← Back to Blog

August 15, 2026 · By JayyRedd

Google Just Made AI Watermarks Optional | Daily AI News Brief (Aug 15)

📺 Watch the video · Today's companion page

Google just made the visible watermark on Gemini images, video, and music optional — the invisible one stays. Alibaba dropped a two-point-four trillion parameter model as open weights, with a version small enough to run on your own machine. And DeepSeek is raising API prices by up to eleven hundred percent — starting tomorrow. I'm Jayy, this is your Daily AI News Brief, let's get into it.

Story 1 — Alibaba opens the weights on Qwen 3.8

Yesterday at three P.M. UTC, Alibaba's Qwen team released the open weights for Qwen three-point-eight Max and Qwen three-point-eight twenty-seven B on Hugging Face and ModelScope. Max is the headline number — two-point-four trillion parameters in a mixture-of-experts design, about ninety-five billion active on any given query, a one million token context window, natively multimodal for text, image, and video.

But Max is not the story for you. Almost nobody can run a two-point-four trillion parameter model. The story is the smaller one: twenty-seven-point-eight billion parameters, dense, multimodal, a two hundred sixty-two thousand token native context window — extendable to a million with YaRN scaling — and — this is the part that matters — Apache 2.0. That's a permissive license. Commercial use, no revenue clause, no gatekeeper.

And the momentum is real: Bloomberg reported today that the Qwen family has now passed three billion cumulative downloads across more than four hundred sixty open models — more than Google's and Meta's open models combined, per Hugging Face data.

Here's the creator angle. A dense twenty-eight billion parameter multimodal model is the size that actually fits on hardware you can buy. That means the class of AI work where you're paying per token today — bulk image tagging, transcript cleanup, first-pass draft generation, anything high-volume and low-stakes — has a credible zero-marginal-cost option as of yesterday.

So here's your move. If you've got a twenty-four gigabyte GPU or a Mac with thirty-two gigs or more: grab a quantized build of the twenty-seven B from Hugging Face, load it in LM Studio or Ollama, and run ten real samples of your highest-volume API task through it this weekend. Compare the output side-by-side. If it holds up, that line on your API bill goes to zero.

Story 2 — Google makes visible watermarks optional on Gemini

Rolling out over the next few days: Gemini has a new setting called Media Watermark that lets you turn off the visible sparkle on AI generations — images from Nano Banana, video from Omni, music from Lyria — in both the Gemini app and the Flow editor. Settings, Media Watermark, toggle off. That's it.

Two things stay embedded no matter what you choose: the invisible SynthID watermark and the C2PA provenance metadata. You cannot remove those. And the toggle isn't for everyone — work and school accounts don't get it at all, and in countries that require visible labels, like India, South Korea, and Vietnam, it's only available on the paid AI Ultra plan.

Two related moves worth knowing. Google also open-sourced Credentio, a C++ library that validates C2PA content credentials locally, on your machine — no sending files to a cloud endpoint. And Anthropic went the other direction this week: every new Claude model now watermarks its text worldwide — a statistical pattern in word choices you can't see but detectors can — plus C2PA metadata on generated files, to comply with the EU AI Act's transparency rules.

Creator angle: cleaner deliverables, starting now. Client work, social posts, marketing assets, portfolio pieces — no more cropping or covering the corner sparkle. But understand what didn't change: the provenance is still in the file, permanently. Turning off the visible mark is a presentation choice, not a disguise. Don't sell AI-generated work as hand-made — anyone with a validator can check.

Two things to do today. One: open Gemini, go to Settings, Media Watermark, and make the call for your own workflow. Two — and this is the one people will skip: if you deliver AI-generated work to clients, add one line to your contract or your delivery email that says how AI content is labeled. The platforms just made disclosure a setting. Make it your policy before a client asks.

Story 3 — Gemini 3.7 Flash lands at half price

Google shipped Gemini three-point-seven Flash on Wednesday, three weeks after three-point-six Flash, and called it their most intelligent workhorse model yet for coding and agents. On Google's own DeepSWE benchmark it scores sixty-five-point-three percent against forty-nine percent for the model it replaces. That's a sixteen point jump in three weeks.

The pricing is the aggressive part. Seventy-five cents per million input tokens, three seventy-five out, with the same roughly one million token context window. That's half the normal rate — and it's introductory. It expires December thirty-first, and on January first it goes to a dollar fifty and seven fifty.

