Report
AI Agent Harness Benchmark, November 2026
Public SWE-bench and Terminal-Bench scores for agent harnesses and coding agents, each with the date on the official source. Read November 5, 2026.
Key takeaways
- This page copies public scores for coding agents and the harnesses around them.
- The tables use the official SWE-bench leaderboard and the official Terminal-Bench 4.0 leaderboard. Each row keeps the date on that source.
- Nothing here is estimated, and a score from another website is not filled in when the official board does not list it.
This page copies public scores for coding agents and the harnesses around them. A harness is the loop, tools, and limits around a model. The tables use the official SWE-bench leaderboard and the official Terminal-Bench 4.0 leaderboard. Each row keeps the date on that source. Nothing here is estimated, and a score from another website is not filled in when the official board does not list it.
What these numbers are
SWE-bench Verified is 500 real GitHub issues. The score is the percent resolved: the share of issues whose tests passed. The bash-only rows all use mini-SWE-agent, so the harness stays the same while the model changes. The SWE-bench site says release 2.x calls tools, while release 1.x reads the action out of the model's text, and that those two generations are not necessarily comparable. The version column is the release named on the row.
Terminal-Bench 4.0 is the current board on tbench.ai. The agent column changes by row, so those scores include Claude Code, Codex, Grok Build, mini-SWE-agent, or Muse Code, not one shared harness. Accuracy and the 95% half-width are the fields on the official leaderboard. The cost cell is copied as printed, including a partial-trial note where the board includes one.
The newest result date on the SWE-bench Verified file, as read on November 5, 2026, is 2026-02-26. The Terminal-Bench 4.0 board record says it was updated 2026-10-23. This November report is a reading of those boards. It is not a claim that either project published a new run in November.
SWE-bench Verified, bash only
47 rows from the Verified list where the harness is mini-SWE-agent. "Team checked" means the SWE-bench page marks the run as performed or checked by the SWE-bench team. A "No" means the row is on the official board without that mark. Source: swebench.com, read November 5, 2026.
| Claude 4.5 Opus (high) | mini-SWE-agent | Claude 4.5 Opus | 76.8% | 2026-02-17 | 2.0.0 | high | $376.95 | No | No |
| MiniMax M2.5 (high) | mini-SWE-agent | MiniMax M2.5 | 75.8% | 2026-02-17 | 2.0.0 | high | $36.64 | No | Yes |
| Gemini 3 Flash (high) | mini-SWE-agent | Gemini 3 Flash | 75.8% | 2026-02-17 | 2.0.0 | high | $177.98 | No | No |
| Claude 4.6 Opus | mini-SWE-agent | Claude 4.6 Opus | 75.6% | 2026-02-17 | 2.0.0 | Not stated | $275.76 | No | No |
