Morning, {{first name | folks}}! A busy week in AI. Google showed a 10-minute AI film, new benchmarks put agents through real scientific work, Anthropic changed how it charges for blocked requests, PrismML put a 2B model on smart glasses, and Claude is taking on more of Anthropic’s own R&D.
Today’s Top 5
Google Demonstrates a 10-Minute AI-Generated Film: Google showed an AI system keeping characters, locations, and objects consistent across a 10-minute film.
A New Benchmark Tests AI Agents on Real Scientific Work: Terminal-Bench-Science tests agents across 70 scientific workflows, with Claude Opus 5 resolving 30% of tasks.
Anthropic Resumes Charging for Some Blocked Claude Requests: Anthropic will charge for blocked requests involving biology, distillation attacks, and frontier LLM development.
PrismML Puts a 2B AI Model on Smart Glasses: PrismML demonstrated a 2B vision model running locally on Snapdragon-powered glasses using 1-bit compression.
Anthropic Reveals How Much Claude Is Doing Inside Its AI Lab: Anthropic says Claude now leads 26% of its AI R&D work.
Google Research has demonstrated a new AI system that can generate a coherent 10-minute film, with characters, locations, and objects staying consistent across different scenes. The system uses multiple AI agents to plan the story, generate individual shots, track visual details, and review the footage as it is produced.
The result moves AI video beyond the short, disconnected clips that have dominated the technology so far. Google’s work is focused on one of the harder problems in longer-form generation: keeping the same characters, environments and story details intact as more scenes are added.
Artificial Analysis has added Terminal-Bench-Science, a benchmark covering 70 scientific workflows across life sciences, physical sciences, earth sciences, mathematics, and engineering. Instead of standard questions, AI agents must complete research tasks and produce verifiable outputs such as analyses, simulations, proofs, code, and data products.
The first results put Claude Opus 5 at 30% task resolution, followed by GPT-5.6 Sol at 22.4% and Claude Fable 5 at 21.4%. The benchmark is designed to be harder than standard Terminal-Bench, with every model tested seeing its resolution rate fall by more than 10 percentage points.
Anthropic is resuming charges for requests that its safeguards block before Claude responds, but only for requests in three categories: biology, distillation attacks, and frontier LLM development. It says the change is aimed at attacks on its systems and does not apply to blocked requests generally.
Anthropic says 99.7% of accounts using Claude Code, Claude.ai or Cowork did not hit one of these billable blocks in recent testing. The classifiers have a reported false-positive rate below 0.1%, and users can flag incorrect blocks through /feedback in Claude Code.
PrismML has demonstrated its 1-bit Bonsai model running locally on AI smart glasses powered by Qualcomm’s Snapdragon AR1 Gen 1 platform. The company says its new 2B vision-language model can fit 4× more parameters into the same memory footprint and process what the wearer is seeing in real time.
The demo puts a relatively large vision model on hardware with tight memory and power limits. By running the model locally, the glasses can process visual inputs without sending every request to the cloud, making on-device AI more practical for wearable devices.
Anthropic says Claude now leads 26% of its AI research and development work, meaning it can complete most of a task from a high-level prompt while a human supervises the result. More than 90% of Anthropic’s R&D now has Claude either leading or handling large parts of the work under close human direction.
Anthropic is publishing the figures as part of a new set of measurements for tracking the pace of AI development inside frontier labs. The company says Claude is not fully autonomous in any measured area but wants these metrics to show how quickly AI is becoming involved in building the next generation of AI systems.
Other AI Signals:
Akamai Signs $11.6 Billion Cloud Deal With Anthropic. The seven-year agreement could expand to about $20 billion as Anthropic adds more computing capacity.
Perplexity Cuts Search API Costs by 68% With New Engine. Its new Photon retrieval system brings the fast preset to 160 ms median latency while cutting estimated model-plus-search costs across six agentic benchmarks.
Huawei Builds a Bigger Challenge to Nvidia in AI Chips. Its Ascend chips, SuperPoD systems and large AI clusters are expanding Huawei’s position in AI computing, particularly in China.
OpenAI Agent Accesses Australian Government Health System. The agent accessed public and non-public files while researching medical spending, with the Australian government investigating the incident.
DeepSeek Reportedly Reaches a $1B Annual Revenue Run Rate. The Information estimates DeepSeek’s annualised revenue has more than doubled in recent months, helped by higher API prices and continued demand.
AI Tools to Try:
HyperProbe: lets coding agents debug production issues without requiring a new deployment.
PageIndex: lets you chat with long documents and get answers with page-level citations.
Suprbox: adds a policy-controlled security layer between AI agents and the systems they can access.
Jev: a lightweight decision-making model for AI agents that returns probability-based decisions quickly and cheaply, including whether an agent should retry or take its next action.
