Frontier Labs × RSI Levels
Where each frontier lab, and each product, sits on the five levels of recursive self-improvement, based only on what it has said publicly. Every dot opens its evidence: the organisation's own post, what the model did and what was reported. Level placements are our reading of those posts, not the labs' own labels. Claude, made by Anthropic, helps write this site.
Shipped products and open tools grouped by the level of self-improvement they put in a customer's hands. L3 and L5 have no commercial products yet. ⚠️ = preview, limited access or details we could not confirm. 🔗✓ = link checked when added; 🔗 = link not re-checked from our build environment.
What the map shows
- Everyone is at L2. Models that optimise kernels, serving and training code are now routine inside frontier labs.
- The race is at L4: how much of a lab's own R&D the model does while deployed. Measures differ (share of merged code, agent-days, throughput), so the dots compare levels, not labs.
- L5 only in bounded settings: one kernel (Google DeepMind), small models (Recursive), an agent framework (MiniMax). The objective and the verifier stay human-set everywhere.
- Caveats: all figures are organisation-reported; post-cutoff months undercount; absence of a dot means nothing public was found, not that nothing exists.
Evidence comes from the industry cases in data/rsi-papers.json (from the CC0 Awesome-RSI catalog) plus three earlier landmark results added by this site (AlphaZero, FunSearch, OpenAI Five). See the RSI Evolution Map for the 524 papers.