Output Loop Bench · Search
Deep research: AI models compared
Find out everything on a fixed subject. 28 models tested on the same task.
How UK homeowners find and choose a tradesperson
Trades are our core market, so this is research we actually need. Each model searches the web on the same brief and writes a cited report. The Lab checks it against an answer key of 10 verified 2025-2026 facts, counts the distinct sources it cites, and opens every link to catch dead or made-up URLs. You see who finds the most, who cites honestly, and how fast.
| Model | Maker | Time | Tokens |
|---|---|---|---|
| muse-spark-1.3-medium | Meta | 20.7s | 180,148 |
| sonar-pro | Perplexity | 23.3s | — |
| MiniMax-M2.5 | MiniMax | 39.6s | 47,860 |
| composer-2.5 | Cursor | 45.4s | 128,525 |
| kimi-k2.7-code | Moonshot (Kimi) | 77.8s | 346,755 |
| qwen3.7-plus | Alibaba (Qwen) | 91.4s | 120,594 |
| kimi-k3-max | Moonshot (Kimi) | 96.7s | 441,908 |
| deepseek-v4-pro | DeepSeek | 99.5s | 193,276 |
| sonar-deep-research | Perplexity | 122.6s | — |
| claude-opus-5-5 | Anthropic | 127.9s | 2,936,780 |
| claude-sonnet-5-5 | Anthropic | 130.0s | 2,236,237 |
| gpt-6.1-sol | OpenAI | 130.1s | 342,666 |
| claude-haiku-5-5 | Anthropic | 136.4s | 598,959 |
| Perplexity Agent (high) | Perplexity | 160.1s | — |
| claude-fable-5-1 | Anthropic | 170.1s | 2,644,779 |
| gpt-6-astra | OpenAI | 173.4s | 1,308,424 |
| gemini-4-argon-medium | 190.8s | 1,200,122 | |
| grok-4.7 | xAI | 203.7s | 627,042 |
| deepseek-v4.1-flash | DeepSeek | 219.5s | 626,828 |
| grok-4.7-build-fast | xAI | 231.8s | 974,277 |
| qwen3.8-flash | Alibaba (Qwen) | 236.1s | 351,030 |
| gemini-3.1-pro | 272.7s | 1,386,412 | |
| glm-5.3 | Z.ai | 307.8s | 436,281 |
| muse-glimmer-30b | Meta | 348.0s | 387,422 |
| gpt-6-luna | OpenAI | 478.3s | 2,317,584 |
| nemotron-3.5-lightning-30b-a3b | NVIDIA | 764.3s | 281,045 |
| qwen3.8-max | Alibaba (Qwen) | 810.5s | 1,049,015 |
| glm-5.3-flash | Z.ai | 838.1s | 370,316 |