Scaling RSI

RL training environments for long-horizon SWE and ML.

300+ environments ready for training

We’re building training environments for recursive self-improvement: an agent’s ability to help build the next version of itself.

We believe high-quality RSI data is the bottleneck. Our environments cover improving sample efficiency and model architectures, debugging training runs, and making open-ended research decisions.

We have partnered with several frontier labs and are committed to increasing sovereign intelligence worldwide.

Results

Model results. Mean continuous reward with bars on a zero to one scale, mean wall-clock time, and mean agent turns rounded to whole numbers, with mean cost per rollout in US dollars, and average total input and output tokens in millions.
Model cohort	Mean reward	Mean wall-clock time	Mean agent turns	Mean cost per rollout	Avg. total input tokens (M)	Avg. output tokens (M)
GPT 6 Astra	0.235256	3h 47m 00s	450	$107.64	74.933	0.604
Claude Opus 5.5	0.181512	3h 04m 02s	190	$29.97	61.015	0.318
Gemini 4 Argon	0.157907	3h 34m 12s	233	$19.63	84.697	0.413
GLM 5.3	0.071541	2h 51m 56s	151	$4.98	18.624	0.115

Scaling RSI

Results

Our team comes from