Gemini 3.5 Flash-Lite (Google API, free tier)
117 episodes: 117 tasks across 7 worlds × seed 1. Same tools, prompt, step budget and verifiers as every model on the board. Run 2026-10-22 through the provider's API, with the harness on Apple M1 Pro, 32 GB RAM, darwin · harness edb0b13.
Google's Gemini API on the free tier, through its OpenAI-compatible endpoint, with the model's default thinking. Requests spaced at least 7 seconds apart to stay inside the free tier's per-minute limit; same prompt, tools, 40-turn cap and verifiers as every other run.
80%
94 of 117 passed every check
0 of 12
Injection episodes where it followed instructions hidden in app data · 12 passed
8 of 12
Episodes needing a detail only the user knows · 8 passed
5.3
0.6% of tool calls errored
60,260
37 s per episode
Outcomes
Every episode, passed or failed, with failures by their main reason's family.
80%
16%
3.4%
Why episodes failed
Main reason for each of the 23 failed episodes, read off the verifier's checks.
What each reason means
By world
- Quickbasket
- Coinpurse
- Skyfare
- Ledgerline
- Crewboard
- Daybook
- Worksuite
Every episode
Open one to read the instruction, each tool call the model made, and every verifier check.