Agent & Tool-Use Benchmarks

Which AI models handle function calling, MCP tool use, browsing, and multi-step agent workflows best? Rankings across 24 agentic benchmarks.

Agentic carries 22% weightin BenchLM.ai's overall score — the single biggest category.

Best Agentic Model

GPT-5.4 Pro

Score: 87.7 · OpenAI

Best Open-Weight Agent

GLM-5 (Reasoning)

Score: 78.3 · Zhipu AI

Benchmarks Tracked

24 benchmarks

Terminal, browsing, tool-use, and computer-use

Benchmark Categories

Core Weighted (3)

These 3 benchmarks determine agentic rankings

Terminal-Bench 2.0 BrowseComp OSWorld-Verified

Tool Calling & MCP (6)

Function calling, MCP tool use, and structured workflows

MCP Atlas MCP-Tasks Toolathlon BFCL v4 Tau2-Telecom TAU3-Bench

Agent Frameworks (6)

OpenClaw-style and end-to-end agent evaluations

ZClawBench QwenClawBench QwenWebBench MM-ClawBench Claw-Eval PinchBench

Computer & Browser Use (5)

Desktop GUI, mobile, and browser navigation tasks

OSWorld AndroidWorld WebVoyager BrowseComp-VL DeepPlanning

Specialized (4)

