← All benchmarks

FrontierML ($200 wallet)

Agents · 2026-10-13

Induction Labs' FrontierML gives a coding agent 21 open-ended, end-to-end machine learning tasks spanning language, vision, audio, computer use and games, inspired by work at frontier labs, neolabs and universities: pretraining a language model on 10M words, converting an autoregressive model to text diffusion, ternary quantization, voice endpointing or a Pokémon battle bot. Each trial has a 12-hour CPU sandbox and a $200 wallet for GPUs and OpenRouter tokens bought through Spin at fixed GPU prices; the agent submits an inference server, graded once on an air-gapped H100 against held-out, out-of-distribution tests, and a Claude Opus 5.5 reviewer zeroes runs that break the task rules. Each task's metric is rescaled linearly between measured floor and ceiling anchors and clamped to 0-1, and the score is the mean over the 21 tasks of one trial per task. Models run in their developer's own agent harness at the effort that did best on other agentic benchmarks, and the agent's own model API calls are not charged to the wallet. The $200 wallet is the leaderboard's headline; the $10 and $50 wallets are components.

Top models (higher is better)