| Claude 4.5 Opus (20251101) (medium) | mini-SWE-agent | Claude 4.5 Opus | 74.4% | 2025-11-24 | 1.16.0 | medium | $360.62 | Yes | No |
| Gemini 3 Pro Preview (2025-11-18) | mini-SWE-agent | Gemini 3 Pro Preview | 74.2% | 2025-11-18 | 1.15.0 | Not stated | $229.98 | Yes | No |
| GPT 5.2 Codex | mini-SWE-agent | GPT 5.2 Codex | 72.8% | 2026-02-19 | 2.0.0 | Not stated | $224.71 | No | No |
| GPT 5.2 (high) | mini-SWE-agent | GPT 5.2 | 72.8% | 2026-02-17 | 2.0.0 | high | $236.78 | No | No |
| GLM 5 (high) | mini-SWE-agent | GLM 5 | 72.8% | 2026-02-17 | 2.0.0 | high | $267.19 | No | Yes |
| GPT 5.2 (2025-12-11) (high) | mini-SWE-agent | GPT 5.2 | 71.8% | 2025-12-11 | 1.17.2 | high | $260.13 | No | No |
| Claude 4.5 Sonnet (high) | mini-SWE-agent | Claude 4.5 Sonnet | 71.4% | 2026-02-17 | 2.0.0 | high | $328.95 | No | No |
| Kimi K2.5 (high) | mini-SWE-agent | Kimi K2.5 | 70.8% | 2026-02-17 | 2.0.0 | high | $73.28 | No | Yes |
| Claude 4.5 Sonnet (20250929) | mini-SWE-agent | Claude 4.5 Sonnet | 70.6% | 2025-09-29 | 1.13.3 | Not stated | $279.17 | Yes | No |
| DeepSeek V3.2 (high) | mini-SWE-agent | DeepSeek V3.2 | 70% | 2026-02-17 | 2.0.0 | high | $223.92 | No | Yes |
| Gemini 3 Pro | mini-SWE-agent | Gemini 3 Pro | 69.6% | 2026-02-26 | 2.0.0 | high | $480.01 | No | No |
| GPT 5.2 (2025-12-11) | mini-SWE-agent | GPT 5.2 | 69% | 2025-12-11 | 1.17.2 | Not stated | $134.83 | No | No |
| Claude 4 Opus (20250514) | mini-SWE-agent | Claude 4 Opus | 67.6% | 2025-08-02 | 1.0.0 | Not stated | $565.64 | Yes | No |
| Claude 4.5 Haiku (high) | mini-SWE-agent | Claude 4.5 Haiku | 66.6% | 2026-02-17 | 2.0.0 | high | $165.46 | No | No |
| GPT 5.1 Codex (medium) | mini-SWE-agent | GPT 5.1 Codex | 66% | 2025-11-24 | 1.16.0 | medium | $294.44 | No | No |
| GPT 5.1 (2025-11-13) (medium) | mini-SWE-agent | GPT 5.1 | 66% | 2025-11-20 | 1.15.0 | medium | $153.08 | No | No |
| GPT 5 (2025-08-07) (medium) | mini-SWE-agent | GPT 5 | 65% | 2025-08-07 | 1.7.0 | medium | $140.19 | Yes | No |
| Claude 4 Sonnet (20250514) | mini-SWE-agent | Claude 4 Sonnet | 64.93% | 2025-07-26 | 1.0.0 | Not stated | $185.73 | Yes | No |
| Kimi K2 Thinking | mini-SWE-agent | Kimi K2 Thinking | 63.4% | 2025-12-10 | 1.17.2 | Not stated | $219.15 | Yes | Yes |
| MiniMax M2 | mini-SWE-agent | MiniMax M2 | 61% | 2025-11-24 | 1.17.0 | Not stated | $214.13 | Yes | Yes |
| DeepSeek V3.2 Reasoner | mini-SWE-agent | DeepSeek V3.2 Reasoner | 60% | 2025-12-01 | 1.17.1 | Not stated | $14.04 | Yes | Yes |
| GPT 5 mini (2025-08-07) (medium) | mini-SWE-agent | GPT 5 mini | 59.8% | 2025-08-07 | 1.7.0 | medium | $17.74 | Yes | No |
| o3 (2025-04-16) | mini-SWE-agent | o3 | 58.4% | 2025-07-26 | 1.0.0 | Not stated | $166.83 | Yes | No |
| Devstral Small (2512) | mini-SWE-agent | Devstral Small (2512) | 56.4% | 2025-12-09 | 1.17.2 | Not stated | Not listed | Yes | Yes |
