Terminal-Bench 4.0
We show this table for reference; we do not rank on it.
The current Terminal-Bench release measures difficult computer work after recalibrating task resources, fixing unstable tasks, and removing tasks that no longer separate frontier systems.
Previous release
Terminal-Bench 3.0 remains available as a historical snapshot. Version 4.0 changes the task set and resource limits, so movement between the two tables is not a like-for-like model comparison.
View Terminal-Bench 3.0Tasks completed on Terminal-Bench 4.0 — October 15, 2026 snapshot
We mirror the published tasks completed view for Terminal-Bench 4.0. Claude Opus 5.5 leads the public snapshot at 64.85%, followed by Claude Sonnet 5.5 (61.82%) and GPT-6 Astra (58.18%). We do not weight the raw table directly. A normalized version can contribute through our external agentic consensus gate.
Claude Opus 5.5
Anthropic
Claude Code · max reasoning
Claude Sonnet 5.5
Anthropic
Claude Code · max reasoning
GPT-6 Astra
OpenAI
Codex · max reasoning
35 modelsAgentic8% of Agentic reference weightCurrentUpdated October 15, 2026 snapshot
Tasks completed table (35 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | Claude Opus 5.5AnthropicClaude Code · max reasoning | 64.85% | Not reported | Closed |
| 2 | Claude Sonnet 5.5AnthropicClaude Code · max reasoning | 61.82% | Not reported | Closed |
| 3 | GPT-6 AstraOpenAICodex · max reasoning | 58.18% | Not reported | Closed |
| 4 | GPT-6.1 SolOpenAICodex · max reasoning | 58.18% | Not reported | Closed |
| 5 | Claude Fable 5.1AnthropicClaude Code · max reasoning | 57.88% | Not reported | Closed |
| 6 | GPT-6 AstraOpenAICodex · xhigh reasoning | 57.88% | Not reported | Closed |
| 7 | GPT-6 AstraOpenAICodex · high reasoning | 57.88% | Not reported | Closed |
| 8 | Claude Fable 5.1AnthropicClaude Code · xhigh reasoning | 57.88% | Not reported | Closed |
| 9 | Claude Fable 5.1AnthropicClaude Code · high reasoning | 54.55% | Not reported | Closed |
| 10 | GPT-6 AstraOpenAICodex · medium reasoning | 54.24% | Not reported | Closed |
| 11 | Claude Fable 5.1AnthropicClaude Code · medium reasoning | 53.94% | Not reported | Closed |
| 12 | Claude Opus 5AnthropicClaude Code · xhigh reasoning | 53.94% | Not reported | Closed |
| 13 | Claude Opus 5AnthropicClaude Code · max reasoning | 51.82% | Not reported | Closed |
| 14 | GPT-6 AstraOpenAICodex · low reasoning | 50.61% | Not reported | Closed |
| 15 | Claude Opus 5AnthropicClaude Code · high reasoning | 50.30% | Not reported | Closed |
| 16 | GPT-6 SolOpenAICodex · max reasoning | 49.39% | Not reported | Closed |
| 17 | Claude Opus 5AnthropicClaude Code · medium reasoning | 44.85% | Not reported | Closed |
| 18 | Claude Fable 5AnthropicClaude Code · max reasoning | 44.55% | Not reported | Closed |
| 19 | Claude Fable 5.1AnthropicClaude Code · low reasoning | 43.33% | Not reported | Closed |
| 20 | GLM-5.3Z.AIClaude Code · max reasoning | 41.82% | Not reported | Open |
| 21 | Grok 4.7xAIGrok Build · xhigh reasoning | 37.58% | Not reported | Closed |
| 22 | GPT-5.6 SolOpenAICodex · max reasoning | 37.27% | Not reported | Closed |
| 23 | GLM-5.3-FlashZ.AIClaude Code · none reasoning | 35.76% | Not reported | Open |
| 24 | Claude Opus 5AnthropicClaude Code · low reasoning | 34.85% | Not reported | Closed |
| 25 | Qwen3.8-Max-0902AlibabaClaude Code · max reasoning | 26.97% | Not reported | Unknown |
| 26 | Claude Opus 4.8AnthropicClaude Code · max reasoning | 23.64% | Not reported | Closed |
| 27 | GPT-5.6 TerraOpenAICodex · max reasoning | 21.52% | Not reported | Closed |
| 28 | Grok 4.6xAIGrok Build · high reasoning | 20.30% | Not reported | Closed |
| 29 | Gemini 3.8 FlashGooglemini-SWE-agent · high reasoning | 19.09% | Not reported | Closed |
| 30 | GPT-5.6 LunaOpenAICodex · max reasoning | 17.27% | Not reported | Closed |
| 31 | GPT-6 LunaOpenAICodex · max reasoning | 16.36% | Not reported | Closed |
| 32 | Muse Spark 1.3MetaMuse Code · xhigh reasoning | 14.55% | Not reported | Closed |
| 33 | Grok 4.5xAIGrok Build · high reasoning | 12.42% | Not reported | Closed |
| 34 | Claude Sonnet 5AnthropicClaude Code · max reasoning | 12.42% | Not reported | Closed |
| 35 | Gemini 3.7 FlashGooglemini-SWE-agent · high reasoning | 11.21% | Not reported | Closed |
AA Terminal-Bench 4.0
Stored separately from the benchmark-owned Terminal-Bench 4.0 lane because Artificial Analysis controls the agent harness and configuration. Display-only; not yet admitted as independent-run evidence.
