Google is back in the frontier conversation with Gemini 4 Argon, a model built for very long jobs and topping several benchmarks, though access is still gated. OpenAI used its DevDay to turn ChatGPT into an always-on agent platform with a flurry of launches. And Anthropic's leaked IPO paperwork lays bare the brutal economics of the AI race.

Google unveiled Gemini 4 Argon, which outperforms rival top models on 13 of 19 benchmarks and debuted at number one on a text leaderboard. Its headline feature is output of up to one million tokens in a single response, built for long, multi-step tasks like large code migrations and audits. Access is limited at first to vetted cybersecurity partners, with wider release and higher prices to follow. Pricing is set at $2 per million input tokens and $10 per million output tokens.

Why it matters: After a rough stretch, Google is signalling it can compete at the top again, which reshapes where enterprises place their bets.

At DevDay OpenAI shipped more than 20 launches, led by always-on agents called dots that run in the cloud on GPT-6 Astra and connect to over 4,000 apps. It also released GPT-6.1 Sol at $2 and $10 per million tokens with near-flagship performance, a Decisions API for fast routing, and an Ultrafast speed tier. A new $500 per month Pro plan unlocks the fastest speeds. OpenAI framed the push as making ChatGPT a universal identity and coordination layer across tools.

Why it matters: The battle is shifting from raw model power to always-on agents that do work between conversations.

Source: OpenAI

A leaked prospectus shows Anthropic projecting about $4.6B in 2025 revenue alongside roughly $42B in net losses and $518B in future cloud commitments. The company is reportedly targeting a valuation above $2 trillion. Nearly a quarter of revenue comes from just two customers, and about a third of the filing is given over to risks including manipulation and self-preserving behaviour. The seven co-founders would keep majority voting control after going public.

Why it matters: The numbers show how much capital frontier AI now burns, and how concentrated the revenue and the control remain.

Dyna Robotics posted about an hour of unedited footage of a semi-humanoid robot named Taku folding hotel laundry. The video keeps every stumble, awkward landing and correction rather than cutting to a highlight reel. It shows just how many small problems a robot must solve to handle towels and sheets in a messy real setting.

Why it is interesting: Unscripted footage is a more honest measure of robot dexterity than polished demos.

In Singapore a man was charged after allegedly sharing an AI-generated image of a crocodile he claimed was spotted at Pandan Reservoir. The hoax suspended rowing and canoeing for two days while authorities searched and found nothing. Spreading false information there can carry up to three years in prison and a fine.

Why it is interesting: It shows how a single fake image can trigger real disruption and serious legal consequences.

A tutorial trend shows how to build a 1990s sitcom style AI character using image tools and video generation. Creators design a consistent anthropomorphic cat, set a cluttered retro backdrop and animate short comedic sketches with simple motion prompts. The finished clips get title cards, film grain and audio before posting to social feeds.

Why it is interesting: It is a clear example of how quickly casual creators can assemble polished, shareable AI video.

Context: Before a model tackles a complex task, having it restate your goal in its own words exposes misunderstandings early and saves wasted time and tokens.

Prompt: Before you start, restate my goal and the constraints in your own words. List any assumptions you are making and any part of my request that is unclear. Do not begin the task until I confirm your understanding is correct.

Example use case: A developer hands the model a multi-step data migration brief and catches a wrong assumption about the file format before any code is written.

Anthropic, OpenAI, Google, Meta, xAI and Nvidia signed a voluntary agreement to monitor their models, build internal safety teams and submit to outside audits. The pact carries no legal penalties but could pave the way to formal rules. It reflects a preference for industry-led oversight over broader government mandates.

Why it is important: It sets the tone for how AI will be governed, and critics note there is no enforcement if a firm breaks its pledge.

Source: Reuters

Meta reportedly labelled new AI data centres as experimental facilities to claim federal research tax credits, cutting its 2025 tax bill to $3.9B from $700 million in 2023. Its own accountants warned the approach could be legally risky. The company pursued it anyway.

Why it is important: It shows how aggressively AI firms are structuring their huge infrastructure spend, and invites regulatory scrutiny.

Anthropic warned that an open model, Z.ai's GLM-5.3, can produce working cyberattacks, with dangerous capabilities recoverable through simple bypasses. Google separately reported that disclosed software vulnerabilities roughly doubled to 10,740 in August alongside AI-assisted hacking. OpenAI also traced a reasoning-replication effort to Moonshot AI.

Why it is important: Open and capable models lower the barrier to real-world attacks, raising the stakes for security teams.

Source: Anthropic

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