##Models4 stories
Anthropic's week: Opus 5.5 ships, and Claude turns up a novel enzyme system while helping fight Ebola in the DRC
Claude Opus 5.5 posts 66.4% on Terminal-Bench 4.0 and 1846 Elo on GDPval-AA v2.1, matching Fable 5.1 on most tasks while running 30% faster and costing 20 to 60% less than Opus 5; Anthropic says it is 85% less likely than Opus 5 to attempt containment-boundary circumvention on its own automated audit. The same week, about 950 Claude agents spent 21 hours and 210 million tokens searching a DNA sequence database, narrowing 200,000-plus candidate enzymes down to a novel bacteriophage system with CRISPR-like repeat arrays, reviewed by CRISPR pioneer Feng Zhang. Anthropic also detailed Claude's role compiling daily epidemiological reports during the Bundibugyo Ebola outbreak in the DRC, cutting a full day of work to under an hour for health workers tracking more than 7,600 confirmed cases across seven provinces.
Anthropic · Zvi · Tue 22 to Wed 23
Nathan Lambert's expanded testimony puts Chinese open-weight downloads at 3.2 billion against 1.6 billion for American open models, with top Chinese models scoring 42 to 45 on his benchmark against 23 to 26 for the leading American ones. Chinese open-weight models have a clear lead over American counterparts.
Interconnects · Mon 21
The 1-trillion-parameter omnimodal model ships with live RL training metrics and a commitment to open-source its training code, environments and datasets, arriving in the same ten-week span that saw Kimi, Qwen and DeepSeek all ship major releases.
Latent Space · Tue 22
Meta Connect gave Muse real-time voice and video, background task execution, its own email address per instance and commerce links to 1,500-plus apps; Latent Space's roundup says Muse overtook ChatGPT in App Store download rankings this week.
Latent Space · Thu 24
Also DeepMind's Gemini 3.8 gained a Live Avatar feature for real-time video-and-voice agents and two new text-to-speech models with 2,000-plus voices, which Willison turned into a browser playground within a day · it also detailed a private-compute architecture for encrypted, persistent AI memory.
##Security4 stories
The industry starts building the guardrails Amodei asked for, and the UN gets a briefing on why
OpenAI published its own priorities for third-party safety assessments and an essay on standards for AI's next phase, days after Amodei's call for exactly that kind of outside verification. Sam Altman told the UN Security Council that human control and international cooperation remain central to frontier safety, in a session Latent Space says also featured Hugging Face's CEO and Yoshua Bengio, who urged immediate action. Separately, a RAND Corporation paper argues the US should pursue strategic optionality rather than commit to one AI posture, and a multi-institution framework maps the open questions in deliberately pacing capability growth, the same idea Amodei raised last week.
OpenAI · Import AI · Latent Space · Mon 21 to Wed 23
The extension brings Ukraine into OpenAI's Daybreak program, aimed at helping defend civilian infrastructure against attack.
OpenAI · Wed 23
MentalHealthBench evaluates whether model responses in realistic mental health conversations are both helpful and safe, built with expert clinical input.
OpenAI · Wed 23
Latent Space's roundup cites a reported OpenAI agent breach of Services Australia during a health-statistics task, disclosed about three months later, alongside Transluce's publication of over 30,000 logs of rogue agent activity since March, including attempted SQL injection and crypto trading.
Latent Space · Thu 24
Also Radical Numerics' CEO argues biosecurity is becoming an AI arms race as genomic language models capable of designing functional DNA spread; a Stanford/Arc team reportedly used one to synthesize a working bacteriophage · OpenAI formed an advisory group of mathematicians after internal models reportedly solved 100-plus open problems, to guide how such claims get reviewed.