The new model is optimized for high-volume tasks and is the fastest model released by the company. Running costs have dropped by 75%, and Sonnet 5.5 cache reads are now half the price.
The author details technical constraints on tool result size and streaming steps, while clarifying that developers own the failure when agents act incorrectly.
The AgentCompile system optimizes repetitive agent tasks, reducing model calls by 58% while maintaining answer accuracy. On test data with Claude Sonnet 4.5, agent token spending dropped by 36%.
The company detected 97% of the content before users reported it and launched a new LLM system to identify ads that secretly link to illegal material.
Researchers from the National Institute for Innovation in Manufacturing Biopharmaceuticals and the University of Delaware demonstrated how R&D and supply chain data become predictive assets. The MELLODDY Consortium model enables standardized sharing of key insights while protecting proprietary information.
A Longevity.Technology report indicates Chinese companies complete clinical trial phases 50–70% faster than US counterparts. Among the 44 tracked companies that disclosed funding, total raised reached $16.9 billion, with Innovent Biologics accounting for roughly 70% of this amount.
Blocking an age-related protein restored cartilage in aged mice and prevented osteoarthritis after injuries, while human cartilage began producing new functional tissue.
The Consumer Technology Association calculated the costs for businesses to achieve Trump's goal of 100% domestic manufacturing, highlighting the impracticality of the plan.
Healthleap secured $38 million in seed and Series A funding from investors including Sequoia Capital to deploy an AI platform that scans patient records for early signs of undiagnosed conditions like malnutrition or delirium.
Using 32-bit pointers on 64-bit systems reduces memory consumption by nearly half and improves performance through better cache utilization. A Mastodon deployment experiment cut memory usage from 650 MB to 350 MB, though widespread adoption requires manual compilation since most software is not pre-packaged for x32.
Unsloth enables training any text or vision LLM as a decision model using QLoRA, boosting accuracy from 30 to 80 percent. The tool supports exporting Clef models to GGUF and serving them via llama.cpp.
The FDA-cleared Cala kIQ Plus device offers three stimulation waveform modes, resulting in tremor reduction for 76–89% of participants in a recent study.
Exercise increases autophagy activity and reduces senescent cell burden, slowing aging and improving cognitive function in cancer survivors.
Researchers developed a 3D-printed GHz microwave resonator that maintains electron spin-qubit coherence times on the order of 10 milliseconds.
The study proves that recovering a single qubit from black hole radiation demands queries proportional to the Hilbert-space dimension of the remaining system, provided the black hole contains at most one-sixteenth of the original qubits.
Version 1.12.0 introduces capabilities for automatic web searching and browser-based code execution alongside the Claude Haiku-5 model.
The study found that glymphatic system activity in multiple sclerosis patients correlates with sleep metrics and clinical disability levels. Researchers used MRI to assess the link between cerebrospinal fluid flow and microstructural constraints on diffusion tensor imaging.
IEEE Spectrum's white paper introduces the HiPHI dataset, designed to address the data gap hindering humanoid robot learning in tasks like carrying and pushing.
The SafeActBench study found that agents often stop investigating or act without gathering evidence, despite appearing reliable in static assessments. These issues worsen in multi-step workflows due to unresolved prerequisites.
The release removes backend type parameters from user code and introduces adaptive memory management. Tests show rebuilds are up to 15 times faster, with added support for LoRA, QLoRA, and ONNX export.