| GPT 5 mini | mini-SWE-agent | GPT 5 mini | 56.2% | 2026-02-17 | 2.0.0 | Not stated | $23.60 | No | No |
| GLM 4.6 (T=1) | mini-SWE-agent | GLM 4.6 | 55.4% | 2025-12-01 | 1.17.1 | Not stated | $48.31 | Yes | Yes |
| Qwen3-Coder 480B/A35B Instruct | mini-SWE-agent | Qwen3-Coder 480B/A35B Instruct | 55.4% | 2025-08-02 | 1.0.0 | Not stated | $123.96 | Yes | Yes |
| GLM 4.5 (2025-08-22) | mini-SWE-agent | GLM 4.5 | 54.2% | 2025-08-22 | 1.9.1 | Not stated | $148.60 | Yes | Yes |
| Devstral (2512) | mini-SWE-agent | Devstral (2512) | 53.8% | 2025-12-09 | 1.17.2 | Not stated | Not listed | Yes | Yes |
| Gemini 2.5 Pro (2025-05-06) | mini-SWE-agent | Gemini 2.5 Pro | 53.6% | 2025-07-26 | 1.0.0 | Not stated | $144.19 | Yes | No |
| Claude 3.7 Sonnet (20250219) | mini-SWE-agent | Claude 3.7 Sonnet | 52.8% | 2025-07-20 | 0.0.0 | Not stated | $177.31 | Yes | No |
| o4-mini (2025-04-16) | mini-SWE-agent | o4-mini | 45% | 2025-07-26 | 1.0.0 | Not stated | $104.99 | Yes | No |
| Kimi K2 Instruct | mini-SWE-agent | Kimi K2 Instruct | 43.8% | 2025-08-07 | 1.7.0 | Not stated | $265.80 | Yes | Yes |
| GPT 4.1 (2025-04-14) | mini-SWE-agent | GPT 4.1 | 39.58% | 2025-07-26 | 1.0.0 | Not stated | $73.16 | Yes | No |
| GPT 5 nano (2025-08-07) (medium) | mini-SWE-agent | GPT 5 nano | 34.8% | 2025-08-07 | 1.7.0 | medium | $19.04 | Yes | No |
| Gemini 2.5 Flash (2025-04-17) | mini-SWE-agent | Gemini 2.5 Flash | 28.73% | 2025-07-26 | 1.0.0 | Not stated | $66.94 | Yes | No |
| gpt-oss-120b | mini-SWE-agent | gpt-oss-120b | 26% | 2025-08-07 | 1.7.0 | Not stated | $28.56 | Yes | Yes |
| GPT 4.1 mini (2025-04-14) | mini-SWE-agent | GPT 4.1 mini | 23.94% | 2025-07-20 | 0.0.0 | Not stated | $217.58 | Yes | No |
| GPT 4o (2024-11-20) | mini-SWE-agent | GPT 4o | 21.62% | 2025-07-20 | 0.0.0 | Not stated | $765.57 | Yes | No |
| Llama 4 Maverick Instruct | mini-SWE-agent | Llama 4 Maverick Instruct | 21.04% | 2025-07-20 | 0.0.0 | Not stated | $156.85 | Yes | No |
| Gemini 2.0 Flash | mini-SWE-agent | Gemini 2.0 Flash | 13.52% | 2025-07-26 | 0.0.0 | Not stated | $0.00 | Yes | No |
| Llama 4 Scout Instruct | mini-SWE-agent | Llama 4 Scout Instruct | 9.06% | 2025-07-20 | 0.0.0 | Not stated | $57.82 | Yes | No |
| Qwen2.5-Coder 32B Instruct | mini-SWE-agent | Qwen2.5-Coder 32B Instruct | 9% | 2025-08-03 | 1.0.0 | Not stated | $34.06 | Yes | Yes |
Other harnesses the SWE-bench team checked
30 Verified rows that use a harness other than mini-SWE-agent and that the board marks as team-checked. The full list also has 103 submitted rows without that mark. The highest of those is Sonar Foundation Agent + Claude 4.5 Opus (79.2% on 2025-12-05) and live-SWE-agent + Claude 4.5 Opus medium (20251101) (79.2% on 2025-12-15). They are named here and left out of the table.
