AI-designed viruses work + Anthropic eases bio limits

A Stanford and Arc Institute team led by Brian Hie used the Evo 1 and Evo 2 genome language models to write complete viral genomes from scratch, then built about 300 of them in the lab. Sixteen produced working bacteriophages — viruses that infect bacteria — and three beat the natural template they were modeled on, while a cocktail of the designs broke through E. coli strains that had already evolved resistance. The models were trained with human viruses excluded from the data, and the authors argue that exclusion is what makes the work publishable; several reviewers responded that the harder question is what happens when a lab without that restraint repeats it.

When Fable 5 shipped, its biology classifier was tuned so cautiously that it pushed a wide range of harmless questions down to Claude Opus 5, and Anthropic accepted that cost because a frontier biology model could give real uplift to someone building a weapon. The company has now rewritten the classifier's constitution, carved out explicit exceptions for benign cases, retrained on new data, and reports about 85 percent fewer biology-related fallbacks across its products. Virology, toxicology and molecular design still get routed away — the loosening covers reading lab results and learning biology, not dual-use research.

WeatherNext predicts a storm's track, intensity and wind structure together rather than trading one against the others, and its three-day forecasts now match what prior models managed at two days — a gain the team frames as roughly a decade of meteorological progress. It runs on a 28-kilometer grid and produces a forecast in under one minute on a single tensor processing unit. The National Hurricane Center, the UK Met Office and the Cooperative Institute for Research in the Atmosphere are already using it, and Google published the paper in Nature alongside open code and model weights.

Taalas, founded in Toronto in 2023 by former Tenstorrent chief executive Ljubisa Bajic, burns model weights into a mask read-only memory fabric instead of streaming them from high bandwidth memory, which removes the memory bottleneck that dominates inference cost. The company said in February that its HC1 test chip served Meta's 8 billion parameter Llama 3.1 at 16,960 tokens per second, and its next chip targets 20 billion parameters. AMD did not disclose terms and expects to close by the fourth quarter, planning to sell the technology as systems alongside its Instinct accelerators. The catch is that a chip built around one model is hard to reuse when the model changes.

Free and Go users move to GPT-5.6 Luna as their default, get unlimited text conversations, and gain a Think button that spends more reasoning on hard questions. OpenAI says Luna makes 62 percent fewer factual errors than the GPT-5.5 Instant model it replaces, measured on its own evaluations. Separate caps still apply to files, images, voice and image generation, and the rollout runs across this week and next. Paying subscribers get an updated GPT-5.6 Sol at the same time, on a service now serving roughly one billion weekly users.

Cloudflare OS lets non-developers describe an internal app in plain language and have an agent write and run it, with each workspace isolated and every reach into a company system passing through a permission layer the company calls Gatekeepers. Cloudflare deployed the first version to thousands of its own employees in May before releasing the code on GitHub under the Apache 2.0 license — a genuinely permissive license, though the README says the team is not seeking outside contributions. A hosted version is planned for the Cloudflare dashboard later.

Suno says it will add watermarking and fingerprinting so its output can be identified once it leaves the platform, plus a download policy meant to block mass uploads to streaming services, and it has licensed Musixmatch's Sentinel system to detect lyric copying. Asked whether the watermark would use Google's SynthID or something proprietary, the company declined to say — which matters, because an identifier only helps if the platforms receiving the music can read it. Suno is defending suits from Universal and Sony coordinated by the Recording Industry Association of America, and lost a copyright ruling to the German rights society GEMA in July.