Benchmark
Terminal-Bench 2.1
Agentic benchmarks measure how well models carry out tasks that take several steps. Scores are percentages. Higher is better.
- Models and systems measured
- 64
- Scores recorded
- 159
- Sources
- 3
- IQ weight
- 2.76
Score distribution
- 75.3%
- Mean
- 82.7%
- Median
- 17.2%
- Std dev
- 56.4%
- 10th percentile
- 88.8%
- 90th percentile
- 2.9%
- Min
This benchmark has a weight of 2.76 for chat models, measured over 53 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.
Leaderboard
| # | Model or system | Score |
|---|---|---|
| 1 | GPT-5.6 SolUltra effort· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 91.9% |
| 2 | DeepSeek V4.1 Flash· Hugging Face Official Benchmarks data· Reported by the model's developer. The source relaying it did not verify it. claimValues from 2 sources
| 90.6% |
| 3 | DeepSeek V4.1 FlashDeepSeek V4.1 launch chart · exact public value · setting not further specified· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 90.6% |
| 4 | DeepSeek V4.1 FlashDSH Minimal · max effort · temp 1.0 · top_p .95 · 1M · max_steps 500 · N=3 · no network· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 90.6% |
| 5 | DeepSeek V4.1 Flashmini-SWE · max effort · temp 1.0 · top_p .95 · 1M · max_steps 500 · N=3 · no network· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 90.3% |
| 6 | Claude Opus 5Atria Dawn Preview official card / report Table 1 · published comparison run · model snapshots and reasoning tiers not specified unless stated · four-run mean · 12× timeout multiplier capped at 12 hours · 500 turns · Claude Code v2.1.224 except GPT models use Terminus 2 · Opus 4.8 fallback on a subset of tasks· Reported by Shanghai AI Laboratory· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 90.2% |
| 7 | MiMo V2.6 Pro· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 89.9% |
| 8 | MiMo V2.6 Pro RL· Hugging Face Official Benchmarks data· Reported by the model's developer. The source relaying it did not verify it. claim | 89.9% |
| 9 | Gemini 3.8 Flash· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 89.4% |
| 10 | Gemini 3.8 FlashGoogle Model Card· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 89.4% |
| 11 | Qwen3.8 MaxAtria Dawn Preview official card / report Table 1 · published comparison run · model snapshots and reasoning tiers not specified unless stated · four-run mean · 12× timeout multiplier capped at 12 hours · 500 turns · Claude Code v2.1.224 except GPT models use Terminus 2· Reported by Shanghai AI Laboratory· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 89.3% |
| 12 | Claude Opus 5DeepSeek V4.1 launch chart · exact public value · setting not further specified· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 89.1% |
| 13 | Claude Opus 5Google Model Card· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 89.1% |
| 14 | GPT-5.6 Sol· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 88.8% |
| 15 | GPT-5.6 SolDeepSeek V4.1 launch chart · exact public value · setting not further specified· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 88.8% |
| 16 | GPT-5.6 SolGoogle Model Card· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 88.8% |
| 17 | GPT-5.6 Solmax · 89 tasks · native coding harness · isolated cloud sandbox · pass@1· Reported by Meta· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 88.8% |
| 18 | GPT-5.6 SolOpenAI GPT-5.6 launch table· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 88.8% |
| 19 | GPT-5.6 SolQwen表:Qwen 采用 Claude Code avg@10、5h、max_tokens 131072;友商取公开最高分· Reported by Qwen· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 88.8% |
| 20 | GPT-5.6 Sol厂商公开表共同值· Reported by Z.ai, DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 88.8% |
| 21 | Muse Spark 1.3· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 88.8% |
| 22 | Muse Spark 1.3max · 89 tasks · native coding harness · isolated cloud sandbox · pass@1· Reported by Meta· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 88.8% |
| 23 | LongCat 2.5· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 88.4% |
| 24 | Kimi K3setting not named by the source· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 88.3% |
| 25 | Kimi K3· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 88.3% |
| 26 | Kimi K3best score across harnesses · max effort· Reported by Moonshot AI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 88.3% |