| Source rank | Source model or configuration | Score | Source setup | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | Claude Sonnet 5.5Anthropic | 63.6% | Artificial AnalysisOriginal result | Not reported | Closed |
| 2 | Claude Opus 5.5Anthropic | 59.6% | Artificial AnalysisOriginal result | Not reported | Closed |
| 3 | GPT-6 AstraOpenAI | 59.1% | Artificial AnalysisOriginal result | Not reported | Closed |
| 4 | Gemini 4 ArgonGoogle | 57.1% | Artificial AnalysisOriginal result | Not reported | Closed |
| 5 | Muse Wonder FullMeta | 57.1% | Artificial AnalysisOriginal result | Not reported | Closed |
| 6 | GPT-6.1 SolOpenAI | 56.1% | Artificial AnalysisOriginal result | Not reported | Closed |
| 7 | Claude Fable 5.1Anthropic | 52.0% | Artificial AnalysisOriginal result | Not reported | Closed |
| 8 | Muse WonderMeta | 52.0% | Artificial AnalysisOriginal result | Not reported | Closed |
| 9 | GLM-5.3Z.AI | 41.9% | Artificial AnalysisOriginal result | Not reported | Open |
| 10 | MiMo-V2.6-ProXiaomi | 34.8% | Artificial AnalysisOriginal result | Not reported | Open |
| 11 | Step 5 PreviewStepFun | 33.3% | Artificial AnalysisOriginal result | Not reported | Open |
| 12 | Muse Spark 1.3Meta | 33.3% | Artificial AnalysisOriginal result | Not reported | Closed |
| 13 | Claude Haiku 5.5Anthropic | 32.8% | Artificial AnalysisOriginal result | Not reported | Closed |
| 14 | GLM-5.3-FlashZ.AI | 32.8% | Artificial AnalysisOriginal result | Not reported | Open |
| 15 | DeepSeek V4.1 Flash 1017DeepSeek | 26.8% | Artificial AnalysisOriginal result | Not reported | Open |
| 16 | Mistral Large 4Mistral | 26.8% | Artificial AnalysisOriginal result | Not reported | Pending |
| 17 | Grok 4.7xAI | 25.8% | Artificial AnalysisOriginal result | Not reported | Closed |
| 18 | Gemini 3.8 FlashGoogle | 19.7% | Artificial AnalysisOriginal result | Not reported | Closed |
| 19 | Kimi K3Moonshot AI | 12.6% | Artificial AnalysisOriginal result | Not reported | Pending |
| 20 | GPT-6 LunaOpenAI | 12.6% | Artificial AnalysisOriginal result | Not reported | Closed |
| 21 | Muse Spark OpenMeta | 8.1% | Artificial AnalysisOriginal result | Not reported | Open |
| 22 | Qwen3.8-27BAlibaba | 5.6% | Artificial AnalysisOriginal result | Not reported | Open |
| 23 | MiniMax M3MiniMax | 2.0% | Artificial AnalysisOriginal result | Not reported | Open |
| 24 | InklingThinking Machines Lab | 1.0% | Artificial AnalysisOriginal result | Not reported | Open |
| 25 | Muse Glimmer 30BMeta | 0.5% | Artificial AnalysisOriginal result | Not reported | Open |
Terminal-Bench 4.0 (Vals)
We keep the Vals-hosted Terminal-Bench 4.0 table on a dedicated key so it never overwrites the official Terminal-Bench 4.0 leaderboard rows. Vals runs its own harness and publishes per-domain splits; this mirror is display only and does not enter weighted rankings.
