Logical Intelligence Aleph tops formal verification benchmarks with

Logical Intelligence’s AI coding agent Aleph solved 99.4% of the PutnamBench problems, the highest score among public formal reasoning benchmarks [2]. The system also posted a 94% success rate on VeriSoftBench, achieved state‑of‑the‑art results on LeanEval, and earned a perfect score on Verina [2].

Founded by Eve Bodina in San Francisco, the startup builds energy‑based reasoning models (EBRMs) that generate machine‑checkable proofs instead of merely plausible code [1]. Aleph’s performance marks a sharp jump from the sub‑2% solve rates seen on PutnamBench a year earlier, with the agent correctly handling 668 of the benchmark’s 672 problems [2]. The company says the agent is already deployed in production verification workflows, including work with the Ethereum Foundation’s cryptographic libraries [1][2].

Bodina argues that as AI‑generated software scales, “being mostly right is effectively the same as being wrong” in mission‑critical environments, making formal verification a prerequisite for sectors such as semiconductor design, finance, and energy systems [2]. She positions Aleph’s benchmark wins as evidence that verified code generation is moving from theory to practice, a shift that could reshape how organizations handle AI‑produced code [2].

The beta launch planned for later this year will target critical infrastructure operators, offering a tool that replaces months‑long manual verification with repeatable, scalable proof generation [1][2]. If Aleph can maintain its benchmark performance in real‑world deployments, it may set a new baseline for AI‑assisted software development, forcing the industry to adopt formal verification as a standard safety net. The open question remains whether other AI labs can match Aleph’s results, and how quickly the broader market will demand provable correctness over mere plausibility.