| OpenHands + GPT-5 | OpenHands | GPT-5 | 71.8% | 2025-08-07 | — | Not stated | Not listed | Yes | No |
| Moatless Tools + Claude 4 Sonnet | Moatless Tools | Claude 4 Sonnet | 70.8% | 2025-06-11 | — | Not stated | Not listed | Yes | No |
| OpenHands + Claude 4 Sonnet | OpenHands | Claude 4 Sonnet | 70.4% | 2025-05-24 | — | Not stated | Not listed | Yes | No |
| SWE-agent + Claude 4 Sonnet | SWE-agent | Claude 4 Sonnet | 66.6% | 2025-05-22 | — | Not stated | Not listed | Yes | No |
| OpenHands | OpenHands | Undisclosed | 65.8% | 2025-04-15 | — | Not stated | Not listed | Yes | No |
| OpenHands + Kimi K2 | OpenHands | Kimi K2 | 65.4% | 2025-07-16 | — | Not stated | Not listed | Yes | Yes |
| SWE-agent + Claude 3.7 Sonnet w/ Review Heavy | SWE-agent | Claude 3.7 Sonnet w/ Review Heavy | 62.4% | 2025-02-25 | — | Not stated | Not listed | Yes | No |
| OpenHands + 4x Scaled (2024-02-03) | OpenHands | 4x Scaled | 60.8% | 2025-02-03 | — | Not stated | Not listed | Yes | No |
| Nemotron-CORTEXA | Nemotron-CORTEXA | Multiple | 58.2% | 2025-04-10 | — | Not stated | Not listed | Yes | No |
| OpenHands + CodeAct v2.1 (claude-3-5-sonnet-20241022) | OpenHands | CodeAct v2.1 (claude-3-5-sonnet-20241022) | 53% | 2024-10-29 | — | Not stated | Not listed | Yes | No |
| Composio SWE-Kit (2024-10-25) | Composio SWE-Kit | Multiple | 48.6% | 2024-10-25 | — | Not stated | Not listed | Yes | No |
| AppMap Navie v2 | AppMap Navie v2 | Multiple | 47.2% | 2024-11-06 | — | Not stated | Not listed | Yes | No |
| Skywork-SWE-32B + TTS(Bo8) | Skywork-SWE-32B | TTS(Bo8) | 47% | 2025-06-16 | — | Not stated | Not listed | Yes | Yes |
| OpenHands + DevStral Small 2505 | OpenHands | DevStral Small 2505 | 46.8% | 2025-05-20 | — | Not stated | Not listed | Yes | Yes |
| Agentless Lite + O3 Mini (20250214) | Agentless Lite | O3 Mini | 42.4% | 2025-02-14 | — | Not stated | Not listed | Yes | No |
| SWE-agent + SWE-agent-LM-32B | SWE-agent | SWE-agent-LM-32B | 40.2% | 2025-05-11 | — | Not stated | Not listed | Yes | Yes |
| SWE-agent + DevStral Small 2507 | SWE-agent | DevStral Small 2507 | 38% | 2025-07-25 | — | Not stated | Not listed | Yes | Yes |
| Skywork-SWE-32B | Skywork-SWE-32B | Qwen2.5 Coder 32B Instruct | 38% | 2025-06-16 | — | Not stated | Not listed | Yes | Yes |
| SWE-agent + Claude 3.5 Sonnet | SWE-agent | Claude 3.5 Sonnet | 33.6% | 2024-06-20 | — | Not stated | Not listed | Yes | No |
| SWE-Fixer (Qwen2.5-7b retriever + Qwen2.5-72b editor) | SWE-Fixer | Qwen2.5 (7B + 72B) | 32.8% | 2025-03-06 | — | Not stated | Not listed | Yes | Yes |
| AppMap Navie + GPT 4o (2024-05-13) | AppMap Navie | GPT-4o | 26.2% | 2024-06-15 | — | Not stated | Not listed | Yes | No |
| SWE-agent + GPT 4o (2024-05-13) | SWE-agent | GPT-4o | 23.2% | 2024-07-28 | — | Not stated | Not listed | Yes | No |
