Terminal-Bench 2.1

Version 2.1 of Terminal-Bench.

Category: Agentic coding
Unit: Percent (%); Higher is better
Comparison groups: 7total; All model rankings (records) · Ranked results: 9

All model rankings (records)

These rankings are recalculated by comparison group including legacy models and models unavailable in Claude Code / Codex. They differ from the rankings above because the set of models differs.

Show rankings including legacy and unavailable models (entries: 9)

official results (Gemini 3.8 Flash model card)

Rank	Model	effort	Score	Source cost (metric)	measured/reported by	source
1	Gemini 3.8 Flash	not stated	89.4%	—	Google	Gemini 3.8 Flash model cardRetrieved 2026-11-12 · Vendor-reported value. Self-reported by Google (model card, as of 2026-09). thinking setting not stated Source excerpt Terminal-bench 2.1 \| 89.4%

official results (Kimi K3 (Hugging Face model card, moonshotai))

Rank	Model	effort	Score	Source cost (metric)	measured/reported by	source
1	Kimi K3	Max	88.3%	—	Moonshot AI	Kimi K3 (Hugging Face model card, moonshotai)Retrieved 2026-11-13 · Vendor-reported value (self-reported by Moonshot AI). Table note: All Kimi K3 results are obtained with reasoning effort set to 'max' and temperature = 1.0. Kimi Code harness. Source excerpt Terminal-Bench 2.1 \| 88.3Location HF model card "3. Evaluation Results" table, Terminal-Bench 2.1 row, Kimi K3 (max) column (leftmost numeric column). Supporting footnote: Coding / Terminal-Bench 2.1

official results (GLM-5.3 (Hugging Face model card, zai-org))

Rank	Model	effort	Score	Source cost (metric)	measured/reported by	source
1	GLM-5.3	not stated	88.2%	—	Z.ai	GLM-5.3 (Hugging Face model card, zai-org)Retrieved 2026-11-13 · Vendor-reported value (self-reported by Z.ai). Card Note: the default for reasoning_effort is max; use the default max to reproduce the benchmarks. Claude Code 2.1.207; 6h timeout. Effort not stated in this footnote. Source excerpt Terminal Bench 2.1 \| 88.2Location HF model card "Benchmark" table, Terminal Bench 2.1 row, GLM-5.3 column (leftmost numeric column). Basis: Footnotes, Terminal-Bench 2.1

official results (Gemini 3.7 Flash model card)

Rank	Model	effort	Score	Source cost (metric)	measured/reported by	source
1	Gemini 3.7 Flash	not stated	85.8%	—	Google	Gemini 3.7 Flash model cardRetrieved 2026-11-13 · Vendor-reported value. Table in the Google model card. methodology PDF (deepmind.com/models/evals-methodology/gemini-3-7-flash): All of the results are run with the Gemini API ... with default sampling settings unless indicated otherwise. Gemini computed in-house; default harness Terminus 2. thinking level not stated Source excerpt Terminal-bench 2.1 Agentic terminal coding \| 85.8%Location Gemini 3.7 Flash model card Evaluation table, Terminal-bench 2.1 row, Gemini 3.7 Flash column. In-house computation basis: methodology PDF, Coding / Terminal-Bench 2.1
2	Gemini 3.6 Flash	not stated	78.0%	—	Google	Gemini 3.7 Flash model cardRetrieved 2026-11-13 · Vendor-reported value. Table in the Google model card. methodology PDF (deepmind.com/models/evals-methodology/gemini-3-7-flash): All of the results are run with the Gemini API ... with default sampling settings unless indicated otherwise. Gemini computed in-house; default harness Terminus 2. thinking level not stated Source excerpt Terminal-bench 2.1 Agentic terminal coding \| 78.0%Location Gemini 3.7 Flash model card Evaluation table, Terminal-bench 2.1 row, Gemini 3.6 Flash column. In-house computation basis: methodology PDF, Coding / Terminal-Bench 2.1

official results (Gemini 3.5 Flash model card)

Rank	Model	effort	Score	Source cost (metric)	measured/reported by	source
1	Gemini 3.5 Flash	not stated	76.2%	—	Google	Gemini 3.5 Flash model cardRetrieved 2026-11-13 · Vendor-reported value. Table in the Google model card. methodology PDF (deepmind.google/models/evals-methodology/gemini-3-5-flash): All of the results are all run with the Gemini API ... with default sampling settings unless indicated otherwise. 3.5 Flash and 3 Flash computed in-house with the Terminus 2 harness (3.1 Pro is not used because its value is from the leaderboard). thinking level not stated Source excerpt Coding \| Terminal-bench 2.1 Agentic terminal coding \| Terminus-2 harness \| 76.2%Location Gemini 3.5 Flash model card Evaluation table, Terminal-bench 2.1 row, Gemini 3.5 Flash column. In-house computation basis: methodology PDF, Coding / Terminal-Bench 2.1
2	Gemini 3 Flash Preview	not stated	58.0%	—	Google	Gemini 3.5 Flash model cardRetrieved 2026-11-13 · Vendor-reported value. Table in the Google model card. methodology PDF (deepmind.google/models/evals-methodology/gemini-3-5-flash): All of the results are all run with the Gemini API ... with default sampling settings unless indicated otherwise. 3.5 Flash and 3 Flash computed in-house with the Terminus 2 harness (3.1 Pro is not used because its value is from the leaderboard). Column heading is Gemini 3 Flash (gemini-3-flash-preview is the only match in the Gemini API model list). thinking level not stated Source excerpt Coding \| Terminal-bench 2.1 Agentic terminal coding \| Terminus-2 harness \| 58.0%Location Gemini 3.5 Flash model card Evaluation table, Terminal-bench 2.1 row, Gemini 3 Flash column. In-house computation basis: methodology PDF, Coding / Terminal-Bench 2.1

official results (Qwen3.8-2.4T-A95B (Hugging Face model card, Qwen))

Rank	Model	effort	Score	Source cost (metric)	measured/reported by	source
1	Qwen3.8-Max	not stated	86.6%	—	Alibaba (Qwen)	Qwen3.8-2.4T-A95B (Hugging Face model card, Qwen)Retrieved 2026-11-13 · Vendor-reported value (self-reported by Alibaba's Qwen team). Table column is Qwen3.8-Max (card: Qwen3.8-Max is the official version based on Qwen3.8-2.4T-A95B). Effort not stated (the default for reasoning_effort is xhigh). Claude Code (avg@10); 5 hour timeout; max_tokens=131,072. Source excerpt Terminal Bench 2.1 \| 86.6Location HF model card "Benchmark Results" table, Terminal Bench 2.1 row, Qwen3.8-Max column (rightmost). Basis: note below the table 2

official results (DeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai))

Rank	Model	effort	Score	Source cost (metric)	measured/reported by	source
1	DeepSeek-V4.1-Flash	not stated	90.6%	—	DeepSeek	DeepSeek-V4.1-Flash (Hugging Face model card, deepseek-ai)Retrieved 2026-11-13 · Vendor-reported value (self-reported by DeepSeek). Card "Instruct Model": All instruct results below use the maximum effort setting (reasoning_effort=100), temperature=1.0, top_p=0.95. The correspondence of reasoning_effort=100 to the API effort names (low/high/max) has no official statement, so effort is null. Evaluated in Minimal mode of DeepSeek Harness with 1M context. Source excerpt Terminal-Bench 2.1 (Pass@1) \| 90.6Location HF model card "Comparison with frontier models (Max reasoning effort)" table, Terminal-Bench 2.1 (Pass@1) row, DS-V4.1-Flash column (rightmost)