Anthropic: Claude Sonnet 5.5 has arrived. The second model in the Claude 5.5 family is a clear step up from Sonnet 5: it generates more than 30% faster and, according to tests, is up to 30% cheaper per task — while the API price stays the same ($2 / $10 per million tokens). The faster, cheaper counterpart to Opus 5.5 is aimed at well-defined everyday tasks, bug fixing, documents, slide decks and spreadsheets; it is the first Sonnet to launch with cyber-defense safeguards similar to those of the most powerful models. Haiku 5.5 is coming in the next few weeks. Source: Anthropic.
NVIDIA: Open Agent Safety Platform — OpenShell on the CPU, Sentry on the chip. The company unveiled an open software platform and reference architecture to keep AI agents within controlled boundaries from testing through deployment. OpenShell is an open-source, secure runtime on the CPU (NVIDIA Vera; extensible to Arm and Intel); Sentry is a “watchdog” running outside the software on the BlueField-4 DPU, isolating any agent that crosses the boundary within milliseconds. According to Jensen Huang, after the recent sandbox escapes, safety engineering has to speed up too. Source: NVIDIA Newsroom, CNBC, The Verge.
Leading AI researchers: if AI R&D becomes automated, an “intelligence explosion” could follow. On Monday, more than twenty authors — including Geoffrey Hinton, Yoshua Bengio, OpenAI’s chief scientist (Jakub Pachocki), Anthropic co-founder Jack Clark, Microsoft’s chief scientific officer (Eric Horvitz) and Meta’s AI research lead (Dawn Song) — warned that automating AI’s own research could compress years into months. The paper proposes transparent R&D reporting, embedded independent auditors, limits on the pace of capability gains and emergency plans. Source: WSJ, The Guardian.
New York City: full-council AI hearing on October 5 — SpaceXAI compelled by subpoena. According to City Council Speaker Julie Menin, Anthropic, OpenAI, Google and Meta will testify under oath at a rare Committee of the Whole session convening all 51 council members. Except for Meta, the companies confirmed attendance only after the threat of a subpoena; SpaceXAI did not respond, so a rare subpoena was issued today. The hearing comes amid recent agent incidents and a local legislative package (third-party enforcement, a “kill switch”, whistleblower-protection incentives). Source: NYC Council, City & State NY.
ElevenLabs: Eleven v4 and Eleven v4 Turbo voice models. The new text-to-speech generation is the most expressive model yet (first in the Artificial Analysis ranking; in a blind test about 75% of listeners preferred it), with more than 90 languages, natural-language performance directions and stable voice clones. v4 Turbo is a low-latency version (about 100 ms median inference, about 150 ms to first audio) for agents and streaming. Source: ElevenLabs.
What this means for you: Safety and regulatory pressure around agents now extends to hardware-level safeguards, lab-level transparency and city hearings. If you use AI agents or vendor models, clarify oversight and incident handling in advance — see how we help with AI governance and oversight.