Output Loop Bench · Artifacts
Workflows: AI models compared
Real workflows on saved inputs, by model and technique. 19 models tested on the same 2 tasks.
Posting assistant: build
Can a fast model write a small, working automation from a plain-English job? Step 1 of 2: the model writes and saves the workflow; the Lab then runs it on a fresh inbox it never saw.
Posting assistant: build: all 7 results
| Model | Maker | Time | Tokens |
|---|---|---|---|
| claude-haiku-5-5 | Anthropic | 287.1s | 3,465,769 |
| nemotron-3.5-lightning-30b-a3b | NVIDIA | 371.9s | 1,040,306 |
| qwen3.8-flash | Alibaba (Qwen) | 417.1s | 648,314 |
| gpt-6-luna | OpenAI | 539.5s | 2,280,003 |
| deepseek-v4.1-flash | DeepSeek | 752.8s | 2,591,772 |
Page manager: build
Frontier goes one step further than Fast: the workflow also runs the page day to day. Step 1 of 2: the model writes and saves the workflow; the Lab then runs it on a simulated day it never saw.
Page manager: build: all 12 results
| Model | Maker | Time | Tokens |
|---|---|---|---|
| kimi-k2.7-code | Moonshot (Kimi) | 169.4s | 603,831 |
| claude-sonnet-5-5 | Anthropic | 203.5s | 1,771,958 |
| MiniMax-M2.5 | MiniMax | 248.8s | 455,507 |
| claude-opus-5-5 | Anthropic | 258.2s | 1,902,646 |
| gpt-6.1-sol | OpenAI | 263.0s | 375,261 |