| SWE-agent + GPT 4 (1106) | SWE-agent | GPT-4 (1106) | 22.4% | 2024-04-02 | — | Not stated | Not listed | Yes | No |
| SWE-agent + Claude 3 Opus | SWE-agent | Claude 3 Opus | 15.8% | 2024-04-02 | — | Not stated | Not listed | Yes | No |
| RAG + Claude 3 Opus | RAG baseline | Claude 3 Opus | 7% | 2024-04-02 | — | Not stated | Not listed | Yes | No |
| RAG + Claude 2 | RAG baseline | Claude 2 | 4.4% | 2023-10-10 | — | Not stated | Not listed | Yes | No |
| RAG + GPT 4 (1106) | RAG baseline | GPT-4 (1106) | 2.8% | 2024-04-02 | — | Not stated | Not listed | Yes | No |
| RAG + SWE-Llama 7B | RAG baseline | SWE-Llama 7B | 1.4% | 2023-10-10 | — | Not stated | Not listed | Yes | No |
| RAG + SWE-Llama 13B | RAG baseline | SWE-Llama 13B | 1.2% | 2023-10-10 | — | Not stated | Not listed | Yes | No |
| RAG + ChatGPT 3.5 | RAG baseline | GPT-3.5 | 0.4% | 2023-10-10 | — | Not stated | Not listed | Yes | No |
SWE-bench Multilingual, same harness
300 instances, 13 mini-SWE-agent rows on the same official file. This is a separate task set from Verified, so the percents are not a second reading of the 500 Verified issues.
| Gemini 3 Flash | mini-SWE-agent | Gemini 3 Flash | 72.7% | 2026-02-13 | 2.0.0a0 | Not stated | $105.55 | Yes | No |
| Claude 4.6 Opus | mini-SWE-agent | Claude 4.6 Opus | 72% | 2026-02-13 | 2.0.0a0 | Not stated | $198.89 | Yes | No |
| Claude 4.5 Opus | mini-SWE-agent | Claude 4.5 Opus | 70.7% | 2026-02-13 | 2.0.0a0 | Not stated | $250.07 | Yes | No |
| GLM 5 | mini-SWE-agent | GLM 5 | 69.7% | 2026-02-13 | 2.0.0a0 | Not stated | $192.74 | Yes | Yes |
| Gemini 3 Pro | mini-SWE-agent | Gemini 3 Pro | 68.7% | 2026-02-13 | 2.0.0a0 | Not stated | $306.33 | Yes | No |
| MiniMax M2.5 | mini-SWE-agent | MiniMax M2.5 | 68.3% | 2026-02-16 | 2.0.0a0 | Not stated | $30.06 | Yes | Yes |
| Kimi K2.5 | mini-SWE-agent | Kimi K2.5 | 67.3% | 2026-02-13 | 2.0.0a0 | Not stated | $207.85 | Yes | Yes |
| Claude 4.5 Sonnet | mini-SWE-agent | Claude 4.5 Sonnet | 67% | 2026-02-13 | 2.0.0a0 | Not stated | $200.86 | Yes | No |
| GPT 5.2 (high) | mini-SWE-agent | GPT 5.2 | 66.7% | 2026-02-13 | 2.0.0a0 | high | $160.92 | Yes | No |
| GPT 5.2 Codex | mini-SWE-agent | GPT 5.2 Codex | 66.3% | 2026-02-20 | 2.0.0 | Not stated | $198.65 | No | No |
| Claude 4.5 Haiku | mini-SWE-agent | Claude 4.5 Haiku | 64.7% | 2026-02-13 | 2.0.0a0 | Not stated | $113.85 | Yes | No |
| DeepSeek V3.2 | mini-SWE-agent | DeepSeek V3.2 | 59% | 2026-02-13 | 2.0.0a0 | Not stated | $115.15 | Yes | Yes |
| GPT 5 mini | mini-SWE-agent | GPT 5 mini | 39.7% | 2026-02-13 | 2.0.0a0 | Not stated | $15.49 | Yes | No |
Terminal-Bench 4.0
35 rows. The official leaderboard for Terminal-Bench 4.0. The board name on Harbor Hub is 4-0-0, and the record was updated 2026-10-23. Read November 5, 2026 from the Terminal-Bench site through the public leaderboard it displays. Click a model or agent name when the board supplied a link.
