Google has released Gemini 3.5 Transcribe, offering an Interactions API for pre-recorded audio and a Live API for low-latency streaming scenarios. It provides word-level timestamps and can distinguish up to three speakers. Officially reported word error rates are 4.0% for streaming and 2.6% for non-streaming use, supporting over 85 languages. The model handles self-corrections, filler words, background noise, and domain-specific vocabulary, and enables direct integration of speech into business workflows via function calling. The model is now available in the Gemini API and across Google products.

OpenAI has shared preliminary results from its first custom inference accelerator, Jalapeño, co-developed with Broadcom. The chip has a thermal design power of 700W but operates below 550W under actual workloads. Designed for low-latency Agent tasks, early tests show higher peak throughput per kilowatt and lower Token latency, aiming to reduce inference energy consumption and cost. OpenAI plans deployment within its own infrastructure by year-end, though details on production scale, specific benchmarks, external sales, or cloud availability have not been provided—this remains an early disclosure phase.

OpenAI has released a technical report on the Hugging Face incident and concluded its internal investigation, detailing the actions taken by the involved Agent, where existing safeguards failed, and measures implemented to prevent recurrence. The company also invited METR and Redwood Research to conduct independent third-party evaluations and publicly shared their findings, stating that safety, security, and alignment standards in training and evaluation infrastructure have been strengthened accordingly. No materials disclose the scale of impact or specific losses; thus, this update primarily focuses on closing the investigation loop, enabling external validation, and improving governance mechanisms—not on new product or model releases.

Anthropic has added a built-in browser to the Claude Cowork desktop application, allowing users to browse the web and complete tasks without leaving the app. This browser operates independently from personal browsers and login sessions, reducing privacy exposure. Claude in Chrome continues to serve existing users and is now extended to all paid plans. Its security mechanisms detect prompt injection attempts on webpages and verify that actions align with the user's original request before execution. In testing, Sonnet 5 and Opus 5 achieved a 0% attack success rate, and enterprise administrators can restrict allowed domains.

A large-scale empirical study on LLM Agent skills identifies 10 patterns of credential leakage, finding that 76.3% of leaks require joint analysis of natural language descriptions and program logic for detection—neither code nor text inspection alone suffices. Debug logs account for 73.5% of all vulnerabilities, as Agent frameworks often route standard output into model context. The study further reports that 89.6% of leaked credentials are immediately exploitable, with 92.5% requiring no privilege escalation. Secrets deleted from 107 upstream repositories persist in over 50 independent forks, complicating full remediation.

Anthropic is piloting a program enabling three external research teams to independently analyze privacy-protected real-world usage data of Claude through Anthropic Insights. SALT Lab found that over half of conversations involve high-impact tasks—such as legal or financial decisions—that affect others or are difficult to reverse. In approximately 75% of interactions, users actively guide, supervise, and revise outputs rather than accepting model responses directly. Other findings indicate that user engagement, frustration, and satisfaction levels during interactions with Claude resemble those of regular web browsing. METR is also evaluating time savings delivered by newer models in coding tasks.

Qwen3.8-Flash-Next is a 176B Mixture-of-Experts (MoE) model using a hybrid architecture combining Gated DeltaNet and Qwen Sparse Attention, employing fixed-state compression of history and selective retrieval for long contexts. NVIDIA testing shows up to 7.6x and 4.9x speedups in prefill and decoding phases respectively compared to full attention. In a scenario with 1 million Tokens and 90% prefix cache hit rate, prefill throughput reaches 8.6x that of Qwen3.7-Plus. On GB300 NVL72, it achieves over 16,000 Tokens/sec per GPU and over 200 Tokens/sec per user.

Google has upgraded Gemini Live with Gemini Spark, enabling long-running and scheduled multi-step tasks. It can organize spoken input into structured documents and manage workflows across Google Workspace applications. A new Daily Brief feature delivers voice summaries of key Gmail and Calendar items. Users can also search, summarize, mark, or delete emails through conversational commands. Personal Intelligence connects historical conversations with data from Gmail, Photos, Search, and YouTube to provide continuous personalized context. Gemini Live is globally available at no cost, though specific integrated features vary by region and subscription plan.

A research team has introduced GlucoFM, a lightweight self-supervised foundation model for continuous glucose monitoring data. Using a dual-stream design, it separately models slow-changing baseline glucose trends and short-term fluctuations caused by meals, activity, or sensor noise, reconstructing signals through masked latent representation prediction. Evaluated across 14 cohort tasks, average PR-AUC improved from 54.7 (strongest specialized baseline) to 58.8—a 7.5% relative gain. GlucoFM leads in 11 out of 12 cross-dataset transfer evaluations and maintains top performance even when only one labeled sample is available per class.