| 27 | Kimi K3DeepSeek V4.1 launch chart · exact public value · setting not further specified· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 88.3% |
| 28 | GLM 5.3setting not named by the source· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 88.2% |
| 29 | GLM 5.3· Hugging Face Official Benchmarks data· Reported by the model's developer. The source relaying it did not verify it. claimValues from 2 sources
| 88.2% |
| 30 | GLM 5.3DeepSeek V4.1 launch chart · exact public value · setting not further specified· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 88.2% |
| 31 | DeepSeek V4.1 FlashClaude Code · max effort · temp 1.0 · top_p .95 · 1M · max_steps 500 · N=3 · no network· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 88% |
| 32 | DeepSeek V4 Pro 0813setting not named by the source· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 87.9% |
| 33 | DeepSeek V4 Pro 0813· Hugging Face Official Benchmarks data· Reported by the model's developer. The source relaying it did not verify it. claimValues from 2 sources
| 87.9% |
| 34 | DeepSeek V4 Pro 0813DeepSeek V4.1 launch chart · exact public value · setting not further specified· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 87.9% |
| 35 | MiMo V2.6 Flash· Hugging Face Official Benchmarks data· Reported by the model's developer. The source relaying it did not verify it. claimValues from 2 sources
| 87.6% |
| 36 | GPT-5.6 Terra· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 87.4% |
| 37 | GPT-5.6 TerraOpenAI GPT-5.6 launch table· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 87.4% |
| 38 | Claude Opus 5Claude Code · max· Reported by Meta· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 86.7% |
| 39 | Claude Opus 5max · 89 tasks · native coding harness · isolated cloud sandbox · pass@1· Reported by Meta· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 86.7% |
| 40 | Qwen3.8 Maxsetting not named by the source· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 86.6% |
| 41 | Qwen3.8 Max· Hugging Face Official Benchmarks data· Reported by the model's developer. The source relaying it did not verify it. claimValues from 2 sources
| 86.6% |
| 42 | Qwen3.8 MaxQwen表:Qwen 采用 Claude Code avg@10、5h、max_tokens 131072;友商取公开最高分· Reported by Qwen· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 86.6% |
| 43 | DeepSeek V4.1 FlashPi · max effort · temp 1.0 · top_p .95 · 1M · max_steps 500 · N=3 · no network· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 86.1% |
| 44 | DeepSeek V4.1 FlashDSH PTC · max effort · temp 1.0 · top_p .95 · 1M · max_steps 500 · N=3 · no network· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 85.8% |
| 45 | DeepSeek V4.1 FlashDSH Standard · max effort · temp 1.0 · top_p .95 · 1M · max_steps 500 · N=3 · no network· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 85.8% |
| 46 | Gemini 3.7 Flash· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 85.8% |
| 47 | Gemini 3.7 FlashTerminus 2· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 85.8% |
| 48 | GPT 5.5OpenAI GPT-5.6 launch table· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 85.6% |
| 49 | GLM 5.3Atria Dawn Preview official card / report Table 1 · published comparison run · model snapshots and reasoning tiers not specified unless stated · four-run mean · 12× timeout multiplier capped at 12 hours · 500 turns · Claude Code v2.1.224 except GPT models use Terminus 2· Reported by Shanghai AI Laboratory· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 85.4% |
| 50 | Hy4 preview· Hugging Face Official Benchmarks data· Reported by the model's developer. The source relaying it did not verify it. claimValues from 2 sources
| 85.4% |
Showing 1–50 of 158 results.
Not ranked
The source marks these results as descriptive, not capability scores, so we list them by name instead of ranking them.
| Model or system | Score |
|---|---|
| Muse Spark (version unspecified)Meta Model API · highest effort · version unspecified · 89 tasks · bash-only · pass@1 · five-attempt mean· Reported by Meta, Meta Superintelligence Labs· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 67.3% |
Data credits: Data by LLM Stats · Hugging Face official benchmark leaderboards; licence varies by benchmark · BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links.