| Source rank | Source model or configuration | Vals Terminal-Bench 4.0 mirror score | Source setup | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | Claude Opus 5.5Anthropic | 65.15% | Vals AIOriginal result | Not reported | Closed |
| 2 | Claude Sonnet 5.5Anthropic | 64.14% | Vals AIOriginal result | Not reported | Closed |
| 3 | GPT-6 AstraOpenAI | 59.60% | max reasoningOriginal result | Not reported | Closed |
| 4 | Claude Fable 5.1Anthropic | 58.08% | Vals AIOriginal result | Not reported | Closed |
| 5 | Muse Wonder FullMeta | 58.08% | max reasoningOriginal result | Not reported | Closed |
| 6 | Gemini 4 ArgonGoogle | 57.58% | high reasoningOriginal result | Not reported | Closed |
| 7 | GPT-6.1 SolOpenAI | 55.05% | max reasoningOriginal result | Not reported | Closed |
| 8 | Claude Opus 5Anthropic | 53.53% | Vals AIOriginal result | Not reported | Closed |
| 9 | Muse WonderMeta | 49.49% | max reasoningOriginal result | Not reported | Closed |
| 10 | GPT-6 SolOpenAI | 44.44% | max reasoningOriginal result | Not reported | Closed |
| 11 | Claude Fable 5Anthropic | 41.41% | Vals AIOriginal result | Not reported | Closed |
| 12 | GLM-5.3Z.AI | 38.89% | max reasoningOriginal result | Not reported | Open |
| 13 | GPT-5.6 SolOpenAI | 37.88% | max reasoningOriginal result | Not reported | Closed |
| 14 | GLM-5.3-FlashZ.AI | 37.37% | max reasoningOriginal result | Not reported | Open |
| 15 | Claude Haiku 5.5Anthropic | 36.36% | Vals AIOriginal result | Not reported | Closed |
| 16 | Qwen3.8 MaxAlibaba | 34.34% | Vals AIOriginal result | Not reported | Open |
| 17 | Step 5 PreviewStepFun | 31.82% | Vals AIOriginal result | Not reported | Open |
| 18 | MiMo-V2.6-ProXiaomi | 31.31% | Vals AIOriginal result | Not reported | Open |
| 19 | Grok 4.7xAI | 28.79% | xhigh reasoningOriginal result | Not reported | Closed |
| 20 | Muse Spark 1.3 MaxMeta | 24.75% | max reasoningOriginal result | Not reported | Unknown |
| 21 | MiMo-V2.6-FlashXiaomi | 24.24% | Vals AIOriginal result | Not reported | Open |
| 22 | Mistral Large 4Mistral | 23.23% | high reasoningOriginal result | Not reported | Pending |
| 23 | Claude Opus 4.8Anthropic | 23.23% | Vals AIOriginal result | Not reported | Closed |
| 24 | GPT-5.6 TerraOpenAI | 22.73% | max reasoningOriginal result | Not reported | Closed |
| 25 | Ember-1Fireworks | 20.71% | Vals AIOriginal result | Not reported | Closed |
| 26 | DeepSeek V4.1 Flash 1017DeepSeek | 19.70% | high reasoningOriginal result | Not reported | Open |
| 27 | Gemini 3.8 FlashGoogle | 19.70% | high reasoningOriginal result | Not reported | Closed |
| 28 | DeepSeek V4 Flash 0731DeepSeek | 18.69% | high reasoningOriginal result | Not reported | Open |
| 29 | Grok 4.6xAI | 17.17% | high reasoningOriginal result | Not reported | Closed |
| 30 | Kimi K3Moonshot AI | 17.17% | max reasoningOriginal result | Not reported | Pending |
| 31 | DeepSeek V4 Pro 0813DeepSeek | 14.14% | max reasoningOriginal result | Not reported | Open |
| 32 | GPT-6 LunaOpenAI | 13.64% | max reasoningOriginal result | Not reported | Closed |
| 33 | Gemini 3.7 FlashGoogle | 12.12% | high reasoningOriginal result | Not reported | Closed |
| 34 | GPT-5.6 LunaOpenAI | 11.62% | max reasoningOriginal result | Not reported | Closed |
| 35 | DeepSeek V4 Pro 0813DeepSeek | 11.11% | max reasoningOriginal result | Not reported | Open |
| 36 | Claude Sonnet 5Anthropic | 10.61% | Vals AIOriginal result | Not reported | Closed |
| 37 | Muse Spark 1.3Meta | 10.61% | xhigh reasoningOriginal result | Not reported | Closed |
| 38 | Grok 4.5xAI | 8.59% | high reasoningOriginal result | Not reported | Closed |
| 39 | Hy4 previewTencent | 8.08% | Vals AIOriginal result | Not reported | Open |
