Anthropic had Claude generate mathematical proofs that existing code is correct across 123 open-source libraries, producing approximately 180 million lines of proof code in the process. The model identified bugs that years of conventional testing had missed, using formal verification to validate code correctness. (The Decoder)
Cohere released Rerank 5, a retrieval ranking model available in Pro and Fast tiers with a 32K context window. The new version adds Context Focus, which extracts passage-level context from documents, and implements calibrated relevance scores. The Pro tier scores 82.2 on the BRIGHT benchmark and costs $0.05 per 1 million tokens. (MarkTechPost) Separately, Liquid AI released RIPL, a real-time inference engine for its LFM models, achieving 14.4 to 28.4 times lower p95 latency on the 1.2B parameter LFM2.5 compared to Qwen3.5-0.8B on vLLM on a single B200 GPU. (MarkTechPost) Gradium open-sourced Phonon, a 31-million-parameter on-device text-to-speech model under Apache 2.0 license that fits in 64.5 MB and achieves approximately 7 millisecond latency on laptop CPUs. (MarkTechPost)