Skip to main content

BenchLM recommendation

Best LLMs for AI Agents in 2026

Data verified

As of July 13, 2026, the top model in best llms for ai agents on the BenchLM leaderboard is Claude Mythos 5 with a score of 100.

Last verified: July 13, 2026

Agentic capability is BenchLM's heaviest-weighted category (22% of the overall score) because it is where models differ most in practice: tool calling, browsing, terminal work, and multi-step task execution. This page ranks models purely by their agentic category score, built from benchmarks like Terminal-Bench 2, BrowseComp, OSWorld-Verified, and tau2-bench.

Unless noted otherwise, ranking surfaces on this page use BenchLM's provisional leaderboard lane rather than the stricter sourced-only verified leaderboard.

Bottom line: Claude Mythos 5 and Claude Fable 5 lead agentic work, with GPT-5.5 the strongest non-Anthropic row. Agent harness quality matters as much as the model — test with your own scaffold.

According to BenchLM.ai, Claude Mythos 5 leads this ranking with a score of 100, followed by Claude Fable 5 (99.4) and GPT-5.5 (96.1). There is meaningful separation between the top models, suggesting genuine performance differences.

The best open-weight option is GLM-5 (Reasoning) (ranked #11 with a score of 87.6). Proprietary models hold a clear advantage in this category, though open-weight options may suffice for less demanding use cases.

This ranking is based on provisional weighted averages across the scoring benchmarks in agentic tracked by BenchLM.ai. For detailed model profiles, click any model name below. To compare two specific models head-to-head, use the "vs #" links.

What changed

Claude Mythos 5 tops the agentic category with a perfect weighted score.

Claude Fable 5 within a point of the lead — the strongest generally available agentic model.

GPT-5.5 the best non-Anthropic agentic row at 96.1.

How to choose

Full Rankings (120 models)

