Today on the Brief: Three Kinds Of Control

Google shipped Gemini Omni 1.1 Flash on August 27, and the headline everybody ran with is "40-second video." Slow that down, because if you repeat it the way it's being repeated, you'll be wrong.

A single clip is still 3 to 10 seconds, at 24 fps. That didn't change. What changed is scene extend: the model now reads up to 10 seconds of your previous footage instead of just the last frame, and adds 10 more seconds on top — repeatable to a 40-second cumulative cap. So it's 40 seconds cumulative, built ten seconds at a time. Not a 40-second generation. Same asterisk on 4K: that's an upscale, not a native 4K render.

Also new: first/last-frame interpolation, and up to 3 seconds of video reference.

Here's the part almost nobody led with — the 360p draft mode. Google says it's up to 60% faster and about a third the cost of 720p. Put real numbers on it: video output runs $17.50 per 1M tokens, and Google bills a different token rate per second by resolution — 1,931 / 5,792 / 8,688 / 17,376 tokens per second at 360p / 720p / 1080p / 4K. That works out to roughly 3¢ / 10¢ / 15¢ / 30¢ per second. Input is $1.50 per 1M; text output $9.00 per 1M.

Be honest about that math: those per-second dollar figures are derived from the token rates, not a price list Google published. The rates are Google's. The multiplication is mine.

The expensive part of AI video was never the final render — it was the eleven versions before it, burning full-price generations to discover the camera move was wrong. Now you storyboard the whole sequence at 3¢ a second and pay real money exactly once.

Tencent released Hy4 preview today — 770B total parameters, 49B active, a 1M-token context, weights on Hugging Face and ModelScope as tencent/Hy4-preview and tencent/Hy4-preview-FP8.

And the license is Apache 2.0. A real one — not a community license with a revenue clause, not research-only. I've spent two episodes telling you to read the license before you build on an open model, because Qwen's release wasn't Apache and plenty of people said it was. So let me be consistent: this one checks out, against Tencent's own README.

But "open source" and "you can use this" are not the same sentence. 770B parameters is not a laptop model, and it's not a good desktop model either — this is multi-GPU server hardware. The weights are free. The electricity and the hardware are not. Realistically you'll hit it through an API: it's on OpenRouter and Tencent Cloud TokenHub today at $0.834 in / $2.501 out per 1M, with $0.042 per 1M on cached input. WorkBuddy and CodeBuddy are free for two weeks from launch, and Hy3 free access is extended to September 30.

Two more caveats before you get excited. It's text-only — no multimodal in this drop. And Tencent calls it a preview with known issues, specifically over-thinking and over-verification. That's the vendor saying it out loud; on a per-token bill, second-guessing isn't a personality quirk, it's a line item.

The benchmark deserves a caution too. Tencent ran a blind comparison — 163 experts, 203 engineering tasks — scoring Hy4 at 2.99/4.00 against 2.92 for GLM-5.3 and 2.94 for Kimi K3. That's five hundredths of a point, on a test the vendor designed and ran with their own experts.

On August 26, Anthropic made Claude in Chrome generally available on every paid Claude plan — and switched on autonomous browser actions. In one sentence: it used to ask you to approve each action; now it doesn't. Back on the 21st I covered Anthropic taking computer use to GA and said the open question was whether it holds up on old software with no API. This is the sequel — and it's running unsupervised.

Instead of per-step approval there are three layers: training against a library of known attacks, probes that scan page content for prompt injections before Claude acts on it, and an approval classifier that validates each action against your original request.

The numbers Anthropic published: on their red-team evaluation, attacks against Claude Opus 5 succeeded 3.8% of the time with no safeguards and 0% with the full stack. Sonnet 5: 0%. Fable 5: 0.3%.

One correction worth making, because it's circulating. People are pairing that 0% against a 17.6% figure. 17.6% is a real number — but it's from the older Opus 4.5 evaluation, a different model generation and a different test. Putting them side by side makes the improvement look bigger than what Anthropic actually claimed this week. And either way, these are Anthropic's own evaluations of Anthropic's own product. Nobody outside has red-teamed this at scale yet.

Same day, Claude Cowork got its own built-in browser on the desktop app — and this is the distinction people are getting wrong. Claude in Chrome uses your browser, with your logins, already signed in. The Cowork browser is separate: you bring logins over site by site (from Chrome, Edge or Firefox on macOS; Firefox on Windows and Linux), and banking, email and SSO are excluded unless you explicitly opt in. Cowork's browser is on Enterprise now, with Pro, Max and Team rolling out this week; Linux is still beta. And Chrome means Chrome — not Edge, not Brave, not mobile.

Claude Code 2.1.248 shipped a --restricted mode. It removes the built-in tools that run commands and code, drops WebFetch unless you name it in --tools, keeps file tools inside the working directory, refuses bypassPermissions, and ignores user/project/local settings files. If you run Claude against a client's repo, that flag is your friend. (2.1.247 also added /claude-api cost-optimize and local-only feedback drafts.)

Nvidia and Hugging Face — status check. I led with this yesterday, so: still no confirmation from either company. The Information's $12.9B, Business Insider's "over $13B and not signed." Nothing moved in 24 hours. It's still a report, and you shouldn't make an irreversible decision on it.

OpenAI, Anthropic, Google, Microsoft and Amazon signed a joint letter calling for a "defensive surge" against AI-enabled cyberattacks — 116 signatories by CNBC's count, though outlets don't fully agree on the number. Named risks were hospitals, water systems and power. For creators specifically: there is no new disclosure rule and no store-policy change in this. It's a position, not a deadline.

Anthropic opened 10,000 Claude Team seats for scientists — standard seats free, premium $15/mo with 5× usage limits, price locked for a year. You need to be a principal investigator at an academic or nonprofit research institution. Skip it unless that's you.

Ant Group put out Ling-3.0-flash-Fin, a finance-tuned model (124B total / 5.1B active, 262K context), free on OpenRouter for a month. The weights are not out — they've said next week. So if the license is what you care about, there's nothing to read yet.

Here's what I'd actually do with this:

1. Switch your AI video iteration loop to 360p today and only upscale the take you're keeping. 3¢ to be wrong, 30¢ to be right — the cheapest workflow change available to you this week.
2. Say it correctly: 40 seconds cumulative, 10-second chunks, 4K is an upscale. Getting that right in public is free credibility.
3. If you were waiting on a permissive big open model, Hy4 is genuinely Apache 2.0 — just be honest about your hardware before you plan around it. For most of you it's an API, not a download.
4. If you try autonomous browsing, pick the most boring, lowest-stakes task you have. Reading and pulling before submitting and sending. Build the trust in that order.
5. Find the tool in your week with no API where your hours actually go. That's the one to point this at — not the tool you like, the tool you dread.

Here's the pattern across all three. Every one of them moved the cost of being wrong. Cheap drafts mean a bad take costs 3¢ instead of 30¢. Open weights mean a bad vendor decision doesn't strand you. An agent that clicks means a boring task doesn't cost you an afternoon.

None of them made the ceiling higher. The best video isn't better today. The smartest model isn't smarter today. What got cheaper is failing, and trying again.