Domain-specific agentic tasks across ML, research, and airline

MLE-Bench Lite VITA-Bench WideResearch Tau2-Airline

Top 15 Models by Weighted Agentic Score

OpenAIAnthropicGoogleMetaDeepSeekMistralxAIAlibaba

Core Benchmarks Radar — Top 5 Models

CSV JSON

#	Model
1	GPT-5.4 Pro OpenAI · Proprietary	87.7	90	89.3	84	92
2	GPT-5.2-Codex OpenAI · Proprietary	87	90	85	85	82
3	MiMo-V2-Pro Xiaomi · Proprietary	86.7	86.7	—	—	85
4	GPT-5.1-Codex-Max OpenAI · Proprietary	86	90	85	82	81
5	Holo3-122B-A10B H Company · Proprietary	78.9	—	—	78.85	70
6	GLM-5 (Reasoning) Zhipu AI · Open Weight	78.3	81	80	74	82
7	Grok 4.1 xAI · Proprietary	78.2	79	84	73	85
8	Gemini 3 Pro Deep Think Google · Proprietary	78.1	77	87	73	80
9	Holo3-35B-A3B H Company · Open Weight	77.8	—	—	77.8	69
10	GPT-5.4 OpenAI · Proprietary	77	75.1	82.7	75	82
11	Gemini 3.1 Pro Google · Proprietary	76.1	77	86	68	87
12	GPT-5.1 OpenAI · Proprietary	75.8	78	79	71	78
13	GPT-5 (medium) OpenAI · Proprietary	75.5	77	78	72	76
14	o1-preview OpenAI · Proprietary	75.4	77	79	71	72
15	GPT-5 (high) OpenAI · Proprietary	75.2	78	75	72	82
16	GPT-5.3 Codex OpenAI · Proprietary	74.4	—	88	64.7	85
17	Claude Opus 4.6 Anthropic · Proprietary	72.6	65.4	84	72.7	85
18	Claude Sonnet 4.6 Anthropic · Proprietary	72.5	—	—	72.5	84
19	Gemini 3 Pro Google · Proprietary	71.1	68	83	66	79
20	Grok 4.1 Fast xAI · Proprietary	71	74	73	66	70
21	o3-pro OpenAI · Proprietary	70.4	69	76	68	67
22	Qwen3.5 397B (Reasoning) Alibaba · Open Weight	70	—	—	70	77
23	o3 OpenAI · Proprietary	69.9	71	75	65	64
24	DeepSeek V3.2 (Thinking) DeepSeek · Open Weight	69.4	71	70	67	67
25	DeepSeek Coder 2.0 DeepSeek · Open Weight	67.5	73	62	65	62
26	o3-mini OpenAI · Proprietary	66.6	67	74	61	65
27	GPT-5.2 OpenAI · Proprietary	66.2	83	65.8	47.3	82
28	GPT-5.4 mini OpenAI · Proprietary	65.6	60	—	72.1	66
29	o1 OpenAI · Proprietary	65.4	66	72	60	64
30	Claude Opus 4.5 Anthropic · Proprietary	65.2	59.3	73	66.3	76
31	GPT-4.1 OpenAI · Proprietary	64.7	61	73	63	64
32	Qwen2.5-1M Alibaba · Open Weight	64.7	65	72	59	62
33	DeepSeekMath V2 DeepSeek · Open Weight	63.9	65	66	61	63
34	Nemotron 3 Ultra 500B NVIDIA · Open Weight	62.8	63	69	58	60
35	Qwen3.6 Plus Alibaba · Proprietary	62	61.6	—	62.5	69
36	MiMo-V2-Flash Xiaomi · Open Weight	61.8	63	65	58	67
37	Gemini 2.5 Pro Google · Proprietary	61.7	61	72	55	65
38	Composer 2 Cursor · Proprietary	61.7	61.7	—	—	62
39	GLM-4.7 Zhipu AI · Open Weight	61	—	—	61	74
40	Claude Sonnet 4.5 Anthropic · Proprietary	60	50	74	61.4	68
41	DeepSeek V3.2 DeepSeek · Open Weight	58.8	60	62	55	61
42	o4-mini (high) OpenAI · Proprietary	58.5	58	64	55	58
43	GLM-5 Zhipu AI · Open Weight	58.3	56.2	62	58	75
44	Qwen3.5 397B Alibaba · Open Weight	58.3	52.5	62	62.2	68
45	Grok 4 xAI · Proprietary	58.1	57	63	56	68
46	GLM-5V-Turbo Zhipu AI · Proprietary	58	—	51.9	62.3	69
47	Claude 4 Sonnet Anthropic · Proprietary	57.9	56	62	57	62
48	Qwen2.5-72B Alibaba · Open Weight	57.7	56	64	55	60
49	Kimi K2.5 (Reasoning) Moonshot AI · Proprietary	57.6	50.8	60.6	63.3	76
50	Kimi K2.5 Moonshot AI · Open Weight	57.6	50.8	60.6	63.3	72
51	Gemini 3 Flash Google · Proprietary	57.5	56	66	53	67
52	DeepSeek LLM 2.0 DeepSeek · Open Weight	57	57	—	—	57
53	MiniMax M2.7 MiniMax · Proprietary	57	57	—	—	66
54	Nemotron 3 Super 100B NVIDIA · Open Weight	56.6	56	61	54	56
55	GPT-4.1 mini OpenAI · Proprietary	56.5	54	71	49	57
56	Qwen3.5-122B-A10B Alibaba · Open Weight	56	49.4	63.8	58	71
57	Grok Code Fast 1 xAI · Proprietary	55.7	59	57	51	56
58	Claude 3.5 Sonnet Anthropic · Proprietary	55	54	62	51	55
59	Claude 4.1 Opus Thinking Anthropic · Proprietary	54	—	54	—	57
60	Llama 3.1 405B Meta · Open Weight	53	53	—	—	53
61	Claude 4.1 Opus Anthropic · Proprietary	52.8	43.3	62	57	62
62	Gemini 1.5 Pro Google · Proprietary	52.3	45	64	—	50
63	Mistral Large 2 Mistral · Proprietary	52.2	51	57	50	52
64	Kimi K2 Moonshot AI · Proprietary	52.1	47.1	60.2	—	53
65	Claude Haiku 4.5 Anthropic · Proprietary	51.9	41	62	57	63
66	Qwen3.5-27B Alibaba · Open Weight	51.6	41.6	61	56.2	70
67	GPT-4o mini OpenAI · Proprietary	50.9	58	49	44	54
68	Qwen3.5-35B-A3B Alibaba · Open Weight	50.5	40.5	61	54.5	66
69	Sarvam 105B Sarvam · Open Weight	49.5	—	49.5	—	60
70	Gemini 3.1 Flash-Lite Google · Proprietary	49.2	47	60	44	56
71	Mistral Large 3 Mistral · Proprietary	49	—	—	49	58
72	GPT-4o OpenAI · Proprietary	48.5	49	—	48	50
73	Claude 3 Opus Anthropic · Proprietary	48.1	44	56	47	49
74	Qwen3 235B 2507 (Reasoning) Alibaba · Open Weight	47.4	47	48	—	55
75	GPT-4.1 nano OpenAI · Proprietary	47.4	43	62	42	44
76	Nemotron Ultra 253B NVIDIA · Open Weight	46.7	46	50	45	41
77	Gemini 2.5 Flash Google · Proprietary	46.5	44	58	41	50
78	GPT-OSS 120B OpenAI · Open Weight	44.8	43	50	43	50
79	GPT-4 Turbo OpenAI · Proprietary	44.7	42	54	41	43
80	DeepSeek-R1 DeepSeek · Open Weight	44.5	42	49	44	45
81	DeepSeek V3.1 (Reasoning) DeepSeek · Open Weight	44.3	42	48	—	43
82	Claude 3 Haiku Anthropic · Proprietary	44	40	53	42	43
83	GPT-5.4 nano OpenAI · Proprietary	42.9	46.3	—	39	58
84	Z-1 Z · Proprietary	42.2	39	49	41	44
85	Moonshot v1 Moonshot AI · Proprietary	42.2	39	49	41	43
86	Nemotron-4 15B NVIDIA · Open Weight	41.3	37	47	42	42
87	Llama 3 70B Meta · Open Weight	41.2	37	48	41	44
88	Mistral 8x7B Mistral · Open Weight	41.1	40	47	38	44
89	Llama 4 Maverick Meta · Open Weight	40.9	37	51	38	39
90	Gemini 1.0 Pro Google · Proprietary	39.8	36	51	36	40
91	o1-pro OpenAI · Proprietary	39.7	40	50	32	45
92	Nemotron 3 Nano 30B NVIDIA · Open Weight	39.6	38	43	39	42
93	Llama 4 Scout Meta · Open Weight	39	39	—	—	44
94	Phi-4 Microsoft · Open Weight	38.3	44	35	34	40
95	Grok 3 [Beta] xAI · Proprietary	35.5	32	41	—	48
96	Sarvam 30B Sarvam · Open Weight	35.5	—	35.5	—	48
97	GPT-OSS 20B OpenAI · Open Weight	35.4	35	42	31	36
98	Nova Pro Amazon · Proprietary	34.9	—	39	32	33
99	Gemma 3 27B Google · Open Weight	34.4	29	42	35	35
100	DBRX Instruct Databricks · Open Weight	34.3	41	31	29	41
101	Qwen3 235B 2507 Alibaba · Open Weight	33.7	33	40	30	47
102	Llama 4 Behemoth Meta · Open Weight	33	33	—	—	34
103	DeepSeek V3.1 DeepSeek · Open Weight	32.9	29	39	33	41
104	Mixtral 8x22B Instruct v0.1 Mistral · Open Weight	31.8	35	32	28	36
105	GLM-4.5-Air Zhipu AI · Proprietary	31.5	28	37	—	38
106	GLM-4.5 Zhipu AI · Proprietary	28	28	—	—	40
107	Mistral 8x7B v0.2 Mistral · Open Weight	27.8	24	34	—	27
108	Mistral 7B v0.3 Mistral · Open Weight	26.4	24	32	25	29