| 1 | Opus 5.5 | Claude Code | max | 64.85% | ± 3.11 | 214 / 330 | $4,718.65 | 2026-09-22 |
| 2 | Sonnet 5.5 | Claude Code | max | 61.82% | ± 2.94 | 204 / 330 | $7,333.64 | 2026-09-28 |
| 3 | GPT-6 Astra | Codex | max | 58.18% | ± 2.79 | 192 / 330 | $3,267.18 | 2026-09-03 |
| 3 | GPT-6.1 Sol | Codex | max | 58.18% | ± 3.11 | 192 / 330 | $634.24 | 2026-09-29 |
| 5 | Fable 5.1 | Claude Code | max | 57.88% | ± 3.76 | 191 / 330 | $6,243.50 | 2026-09-01 |
| 5 | Fable 5.1 | Claude Code | xhigh | 57.88% | ± 3.36 | 191 / 330 | $4,872.04 | 2026-09-01 |
| 5 | GPT-6 Astra | Codex | high | 57.88% | ± 2.97 | 191 / 330 | $2,269.42 | 2026-09-03 |
| 5 | GPT-6 Astra | Codex | xhigh | 57.88% | ± 2.72 | 191 / 330 | $2,350.51 | 2026-09-03 |
| 9 | Fable 5.1 | Claude Code | high | 54.55% | ± 3.44 | 180 / 330 | $3,985.38 | 2026-09-01 |
| 10 | GPT-6 Astra | Codex | medium | 54.24% | ± 2.66 | 179 / 330 | $1,914.80 | 2026-09-03 |
| 11 | Fable 5.1 | Claude Code | medium | 53.94% | ± 3.39 | 178 / 330 | $2,832.90 | 2026-09-01 |
| 11 | Opus 5 | Claude Code | xhigh | 53.94% | ± 3.17 | 178 / 330 | $6,086.22 | 2026-07-24 |
| 13 | Opus 5 | Claude Code | max | 51.82% | ± 3.39 | 171 / 330 | $5,969.11 | 2026-07-24 |
| 14 | GPT-6 Astra | Codex | low | 50.61% | ± 2.75 | 167 / 330 | $1,557.30 | 2026-09-03 |
| 15 | Opus 5 | Claude Code | high | 50.3% | ± 3.73 | 166 / 330 | $4,662.27 | 2026-07-24 |
| 16 | GPT-6 Sol | Codex | max | 49.39% | ± 3.23 | 163 / 330 | $1,451.32 | 2026-09-22 |
| 17 | Opus 5 | Claude Code | medium | 44.85% | ± 3.83 | 148 / 330 | $3,191.59 | 2026-07-24 |
| 18 | Fable 5 | Claude Code | max | 44.55% | ± 3.85 | 147 / 330 | $7,265.01 | 2026-06-09 |
| 19 | Fable 5.1 | Claude Code | low | 43.33% | ± 3.61 | 143 / 330 | $2,358.72 | 2026-09-01 |
| 20 | GLM-5.3 | Claude Code | max | 41.82% | ± 3.23 | 138 / 330 | $2,727.63 | 2026-08-14 |
| 21 | Grok 4.7 | Grok Build | xhigh | 37.58% | ± 3.54 | 124 / 330 | $3,683.29 (partial: 324/330 trials) | 2026-09-21 |
| 22 | GPT-5.6 Sol | Codex | max | 37.27% | ± 3.78 | 123 / 330 | $2,541.70 | 2026-06-26 |
| 23 | GLM-5.3-Flash | Claude Code | none | 35.76% | ± 3.54 | 118 / 330 | $118.74 | 2026-08-26 |
| 24 | Opus 5 | Claude Code | low | 34.85% | ± 3.94 | 115 / 330 | $2,393.88 | 2026-07-24 |
| 25 | Qwen3.8-Max-0902 | Claude Code | max | 26.97% | ± 3.78 | 89 / 330 | $4,635.67 | 2026-09-02 |
| 26 | Opus 4.8 | Claude Code | max | 23.64% | ± 3.56 | 78 / 330 | $6,481.26 | 2026-05-28 |
| 27 | GPT-5.6 Terra | Codex | max | 21.52% | ± 3.25 | 71 / 330 | $1,733.52 | 2026-06-26 |
| 28 | Grok 4.6 | Grok Build | high | 20.3% | ± 3.09 | 67 / 330 | $3,591.58 | 2026-08-12 |
| 29 | Gemini 3.8 Flash | mini-SWE-agent | high | 19.09% | ± 3.36 | 63 / 330 | $1,828.77 | 2026-09-02 |
| 30 | GPT-5.6 Luna | Codex | max | 17.27% | ± 2.85 | 57 / 330 | $346.67 | 2026-06-26 |