1
Claude Mythos 5
Anthropic·Proprietary·1M+

100

prov. avg

2
Claude Fable 5
Anthropic·Proprietary·1M+

99.4

prov. avg

3
GPT-5.5
OpenAI·Proprietary·1M

96.1

prov. avg

Claude Sonnet 5
Anthropic·Proprietary·1M

95.7

tracked

GPT-5.6 Sol
OpenAI·Proprietary·1M

94.7

tracked

6
Claude Opus 4.8
Anthropic·Proprietary·1M

93.6

prov. avg

GPT-5.6 Terra
OpenAI·Proprietary·1M

93.6

tracked

8
Gemini 3 Pro Deep Think
Google·Proprietary·2M

91.7

prov. avg

Grok 4.1
xAI·Proprietary·1M

89.9

tracked

GPT-5.6 Luna
OpenAI·Proprietary·1M

89.6

tracked

11
GLM-5 (Reasoning)
Z.AI·Open Weight·200K

87.6

prov. avg

12
GPT-5.4
OpenAI·Proprietary·1.05M

87.5

prov. avg

13
Gemini 3.5 Flash
Google·Proprietary·1M

86.5

prov. avg

o1-preview
OpenAI·Proprietary·200K

84.4

tracked

15
Claude Opus 4.7 (Adaptive)
Anthropic·Proprietary·1M

83.3

prov. avg

Muse Spark 1.1
Meta·Proprietary·1M

82.4

tracked

17
Gemini 3 Pro
Google·Proprietary·2M

81.7

prov. avg

18
Kimi K2.6
Moonshot AI·Open Weight·256K

80.2

prov. avg

19
Claude Opus 4.6
Anthropic·Proprietary·1M

79.2

prov. avg

20
MiniMax M3
MiniMax·Open Weight·1M

78.7

prov. avg

21
GPT-5.3 Codex
OpenAI·Proprietary·400K

78.1

prov. avg

22
DeepSeek V4 Pro (Max)
DeepSeek·Open Weight·1M

77.9

prov. avg

23
Gemini 3.1 Pro
Google·Proprietary·1M

77.2

prov. avg

24
GPT-5.1-Codex-Max
OpenAI·Proprietary·400K

75

prov. avg

25
GPT-5.2-Codex
OpenAI·Proprietary·400K

74

prov. avg

26
GPT-5 (high)
OpenAI·Proprietary·128K

73.4

prov. avg

Grok 4.1 Fast
xAI·Proprietary·1M

72.2

tracked

28
GPT-5.1
OpenAI·Proprietary·200K

71.7

prov. avg

29
Claude Opus 4.5
Anthropic·Proprietary·200K

71.4

prov. avg

30
GPT-5.4 mini
OpenAI·Proprietary·400K

71.1

prov. avg

GPT-5 (medium)
OpenAI·Proprietary·128K

70.8

tracked

32
Claude Sonnet 4.6
Anthropic·Proprietary·200K

70.4

prov. avg

33
DeepSeek V3.2 (Thinking)
DeepSeek·Open Weight·128K

70.3

prov. avg

34
Qwen3.7 Plus
Alibaba·Proprietary·1M

69.1

prov. avg

35
DeepSeek V4 Pro (High)
DeepSeek·Open Weight·1M

68.8

prov. avg

Qwen3.5 397B (Reasoning)
Alibaba·Open Weight·128K

66

tracked

37
DeepSeek V4 Flash (Max)
DeepSeek·Open Weight·1M

65.9

prov. avg

38
o1
OpenAI·Proprietary·200K

65.1

prov. avg

39
GPT-4.1
OpenAI·Proprietary·1M

62.5

prov. avg

40
GLM-5.1
Z.AI·Open Weight·203K

60.6

prov. avg

41
o3
OpenAI·Proprietary·200K

60.2

prov. avg

42
o3-mini
OpenAI·Proprietary·200K

59.7

prov. avg

43
Qwen3.6 Plus
Alibaba·Proprietary·1M

59.3

prov. avg

o3-pro
OpenAI·Proprietary·200K

59.3

tracked

Step 3.7 Flash
StepFun·Open Weight·256K

58.6

tracked

46
Kimi K2.5 (Reasoning)
Moonshot AI·Proprietary·128K

57.7

prov. avg

DeepSeek Coder 2.0
DeepSeek·Open Weight·128K

57

tracked

48
DeepSeek V3.2
DeepSeek·Open Weight·128K

55.8

prov. avg

49
Gemini 2.5 Pro
Google·Proprietary·1M

53.2

prov. avg

50
GPT-5.2
OpenAI·Proprietary·400K

52.6

prov. avg

Qwen2.5-1M
Alibaba·Open Weight·1M

52.6

tracked

52
Grok 4
xAI·Proprietary·128K

52.5

prov. avg

DeepSeekMath V2
DeepSeek·Open Weight·128K

51.3

tracked

54
GLM-5
Z.AI·Open Weight·200K

50.9

prov. avg

55
Claude Sonnet 4.5
Anthropic·Proprietary·200K

50.3

prov. avg

GPT-5.4 nano
OpenAI·Proprietary·400K

49.8

tracked

57
GLM-4.7
Z.AI·Open Weight·200K

49.7

prov. avg

58
Gemini 3 Flash
Google·Proprietary·1M

49.7

prov. avg

59
Kimi K2.5
Moonshot AI·Open Weight·256K

49.5

prov. avg

60
Qwen3.5 397B
Alibaba·Open Weight·128K

49.4

prov. avg

61
Qwen3.5-122B-A10B
Alibaba·Open Weight·262K

49.4

prov. avg

62
MiMo-V2-Flash
Xiaomi·Open Weight·256K

49.2

prov. avg

GLM-5V-Turbo
Z.AI·Proprietary·200K

48.6

tracked

DeepSeek LLM 2.0
DeepSeek·Open Weight·128K

47.7

tracked

65
Nemotron 3 Ultra
NVIDIA·Open Weight·1M

46.8

prov. avg

Qwen2.5-72B
Alibaba·Open Weight·128K

44.9

tracked

67
o4-mini (high)
OpenAI·Proprietary·200K

44.3

prov. avg

68
DeepSeek V4 Flash (High)
DeepSeek·Open Weight·1M

44.2

prov. avg

69
Claude Haiku 4.5
Anthropic·Proprietary·200K

43.5

prov. avg

70
Qwen3.5-27B
Alibaba·Open Weight·262K

43.3

prov. avg

71
Nemotron 3 Super 100B
NVIDIA·Open Weight·1M

41.9

prov. avg

72
Claude 4 Sonnet
Anthropic·Proprietary·200K

41.5

prov. avg

73
Qwen3.5-35B-A3B
Alibaba·Open Weight·262K

40.2

prov. avg

Mistral Large 3
Mistral·Proprietary·128K

39.8

tracked

Llama 3.1 405B
Meta·Open Weight·128K

39.3

tracked

76
Gemini 3.1 Flash-Lite
Google·Proprietary·1M

38.9

prov. avg

Grok Code Fast 1
xAI·Proprietary·256K

37.5

tracked

78
Qwen3 235B 2507 (Reasoning)
Alibaba·Open Weight·128K

37.4

prov. avg

79
GPT-4o
OpenAI·Proprietary·128K

37.3

prov. avg