Creator angle: you have four and a half months of half-price frontier-grade coding. So do this: write down the one internal tool you keep wishing existed — the automation, the agent that does the boring part of your workflow — and scope it this weekend while the discount is live. The cheapest window to build it is open right now, and it closes December thirty-first. Build it before January.

Story 4 — OpenAI previews Ultrafast mode

Also Wednesday: OpenAI previewed Ultrafast mode, a new API service tier that runs GPT-5.6 Sol at up to seven hundred fifty output tokens per second — up to fourteen times faster than standard processing. It's powered by Cerebras and their wafer-scale chips, which is the follow-through on a multi-year agreement the two signed in January to deploy up to seven hundred fifty megawatts of Cerebras inference systems through 2028.

Same model, same intelligence. Just fast.

Two caveats. It's a limited preview to a small set of customers, and OpenAI published no pricing and no general availability date.

Creator angle: speed is about to become a product category, not a spec. At seven hundred fifty tokens a second, a long-form response finishes before you can look away. Your action here is cheap: make a list of every place in your product or workflow where someone stares at a spinner waiting on a model. That's your upgrade list for the day this goes GA — and the people who have the list ready will ship the fast version first.

Story 5 — Apple builds its own model for China

Reuters reported yesterday that Apple has trained its own large language model for the Chinese market, with support from Alibaba. That's a real strategy reversal — the plan had been to license a domestic Chinese model to power Apple Intelligence there.

The detail that makes this notable: Apple is now the first foreign company approved by the Chinese government to offer its own proprietary AI model inside the country. Apple Intelligence is expected to launch in China in the coming months, after years of regulatory delay, and the competitive target is Huawei.

Creator angle: this is the clearest signal yet that "one AI product, one global market" is over. If any part of your audience is outside the US, the model under your product may be a different model for them, with different capabilities and different rules. Quick action: open your analytics and actually look at your geography split. If a meaningful slice of your audience or customers sits in a market with its own AI rules, that's a thing you plan for now — not a thing you discover at launch.

Quick Hits

Five fast ones.

That DeepSeek price change I flagged Friday turned out to be the small one. DeepSeek launched V4-Pro this week — stronger agent and cybersecurity performance, plus native OpenAI Responses API compatibility — and starting tomorrow, August sixteenth, the whole V4 API moves to peak and off-peak pricing with increases of up to eleven hundred percent. V4-Flash output goes from twenty-eight cents per million to a dollar thirty-two at peak — but half that off-peak. So two things tonight: re-quote any DeepSeek batch job at the new rates, and if the job can wait, schedule it into the off-peak window — that one scheduling change cuts the hike in half.

Anthropic reportedly cleared eleven-point-five billion dollars in second quarter revenue — a fourteen-fold jump from a year ago — and had its first profitable quarter on an adjusted basis, with an IPO targeted for this fall. That's the same window OpenAI is targeting: its public S-1 is expected within weeks, at a reported valuation between eight hundred fifty-two billion and a trillion.

Anthropic also published a risk report raising its catastrophic misalignment rating from very low to low — citing broader uncertainty, not new evidence of danger — and disclosed it's shelving a more capable internal model, called Model 2, with no release plans.

OpenAI told investors enterprise revenue has overtaken consumer ChatGPT for the first time, driven heavily by Codex. And Grok four-point-six is now live in GitHub Copilot and Cursor.

Last one. A Connecticut judge sanctioned a self-represented litigant for hiding white-on-white text in his court filings — invisible instructions telling any AI system that read the document to rule in his favor. It's the first documented prompt-injection attempt aimed at a U.S. court. It did not work.

Your Moves Today

Before I let you go — the brief is only useful if you do something with it, so here are your three moves for today, smallest first.

One: open Gemini, Settings, Media Watermark, and set it the way your work needs. Thirty seconds.

Two: if you use DeepSeek, re-quote your batch jobs at the new rates and shift what you can to off-peak — the hike lands tomorrow, so tonight is the deadline.

Three, the weekend project: download Qwen twenty-seven B, run ten samples of your most expensive recurring API task through it, and find out whether that bill is still worth paying. Pick one. Do it today.

The Takeaway

Join the Roundtable

Quick thing before you go. If you're building or monetizing with AI and you're doing it alone, I run a private community called the AI Creators Roundtable. Inside you get a vetted library of prompts, tools, and workflows that actually work — not the recycled listicle stuff. You get a real creator network with people who'll give you honest feedback on your work. Weekly live AMAs where you can bring your actual problems. And exclusive job and client leads we surface for members first. Founding member pricing is locked in for everyone who joins now.

Sources