Anthropic has announced the general availability of Computer Use, Browser Tool, Skills API, and Files API on the Claude Platform, enabling developers to let agents operate applications lacking APIs, upload team workflows, version them, or pin to a specific version_id. The Files API now includes file expiration controls, rate limits increased to 500 RPM, and storage capacity expanded up to 1 TB per organization. The official team also provides examples of integrating Claude Managed Agents with the open AG-UI protocol, covering streaming interactions and frontend integration.

Mistral has introduced Agentic Search, a multi-step loop composed of search, open, navigate, read, and text matching, allowing models to repeatedly locate and verify evidence across long documents and multiple data sources. In FinanceBench testing, accuracy for both Mistral Medium 3.5 and GLM-5.2 improved from approximately 27% to about 86%. Targeted navigation also reduced p90 latency by up to 39.6% and cut token usage by up to one-third. The tool integrates with existing indexes and requires no model fine-tuning.

Liquid AI has released DSpark draft models for its LFM2.5 series, adding approximately 300 million parameters. Using speculative decoding, it batch-generates and validates candidate tokens, achieving up to 3.18x speedup on H100 and up to 2.87x on MacBook edge devices. LFM2.5-2.6B achieves average speedups of 2.67x on H100 and 2.27x on M4 Max, with function call latency decreasing by an average of 57%. Greedy decoding outputs remain consistent with the original model, and integrations via llama.cpp and SGLang are now available.

Superwhisper has released S1-mini, an open-weights text normalization model with 0.6B parameters, designed to convert raw transcripts from any ASR system into polished written text—removing fillers, correcting speech errors, and handling punctuation, capitalization, numbers, dates, currency, and email formatting. Its control fields allow separate specification of tone, structure, and use case. The Q4_K_M quantized version is only 462 MB, runnable on laptop CPUs. The model achieves 94.8% token accuracy on 7,519 held-out samples.

NVIDIA has introduced a generative recommendation solution based on HSTU and Semantic IDs, reframing recommendation tasks as next-item prediction in user behavior sequences, unifying retrieval and ranking in traditional systems. HSTU employs SiLU weights, relative positional bias, and gating mechanisms, optimized for long-sequence training and low-latency serving. Accompanying recsys-examples and nv-embedding-cache enable dynamic embeddings, KV caching, and hierarchical storage management across HBM, DRAM, and remote layers. Official benchmarks report over 2x speed improvement overall.

LangChain has launched public testing for LangSmith Deployment's Preview Builds, enabling temporary agent deployments close to production environments for pull request branches, allowing teams to validate tool calls, behaviors, and modification outcomes before merging. New commits automatically update the corresponding preview, while merging triggers cleanup of temporary environments and initiates standard production deployment. Teams can set triggers by branch or tag, and manage resources via idle timeout and concurrency limits; the demo uses a default 72-hour lifespan and up to five concurrent previews.

Lease End revealed its fine-tuned intent classifier contributed $12 million in revenue, yielding a 50x ROI, but offline metrics failed to expose production issues such as confirmation messages and greetings. While model training took about an hour, the full cycle of data curation, evaluation, and regression feedback typically approached one week, further constrained during migration by vendor-specific artifacts. The team transitioned to a system using Skills, tools, and context, deploying fixes by adjusting Markdown workflows and running targeted evaluations, while retaining the ability to switch between model providers.

A practical analysis shows that when used as judges, LLMs tend to favor content generated by models from the same family, as familiar text styles often yield lower perplexity and thus receive higher scores even when flawed. Judging outcomes are also influenced by answer length and option position, with longer or earlier responses gaining unintended advantages. The analysis recommends using arbiters from different model families and explicitly penalizing irrelevant verbosity in scoring rules. Before formal deployment, calibration with held-out samples and human judgment is advised, with low-consistency categories defaulted to human review.