Mistral Large 2
Mistral·Proprietary·128K

37.3

tracked

81
Claude 3.5 Sonnet
Anthropic·Proprietary·200K

36.9

prov. avg

82
Gemini 2.5 Flash
Google·Proprietary·1M

36.7

prov. avg

GPT-4o mini
OpenAI·Proprietary·128K

36.1

tracked

84
GPT-4.1 mini
OpenAI·Proprietary·1M

35.7

prov. avg

85
Claude 4.1 Opus
Anthropic·Proprietary·200K

35.4

prov. avg

DeepSeek V3.1 (Reasoning)
DeepSeek·Open Weight·128K

34.8

tracked

Gemini 1.5 Pro
Google·Proprietary·2M

32.9

tracked

Claude 3 Opus
Anthropic·Proprietary·200K

32.5

tracked

DeepSeek-R1
DeepSeek·Open Weight·128K

32.1

tracked

90
GPT-OSS 120B
OpenAI·Open Weight·128K

31.7

prov. avg

91
Kimi K2
Moonshot AI·Proprietary·128K

30.1

prov. avg

Claude 4.1 Opus Thinking
Anthropic·Proprietary·200K

23.9

tracked

Nemotron Ultra 253B
NVIDIA·Open Weight·32K

23.5

tracked

GPT-4 Turbo
OpenAI·Proprietary·128K

22.3

tracked

Llama 3 70B
Meta·Open Weight·128K

21.3

tracked

96
GPT-4.1 nano
OpenAI·Proprietary·1M

20.7

prov. avg

DeepSeek V3.1
DeepSeek·Open Weight·128K

20.6

tracked

98
Grok 3 [Beta]
xAI·Proprietary·128K

20.4

prov. avg

Claude 3 Haiku
Anthropic·Proprietary·200K

19.6

tracked

Nemotron 3 Nano 30B
NVIDIA·Open Weight·32K

19.1

tracked

Z-1
Z·Proprietary·128K

18.5

tracked

Moonshot v1
Moonshot AI·Proprietary·128K

17.8

tracked

GLM-4.5
Z.AI·Proprietary·128K

17.6

tracked

Nemotron-4 15B
NVIDIA·Open Weight·32K

17

tracked

105
GPT-OSS 20B
OpenAI·Open Weight·128K

16.7

prov. avg

106
Llama 4 Scout
Meta·Open Weight·10M

16.4

prov. avg

Mistral 8x7B
Mistral·Open Weight·32K

16.4

tracked

108
Qwen3 235B 2507
Alibaba·Open Weight·128K

15.7

prov. avg

o1-pro
OpenAI·Proprietary·200K

15.2

tracked

Phi-4
Microsoft·Open Weight·16K

15.2

tracked

Gemini 1.0 Pro
Google·Proprietary·32K

14.1

tracked

GLM-4.5-Air
Z.AI·Proprietary·128K

13.8

tracked

Gemma 3 27B
Google·Open Weight·32K

12.2

tracked

114
Llama 4 Maverick
Meta·Open Weight·1M

12.1

prov. avg

DBRX Instruct
Databricks·Open Weight·32K

6.8

tracked

Mixtral 8x22B Instruct v0.1
Mistral·Open Weight·64K

6.8

tracked

Llama 4 Behemoth
Meta·Open Weight·32K

4.7

tracked

Nova Pro
Amazon·Proprietary·128K

4.5

tracked

Mistral 7B v0.3
Mistral·Open Weight·32K

2.6

tracked

Mistral 8x7B v0.2
Mistral·Open Weight·32K

0.9

tracked

These rankings update weekly

Get notified when models move. One email a week with what changed and why.

Free. No spam. Unsubscribe anytime.

Key Takeaways

The top model is Claude Mythos 5 by Anthropic with a provisional score of 100.

The best open-weight model is GLM-5 (Reasoning) at position #11.

120 models are included in this ranking.

Score in Context

What these scores mean

The agentic score blends tool-use, browsing, computer-use, and long-horizon execution benchmarks. It carries 22% of the overall BenchLM score — the largest single category weight.

Known limitations

Agentic benchmarks bundle the model with a harness (scaffold, tools, retries). Your agent stack can move results by more than the gap between adjacent models here.

Best LLMs for AI Agents FAQ

What is the best LLM for AI agents?

Claude Mythos 5 and Claude Fable 5 currently top BenchLM's agentic category, with GPT-5.5 the strongest non-Anthropic option. The leaderboard above recomputes on every data refresh. For agent work specifically, weight Terminal-Bench 2, BrowseComp, and tau2-bench scores over general benchmarks.

What is the best AI agent?

This page ranks the models that power agents rather than packaged agent products. An agent product is a model plus a harness — tools, memory, retries — so the same model can perform very differently across products. Start from the strongest agentic model, then evaluate harnesses on your own tasks.

Do agentic benchmarks predict real-world agent performance?

Directionally yes, precisely no. Benchmarks like OSWorld-Verified and tau2-bench correlate with practical reliability, but they test specific environments with specific scaffolds. A 10-point gap usually matters; a 2-point gap can vanish under your own harness.

What is the cheapest good model for agents?

Agent loops burn tokens fast, so effective cost matters more than list price. Check the price-vs-performance view and the provider pricing hubs for cached-input rates — caching discounts of up to 90% dominate agent economics because loops resend the same context every step.

Last updated: July 13, 2026

The AI models change fast. We track them for you.

A weekly brief for engineers and researchers covering new models, ranking shifts, and pricing changes.

Free. No spam. Unsubscribe anytime.