AI benchmarks and leaderboards are different things
Learn the difference between a benchmark, an evaluation run and a leaderboard, and use each to answer the right question.
Model reports
Terminal-Bench
Environment image, timeout and permitted commands change results independently of the model. Version and harness both matter: 2.0 and 2.1 are different instruments, and cards report Terminus, Claude Code and Codex harness numbers for the same version. Frontier-Bench is this series' next version rather than a separate benchmark: frontierbench.ai now bills it as "Terminal-Bench 3.0 (formerly Frontier-Bench)", built by the makers of Harbor and Terminal-Bench, so its mentions are counted here. Merging it did not raise this count: all three cards that named Frontier-Bench (Claude Opus 5, Claude Fable 5 and Mythos 5, Kimi K3) also report Terminal-Bench in the same document, and the counting unit is the document, so each still adds one. Being a new version, its task set is not comparable to a 2.x number. Not to be confused with Cognition's FrontierCode, a separate instrument recorded separately despite the shared prefix.
Evaluation context and evidence
- Source
- Model reports
- Publisher
- Not recorded
- Benchmark release
- 2025-05-19
- Collected
- Not recorded
- Modality
- Not recorded
These are source-reported observations, not independent reruns by this site. Check the task version, evaluation protocol, unit and reporting date before comparing models. Results from different setups are not a universal ranking.
Original sources
Reported results 15
Shown in source observation order, not ranked. Dates retain the precision provided by the source.
| Model | Value | Unit | Protocol | Reported date and basis | Source |
|---|---|---|---|---|---|
| Claude Opus 4 | 43.2 | percent | no extended thinking | 2025-05-22 · document_publication | Claude Opus 4 and Sonnet 4 |
| DeepSeek-V4-Pro | 59.1 | percent | non-think | 2026-04-22 · document_publication | DeepSeek-V4-Pro |
| DeepSeek-V4-Pro | 63.3 | percent | think high | 2026-04-22 · document_publication | DeepSeek-V4-Pro |
| DeepSeek-V4-Pro | 67.9 | percent | think max | 2026-04-22 · document_publication | DeepSeek-V4-Pro |
| DeepSeek-V4-Pro-0813 | 87.9 | percent | max reasoning effort, temp 1.0, top_p 0.95 | 2026-08-13 · document_publication | DeepSeek-V4-Pro-0813 |
| DeepSeek-V4-Flash | 56.9 | percent | think max | 2026-06-25 · document_publication | DeepSeek-V4 |
| Gemini 3.1 Pro | 68.5 | percent | Thinking (High), Terminus-2 harness | 2026-02-19 · document_publication | Gemini 3.1 Pro |
| Kimi K3 | 88.3 | percent | reasoning=max | 2026-06-13 · document_publication | Kimi K3 |
| Qwen3.5-397B-A17B | 52.5 | percent | thinking (temp 0.6, top_p 0.95, top_k 20) | 2026-02-16 · document_publication | Qwen3.5-397B-A17B |
| Hy4 preview | 85.4 | percent | Terminal-Bench 2.1 | 2026-08-28 · document_publication | Hy4 preview |
| Grok 4.5 | 83.3 | percent | as reported | 2026-07-08 · document_publication | Grok 4.5 |
| GLM-5.1 | 63.5 | percent | Terminal-Bench 2.0, Terminus-2 harness, as reported | 2026-04-03 · document_publication | GLM-5.1 |
| GLM-5.2 | 81.0 | percent | Terminal-Bench 2.1, Terminus-2, parser json, timeout 4h, temp 1.0, top_p 1.0, max_new_tokens 48k, max_episodes 500, 256K context | 2026-06-16 · document_publication | GLM-5.2 |
| GLM-5.3-Flash | 84.3 | percent | Terminal-Bench 2.1, Claude Code 2.1.207, temp 1.0, top_p 1, max_new_tokens 65536, 6h timeout | 2026-08-27 · document_publication | GLM-5.3-Flash |
| GLM-5 | 56.2 | percent | Terminal-Bench 2.0, Terminus-2 harness, timeout 2h, temp 0.7, top_p 1.0, 128K context | 2026-02-11 · document_publication | GLM-5 |