| 40 | Gemini 3.5 FlashGoogle | 6.06% | high reasoningOriginal result | Not reported | Closed |
| 41 | Muse Spark 1.2Meta | 6.06% | xhigh reasoningOriginal result | Not reported | Closed |
| 42 | Gemini 3.1 Pro PreviewGoogle | 2.52% | high reasoningOriginal result | Not reported | Unknown |
| 43 | GPT-5.4 miniOpenAI | 2.52% | xhigh reasoningOriginal result | Not reported | Closed |
| 44 | MiniMax M3MiniMax | 1.01% | Vals AIOriginal result | Not reported | Open |
| 45 | InklingThinking Machines Lab | 0.51% | 0.99 reasoningOriginal result | Not reported | Open |
| 46 | MiMo-V2.5-ProXiaomi | 0.51% | Vals AIOriginal result | Not reported | Closed |
| 47 | Inkling-SmallThinking Machines Lab | 0.00% | 0.99 reasoningOriginal result | Not reported | Open |
| 48 | Mercury 2.5Inception | 0.00% | high reasoningOriginal result | Not reported | Closed |
How to read this leaderboard
Operator receipt: 35 sourced rows are currently displayable on this page; the leading published row is Claude Opus 5.5 at 64.85%.
Honest limit: The table compares complete model-and-agent systems, not isolated base models. Version 4.0 also changes resources and the task set, so score changes from older releases mix system progress with protocol changes.
How we use Terminal-Bench 4.0
We mirror the Terminal-Bench 4.0 table across 66 professional computer-work tasks, with 5 trials per task and an 8-hour agent timeout. Version 4.0 fixes 19 tasks and removes 8 tasks that were saturated, refusal-prone, publicly solved, or still affected by quality and platform issues.
Each row keeps the harness and settings the source published. BenchLM shows the published table for reference; the per-model scores it stores feed the ranking.
Snapshot
The published Terminal-Bench 4.0 snapshot places Claude Opus 5.5 first at 64.85%. The third row is 6.67 points behind. The broader top-10 range is 10.61 points, so the table still separates the published systems.
35 models have been evaluated on Terminal-Bench 4.0. The benchmark falls in the Agentic category. BenchLM shows the published table for reference; the per-model scores it stores feed the ranking. BenchAlign v5.8 gives Terminal-Bench 4.0 8% of the Agentic reference weight, so it moves the Agentic leaderboard and the overall ranking. Reference weights are relative weights in the calibrated model, not fixed shares of a score.
About Terminal-Bench 4.0
Year
2026
Tasks
66 professional computer-work tasks
Format
Task completion rate across 5 trials per task
Difficulty
Frontier autonomous knowledge work
Version 4.0 contains 66 tasks, gives every agent up to eight hours, fixes 19 tasks, and removes eight tasks for saturation, refusals, public solutions, or unresolved quality and platform issues. Those breaking changes required fresh trials, so 4.0 scores are not directly comparable with 3.0 or 2.x. Each row still combines a model with an agent harness.
Freshness and provenance
Version
Terminal-Bench 4.0.0
Refresh cadence
Continuous
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does Terminal-Bench 4.0 measure?
The current Terminal-Bench release measures difficult computer work after recalibrating task resources, fixing unstable tasks, and removing tasks that no longer separate frontier systems.
Which model leads the published Terminal-Bench 4.0 snapshot?
Claude Opus 5.5 currently leads the published Terminal-Bench 4.0 snapshot with 64.85% tasks completed. BenchLM shows the published table for reference; the per-model scores it stores feed the ranking.
How many models are evaluated on Terminal-Bench 4.0?
The October 15, 2026 snapshot snapshot contains 35 AI models.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 5,600+ readers.
One email each week. Unsubscribe anytime.