Sapient Open-Sources PRAXIST: 49 Gold Medals on MLE-bench at a Fraction of the Cost

Sapient Intelligence (the team behind the HRM architecture) released and open-sourced PRAXIST Beta, an autonomous R&D framework built around "lineage": it turns reproducible artifacts and evaluation results into a typed evidence graph, separating local artifact construction from fleet-level evidence synthesis so later attempts inherit verified mechanisms instead of starting from scratch.

Confirmed

• Users define goals, constraints and success criteria; PRAXIST then takes over the research loop, with multiple Research Peers exploring competing paths in parallel, sharing findings, and optimizing on accumulated evidence
• The system retains all attempts rather than only winners, stores learnings in a shared memory (research graph), and uses a PI panel to select the next tests
• It achieved 100% safe landing in a partner-provided rocket landing simulation, versus a 17.1% baseline; industrial SLAM applications were also mentioned
• Across all 75 MLE-bench tasks, it earned 49 gold medals using open-source models for $3,054
• The project is open-source, positioned to turn experiments into verified production-grade solutions

Unconfirmed

• Claims that the cost is roughly one-twelfth of Claude Code (per @SignalCompetitive582) and that results beat Claude Opus 4.8 (per @rohanpaulai) are community retellings without official comparison data in the posts

Why it matters

• PRAXIST demonstrates multi-agent parallel exploration plus evidence inheritance as a viable paradigm for autonomous R&D agents, structuring exploration itself as an inheritable lineage
• Its MLE-bench results and cost figures suggest open-source models with the right architecture can compete on hard ML engineering benchmarks at very low cost, contrasting with expensive closed-source agent solutions

2026-08-28 ~ 2026-08-29 · 8 related posts

• [source] PRAXIST Multi-Agent Achieves 100% Rocket Landing Success Rate — JaynitMakwana · 2026-08-28
• Praxist Agent Achieves High Cost-Efficiency on MLE-bench — iamfakhrealam · 2026-08-28
• PRAXIST: Autonomous research framework with multi-peer parallelism — Aiden_Tech_Ai · 2026-08-28
• Open Source PRAXIST: Autonomous Research Team Achieves 100% Safe Landing in Rocket Sim — iamfakhrealam · 2026-08-28
• [source] PRAXIST open-sourced: 49 golds on MLE-bench at 1/12th the cost of Claude Code — SignalCompetitive582 · 2026-08-28
• Sapient Open-Sources PRAXIST: 49 Gold Medals on MLE-bench, Beating Claude Opus 4.8 — rohanpaul_ai · 2026-08-28
• [source] PRAXIST: Improving AI Research via Multi-Agent Collaboration and Shared Memory — iamfakhrealam · 2026-08-29
• PRAXIST coordinates AI agents for complex tasks like rocket landing — iamfakhrealam · 2026-08-29