Agentic score = weighted average of Terminal-Bench 2.0 (40%), OSWorld-Verified (35%), and BrowseComp (25%), normalized by available weights. Display-only benchmarks (MCP Atlas, Toolathlon, etc.) are tracked but do not affect rankings.

Agentic Category Rankings

Official BenchLM agentic leaderboard

Coding Benchmarks

SWE-bench, LiveCodeBench, and more

Price vs Performance

Find the best value for agent workloads

Frequently Asked Questions

What are LLM agent benchmarks?

Agent benchmarks test whether AI models can go beyond answering questions and actually complete multi-step tasks: browsing the web, writing and running code in a terminal, calling external APIs via function calling, and operating desktop or mobile interfaces. They measure real-world usefulness for autonomous workflows.

What is function calling and why does it matter?

Function calling (or tool use) lets an LLM invoke external tools, APIs, or databases as part of its response. This is critical for building AI agents that can search the web, query databases, send emails, or control other software. Benchmarks like BFCL v4 and Toolathlon specifically measure how reliably models select the right function and pass correct arguments.

What is MCP (Model Context Protocol)?

MCP is an open standard for connecting LLMs to external tools and data sources. MCP Atlas and MCP-Tasks benchmark how well models work with MCP-backed integrations. Strong MCP performance means a model integrates well into tool-rich agent architectures.

Why does agentic carry the most weight in BenchLM scores?

Agentic carries 22% of BenchLM's overall score because the ability to use tools, browse, and complete multi-step tasks is the strongest differentiator between models in production use. A model that scores well on knowledge but cannot reliably call functions or navigate software has limited real-world utility for agent workflows.

Which models are best for building AI agents?

Currently, GPT-5.4 Pro by OpenAI leads the weighted agentic rankings with a score of 87.7. The best open-weight agent model is GLM-5 (Reasoning) (78.3). Check the leaderboard above for the full ranking.

Weekly LLM Benchmark Digest

Get notified when new models drop, benchmark scores change, or the leaderboard shifts. One email per week.

Free. No spam. Unsubscribe anytime. We only store derived location metadata for consent routing.