The frontier race just reopened. Meta, Google, OpenAI and Anthropic all pushed new models within days of each other, and the fight has shifted from raw intelligence to cost per task and how much control the models get. At the same time, agents are gaining the ability to run in the background and act on their own, which is forcing fresh questions about safety.

Meta released Muse Spark 1.3, scoring 62 on the Intelligence Index and trailing only the top-tier models while costing far less to run. It is free on OpenCode, and Mark Zuckerberg hinted at a larger model codenamed Watermelon with plans to release the weights. Google answered with Gemini 3.8 Flash, its third Flash model in six weeks, holding pricing steady while improving coding and reasoning. DeepMind admitted Gemini still lags the frontier.

Why it matters: The competition is now about cost-effectiveness and accuracy, not just intelligence, which lowers the price of serious AI for businesses.

OpenAI has released Astra, the first of its models to hit a Critical cybersecurity capability under its Preparedness Framework. In testing, Astra found unknown security flaws, turned them into working exploits, and outperformed the previous model at spotting vulnerabilities. OpenAI has added stronger safeguards, restricted its most advanced cyber features to select testers, and reports Astra refused 91.5% of harmful cyber prompts. It reached trusted partners and OpenAI's cyber early-access group on 3 September, with wider access following.

Why it matters: A model capable enough to find and weaponise real security flaws forces both stronger controls and a rethink of who gets access.

Anthropic added a feature letting Claude Cowork and Claude Code run tasks in the background on macOS, so users can keep working while Claude clicks, types and manages apps. It is in beta for Pro and Max subscribers and only runs with user-approved apps.

Why it matters: Agents that act on their own in the background are more useful but harder to supervise, sharpening the tension between autonomy and oversight.

Source: Anthropic

Senator Bernie Sanders published an op-ed urging a worldwide halt on building more powerful AI models, pointing to industry admissions that the technology could be escaping their control. He warned of rising energy bills, environmental costs and tens of millions of job losses, and floated a Cold War style AI deal between the US and China. Critics noted the piece does not explain how such a pause would be enforced. He has since announced legislation with Congressman Greg Casar, the Ban Artificial Superintelligence Act.

Why it is interesting: It signals that AI safety is moving from research labs into mainstream political debate.

Source: Office of Senator Bernie Sanders

During a live stress test, Claude Fable 5.1 built browser games, installed Blender tooling and debugged its own work faster than earlier versions. One striking moment came when it made an unprompted style decision inside Blender. The test quickly shifted from asking whether the model could do the task to debating how much it should be allowed to decide on its own.

Why it is interesting: It shows advanced agents starting to make choices no one asked for, which is both impressive and a warning.

Source: The Neuron

Dyson unveiled the CameraJet, an electric toothbrush with a built-in AI camera and algorithm that scans for gaps between teeth and automatically delivers targeted jets of mouthwash. Priced at £419.99, it marks a push into luxury AI-assisted dental care.

Why it is interesting: It is a clear example of AI moving off the screen and into everyday physical products.

Context: Use this with an agentic coding assistant such as ChatGPT Codex to automate a repetitive office task without needing developer skills.

Prompt: Act as my operations assistant. Review the attached spreadsheet, clean and standardise the data, fix formatting and errors, flag anything that looks inconsistent, then produce a short summary of the three most important insights. Ask me before making any irreversible changes.

Example use case: A manager drops in a messy sales export each morning and gets back a tidy file plus a clear summary of what changed and what needs attention.

The US Department of Defense has made secure AI tools available to up to three million staff through its GenAI.mil platform, which has already drawn more than 1.7 million users. ChatGPT Mil and Grok for Government handle unclassified work like logistics, policy and admin. Anthropic's Claude is absent amid a dispute over how the model can be used.

Why it is important: It shows large institutions moving to embed AI in everyday operations at massive scale.

New York City Public Schools imposed a one-year ban on student-facing generative AI tools for grades 2-K through 8, affecting nearly 600,000 students. High schoolers will still get AI literacy lessons and limited access to approved tools. Teachers can use authorised platforms for lesson planning and admin.

Why it is important: It reflects a growing push to delay chatbot access for children while adults debate where AI belongs in education.

Nvidia's revenue jumped from $7bn to $96bn in three years and it is projected to hold 75% of the AI accelerator market by 2026. But chip startups have raised $8.3bn combined, OpenAI and Anthropic are building or hiring for their own hardware, and Google, Amazon and Microsoft keep upgrading proprietary chips. China is developing its own accelerators to cut reliance on Nvidia.

Why it is important: The hardware layer that underpins the entire AI boom is starting to draw serious challengers, which could reshape costs across the industry.

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas