Benchmarks — LLM Agent Capability Dashboard

Agent-Tier Benchmark Rankings

Terminal-Bench 2.x

shell agent loops

Claude Mythos 5 (v) restricted88

Claude Fable 5 ★ restored84.3

GPT-5.5 84

GLM-5.2 81

Claude Sonnet 5 ★ NEW80.4

Claude Opus 4.8 ~ (v: 85.0)79

Gemini 3.5 Flash 76.2

DeepSeek V4 Pro 67.9

MiniMax M3 (v)66

Grok 4.3 63.8

Step 3.7 Flash 59.5

DS V4 Flash 56.9

Nemotron 3 Ultra 56.4

Hy3 54.4

Gemini 3.1 Pro ✗54.2

Step 3.5 Flash (v)51

Mistral Large 3 ✗ worst agent12

Missing scores:Kimi K2.7 Code, Grok Build 0.1, MiMo v2.5 Pro, Ring 2.6 1T, ERNIE 5.1, Llama 4 Maverick, Qwen3-Coder-Next, Claude Haiku 4.5, DS V4 Pro Max, Agnes 2.0 Flash, Fugu Ultra, Cohere North Mini Code

SWE-bench Pro

single-pass code review

Claude Fable 5 (v) classifier reroutes80

Claude Mythos 5 (v) restricted77.8

DeepSeek V4 Pro (a)76.2

Fugu Ultra (v)73.7

Claude Opus 4.8 (v)69.2

Claude Sonnet 5 NEW63.2

GLM-5.2 (v)62.1

Qwen3.7 Max 60.6

MiniMax M3 ~ (v)59

GPT-5.5 (v)58.6

Step 3.7 Flash 56.3

Gemini 3.5 Flash (v)55.1

DS V4 Flash 52.6

Gemini 3.1 Pro ✗ (s)46.1

Claude Haiku 4.5 (s)39.5

Missing scores:Kimi K2.7 Code (has SWE-V 62.0 not Pro), Hy3 (has SWE-V 74.0 not Pro), Llama 4 Maverick (has SWE-V 70.4 not Pro), Mistral Large 3 (has SWE-V 76.2 not Pro), Nemotron 3 Ultra (has SWE-V 71.9 not Pro), Step 3.5 Flash (has SWE-V 74.4 not Pro), DS V4 Pro Max (42 benchmarks, no Pro breakdown)

MCP-Atlas

multi-tool use

Gemini 3.5 Flash ★★83.6

Claude Fable 5 ★83.3

Claude Opus 4.8 77.8

GLM-5.2 76.8

Qwen3.7 Max 76.4

Kimi K2.7 Code 76

GPT-5.5 75.3

DeepSeek V4 Pro 73.6

Mistral Large 3 70.4

DS V4 Flash 69

Missing scores:Claude Sonnet 5, Claude Mythos 5, Claude Haiku 4.5, Gemini 3.1 Pro, Grok 4.3, Grok Build 0.1, Nemotron 3 Ultra, Step 3.7 Flash, Step 3.5 Flash, MiniMax M3, Llama 4 Maverick, Qwen3-Coder-Next, Hy3, Ring 2.6 1T, MiMo v2.5 Pro, ERNIE 5.1, Agnes 2.0 Flash, DS V4 Pro Max

SWE-Marathon & Long-Horizon Evidence

sustained multi-hour agent quality

Claude Fable 5 DeepSWE #1 ★70%

Claude Opus 4.8 Marathon ★26.0

Qwen3.7 Max 35h proven ★1,158 calls

GLM-5.2 Marathon13.0

GPT-5.5 Marathon12.0

DeepSeek V4 Pro DeepSWE ✗8%

Missing: Gemini 3.5 Flash, Grok 4.3, Step 3.7, Hy3, Sonnet 5, MiniMax M3 — no Marathon/DeepSWE scores yet. Sonnet 5 TB 80.4% suggests strong long-horizon potential.

Cost-Performance Map

Terminal-Bench score vs $/M output (log scale)

●80+ excellent●65-79 strong●50-64 moderate●25-49 weak●<25 poor●no score

dashed line = optimal frontier · point color = score band (same as bar charts)