Skip to content

Benchmark

TBench

Agentic benchmarks measure how well models carry out tasks that take several steps. Scores are percentages.

Models measured
89
Scores recorded
234
Sources
1
IQ weight
0.93

Score distribution

42.9%
Mean
42.2%
Median
22.9%
Std dev
14.2%
10th percentile
75.1%
90th percentile
0%
Min
0%91%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.93 for chat models, measured over 89 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#Model
201MiMo V2.6 Flashterminal-bench-4.0, mini-swe-agent· Terminal-Bench16.1%
202MiMo V2.6 Flashterminal-bench-4.1, mini-swe-agent· Terminal-Bench16.1%
203Grok 4.5terminal-bench-3.0, Cursor CLI· Terminal-Bench15.7%
204Gemini 2.5 Flashterminal-bench-2.0, Gemini CLI· Terminal-Bench15.4%
205GPT-6 Lunaterminal-bench-4.1, Codex· Terminal-Bench15.2%
206Muse Spark 1.3terminal-bench-4.1, Muse Code· Terminal-Bench14.9%
207Muse Spark 1.3terminal-bench-4.0, Muse Code· Terminal-Bench14.6%
208GPT-5.6 Lunaterminal-bench-3.0, Codex· Terminal-Bench14.3%
209GPT OSS 120Bterminal-bench-2.0, Mini-SWE-Agent· Terminal-Bench14.2%
210Grok Code Fast 1terminal-bench-2.0, Terminus 2· Terminal-Bench14.2%
211Claude Haiku 4.5terminal-bench-2.0, OpenHands· Terminal-Bench13.9%
212Gemini 3.7 Flashterminal-bench-4.1, mini-SWE-agent· Terminal-Bench12.4%
213Grok 4.5terminal-bench-4.0, Grok Build· Terminal-Bench12.4%
214Grok 4.5terminal-bench-4.1, Grok Build· Terminal-Bench12.4%
215Hy4 Flashterminal-bench-4.1, mini-swe-agent· Terminal-Bench11.8%
216GPT-5 nanoterminal-bench-2.0, Codex CLI· Terminal-Bench11.5%
217Gemini 3.7 Flashterminal-bench-4.0, mini-SWE-agent· Terminal-Bench11.2%
218GPT-5 nanoterminal-bench-2.0, OpenHands· Terminal-Bench9.9%
219Qwen3.5 9Bterminal-bench-2.0, little-coder· Terminal-Bench9.2%
220GPT-5 nanoterminal-bench-2.0, Terminus 2· Terminal-Bench7.9%
221MiniMax M3terminal-bench-4.1, mini-swe-agent· Terminal-Bench7.6%
222MiniMax M3terminal-bench-4.0, mini-swe-agent· Terminal-Bench7.3%
223GPT-5 nanoterminal-bench-2.0, Mini-SWE-Agent· Terminal-Bench7%
224DeepSeek V4 Pro 0813terminal-bench-4.0, Claude Code· Terminal-Bench6.7%
225DeepSeek V4 Pro 0813terminal-bench-4.1, Claude Code· Terminal-Bench6.7%
226GLM 5.2terminal-bench-3.0, Claude Code· Terminal-Bench4.6%
227GPT OSS 20bterminal-bench-2.0, Mini-SWE-Agent· Terminal-Bench3.4%
228GPT OSS 20bterminal-bench-2.0, Terminus 2· Terminal-Bench3.1%
229Gemini 3.1 Pro Previewterminal-bench-4.0, Antigravity CLI· Terminal-Bench2.7%
230Gemini 3.1 Pro Previewterminal-bench-4.1, Antigravity CLI· Terminal-Bench2.7%
231GPT-5.4 miniterminal-bench-4.0, Codex· Terminal-Bench1.8%
232GPT-5.4 miniterminal-bench-4.1, Codex· Terminal-Bench1.8%
233Inkling FP4terminal-bench-4.0, mini-swe-agent· Terminal-Bench0%
234Inkling FP4terminal-bench-4.1, mini-swe-agent· Terminal-Bench0%

Showing 201–234 of 234 results.

Tokens the model generated per task, as measured by the source. Comparable within this benchmark only.

Data credits: Terminal-Bench (Laude Institute), Apache-2.0 (licence)