctness.
MLE-Bench LiteDisplay onlyA lightweight machine-learning competition benchmark that measures whether models can iteratively train, evaluate, and improve ML systems in low-resource settings.
MM-ClawBenchDisplay onlyAn OpenClaw-derived agent benchmark covering practical work and life tasks such as office document delivery, research, planning, and code maintenance.
Gert LabsDisplay onlyComposite game-environment leaderboard score across Gert Labs agentic coding, one-shot coding, and social decision-making modes.

About Agentic Benchmarks

Agentic software engineering and terminal task completion benchmark

Common questions

What is an agentic LLM benchmark?

Agentic benchmarks evaluate whether AI models can complete multi-step workflows using tools, browsers, terminals, or software interfaces instead of only answering in chat.

Which benchmarks matter for AI agents?

Key agentic benchmarks include Terminal-Bench 2.0 for terminal tasks, BrowseComp for web research, and OSWorld-Verified for computer-use workflows.

Why do agentic benchmarks matter in 2026?

Agentic benchmarks matter because many modern products rely on models that can browse, plan, use tools, and complete end-to-end tasks rather than only generate text.

Agentic benchmark updates

Agentic is the fastest-moving category. Don't fall behind.

Read a sample issue

Join 2,000+ readers.

One email each week. Unsubscribe anytime.

Related