| 31 | GPT-6 Luna | Codex | max | 16.36% | ± 2.72 | 54 / 330 | $72.14 | 2026-09-22 |
| 32 | Muse Spark 1.3 | Muse Code | xhigh | 14.55% | ± 2.94 | 48 / 330 | $0.00 | 2026-09-02 |
| 33 | Grok 4.5 | Grok Build | high | 12.42% | ± 2.62 | 41 / 330 | $2,094.11 | 2026-07-16 |
| 33 | Sonnet 5 | Claude Code | max | 12.42% | ± 3.06 | 41 / 330 | $9,603.86 | 2026-06-30 |
| 35 | Gemini 3.7 Flash | mini-SWE-agent | high | 11.21% | ± 2.45 | 37 / 330 | $1,261.87 | 2026-08-13 |
Left out
- Third-party score roundups, including pages that list SWE-bench results above the official board, are not used. If the official file does not contain the row, it is not on this page.
- SWE-bench Full, Lite, and Multimodal are on swebench.com and were not copied into this month's tables.
- OSWorld, WebArena, GAIA, and SWE-bench Pro were not read from an official board for this edition, so they are not given scores here.
- The November 3, 2026 paper "You don't need long context" reports its own SWE-bench Verified and Terminal-Bench 2.1 runs. Those are the authors' runs, not rows on the public boards, and they stay on the research note.
FAQ
Does a higher percent mean a better product?
Only inside one table, and only for the harness named on the row. A bash-only SWE-bench score and a Terminal-Bench score are different tests. A mini-SWE-agent 1.x row and a 2.x row are not described by SWE-bench as a clean comparison.
Why are some famous models missing?
They were not on the official file when it was read on November 5, 2026. A score that exists only on a recap site is left out. The newest SWE-bench Verified date in that file is 2026-02-26.
Who checked the runs?
The SWE-bench tables have a team-checked column taken from the official file. Terminal-Bench rows are the rows on the official 4.0 board. This site did not re-run the tasks.
How often does this page change?
It is the November 2026 edition. A later month gets its own page when the boards are read again. The reports index is where that page will be linked.
Cite or embed this
The citation names this page and the date above. The embed is a blockquote with a link back here.
Citation
AI Agents Info. "AI Agent Harness Benchmark, November 2026." Last updated November 5, 2026. https://www.aiagentsinfo.com/reports/agent-harness-benchmark-november-2026.
Embed
<blockquote cite="https://www.aiagentsinfo.com/reports/agent-harness-benchmark-november-2026"> <p>Public SWE-bench and Terminal-Bench scores for agent harnesses and coding agents, each with the date on the official source. Read November 5, 2026.</p> <footer><a href="https://www.aiagentsinfo.com/reports/agent-harness-benchmark-november-2026">AI Agent Harness Benchmark, November 2026</a>, AI Agents Info, last updated November 5, 2026.</footer> </blockquote>