activeCodingAgents

Terminal-Bench 4.0

Metric: Pass@1 (higher is better)Introduced: 2026

Terminal-Bench 4.0 is the latest Harbor Framework terminal agent benchmark. Keep separate from Terminal-Bench 2.1 and 3.0; DeepSeek reports scores with DeepSeek Harness Minimal mode and a 1M-token context window.

Models ranked

tracked on this benchmark

Score band

65.7 – 28.8

best → lowest tracked

Snapshot trend

+20.90

Oct 17 → Oct 23 · 1 models

Leaderboard

Tracked models ranked by Pass@1 (higher is better).

Compare candidates

#Model variant and provenanceRelative to leaderScore

Kimi K3.1 Pro

Version: Terminal-Bench 4.0; Claude-Code harness; max reasoning effortHarness: Not recordedEvaluator: Not recordedObserved: Oct 23, 2026Confidence: Not recordedSource

Notes: Moonshot self-reported launch table, row 'Terminal-Bench 4.0', column 'Kimi K3.1 Pro (max)'. Settings: reasoning effort max, temperature 1.0, top-p 0.95 (blog footnote). Footnote: 'Kimi K3.1 Pro is evaluated with the Claude-Code harness.'

65.7

Claude Mythos 5.1

Configuration: Claude Mythos 5.1 (max thinking effort)

Version: Terminal-Bench 4.0Harness: Claude Code (--bare mode)Evaluator: AnthropicObserved: Oct 17, 2026Confidence: self-reportedCost/task: $18.00Source

Notes: System card §8.6 (p.171): Claude Code --bare mode, max thinking effort, averaged over 10 trials per task (660 trials), SE ~±1.6-2 pts. Launch-page chart data (https://www.anthropic.com/claude-fable-and-mythos-5-1) lists Mythos 5.1 max = 60.9 at $18.0/task. Table 8.1.A shows shared Fable 5.1/Mythos 5.1 column 56% (61%); the 60.9 figure is the Mythos-specific value. Retrieved 2026-10-17.

60.9

DeepSeek V4.1 Flash 1017

Version: Terminal-Bench 4.0; Claude Code harnessHarness: Not recordedEvaluator: Not recordedObserved: Oct 17, 2026Confidence: Not recordedSource

42.3

Claude Haiku 5.5

Version: Terminal-Bench 4Harness: Not recordedEvaluator: Not recordedObserved: Oct 13, 2026Confidence: Not recordedSource

39.2

MiniMax M3.1 Flash

Configuration: MiniMax M3.1 Flash

Version: Terminal-Bench 4.0Harness: Codex 0.149.1Evaluator: MiniMaxObserved: Oct 19, 2026Confidence: self-reportedSource

Notes: MiniMax M3.1 Flash launch blog (2026-10-19) main benchmark table image m31-flash-en-2 (webp converted to png; values transcribed by eye from table) incl. footnotes. Footnote: 'using the Codex 0.149.1 harness.' Retrieved 2026-10-19.

38.1

MiMo-V2.6-Pro

Version: Terminal Bench 4.0Harness: Not recordedEvaluator: Not recordedObserved: Oct 9, 2026Confidence: Not recordedSource

34.9

DeepSeek V4.1 Flash

Version: Terminal-Bench 4.0 Pass@1; DeepSeek Harness Minimal mode; 1M contextHarness: Not recordedEvaluator: Not recordedObserved: Oct 17, 2026Confidence: Not recordedSource

Notes: Harness conflict: the September V4.1 Flash card names DeepSeek Harness Minimal mode for Terminal-Bench 4.0, while the 1017 card and the 2026-10-17 changelog name Claude Code for the comparison table. Version kept per the base card.

31.2

MiMo-V2.6-Flash

Version: Terminal Bench 4.0Harness: Not recordedEvaluator: Not recordedObserved: Oct 9, 2026Confidence: Not recordedSource

28.8

How to read this benchmark

This benchmark scores models where higher is better. Use scores for directional filtering and shortlisting, not universal quality ranking; then validate pricing, context window, provider availability, and fit for your workload.

Trust this score when

There is a fresh timestamped snapshot (or multiple snapshots) for this benchmark.
The model list covers the same version family you can actually deploy today.
Top candidates overlap with your required routing and feature requirements.

Be cautious when

There is only one benchmark snapshot or the dataset appears stale.
The benchmark metric direction is opposite of your decision objective.
The score difference between options is narrow and likely within implementation variance.

Related benchmarks

AA Coding Agent Index

Artificial Analysis Coding Agent Index

Agents' Last Exam

Code editing across 8 programming languages

Last reviewed: Sep 10, 2026

Resources

Website