Today on the Brief: Watermarks Go Optional

On August 14 the Qwen team released open weights for both Qwen3.8-Max and Qwen3.8-27B on Hugging Face and ModelScope. Max is the headline: a 2.4-trillion-parameter mixture-of-experts model activating roughly 95 billion parameters per query, natively multimodal across text, image, and video, with a 1-million-token context window.

The more consequential release is the smaller one. Qwen3.8-27B is a dense 27.78-billion-parameter multimodal model with a 262,144-token native context window (extendable to 1M via YaRN scaling), shipped under Apache 2.0 — a permissive license with no revenue clause and no gatekeeper. Bloomberg also reported today that the Qwen family has passed 3 billion cumulative downloads across 460+ open models — more than Google's and Meta's open models combined, per Hugging Face data.

Rolling out over the next few days: a new Settings > Media Watermark toggle in the Gemini app and the Flow editor lets you turn off the visible "sparkle" watermark on AI generations — images from Nano Banana, video from Omni, and music from Lyria. The invisible SynthID watermark and C2PA provenance metadata always stay embedded and cannot be removed.

The toggle isn't universal: work and school accounts don't get it, and in countries that require visible labels — India, South Korea, Vietnam — it's available only on the paid AI Ultra plan. Google also open-sourced Credentio, a C++ library for validating C2PA content credentials locally, with no cloud round-trip. Meanwhile Anthropic went the other way: all new Claude models now watermark their text worldwide (a statistical pattern in word choice) plus C2PA on files, under the EU AI Act's transparency rules.

Google shipped Gemini 3.7 Flash on August 13, just three weeks after 3.6 Flash, calling it its most intelligent workhorse model yet for coding and agents. On Google's own DeepSWE v1.1 evaluation it scores 65.3% against 49.0% for the model it replaces.

Pricing is the aggressive part: $0.75 per million input tokens and $3.75 per million output, with the same ~1,048,576-token context window. That rate is introductory and expires December 31, 2026 — on January 1 it becomes $1.50 and $7.50. Access is live through the Gemini API, Google AI Studio, Android Studio, Google Antigravity, and the Gemini Enterprise Agent Platform.

Also on August 13, OpenAI previewed Ultrafast mode — a new API service tier running GPT-5.6 Sol at up to 750 output tokens per second, up to 14× faster than Standard processing. It runs on Cerebras wafer-scale hardware, the follow-through on a multi-year agreement the two companies signed in January to deploy up to 750 megawatts of Cerebras inference systems through 2028.

Same model, same intelligence — just dramatically faster. Two caveats: it's a limited preview open to a small set of customers, and OpenAI published neither pricing nor a general-availability date.

Reuters reported on August 14 that Apple has trained its own large language model for the Chinese market with support from Alibaba — a reversal of its plan to license a domestic Chinese model to power Apple Intelligence there.

The notable detail: 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, with Huawei as the competitive target.

DeepSeek launched V4-Pro — stronger agent and cybersecurity performance, plus native OpenAI Responses API support — and raises V4 API prices by up to 1,100% starting August 16 (16:00 UTC), moving to peak/off-peak billing. V4-Flash output goes from a flat $0.28/M to $1.32/M at peak ($0.66 off-peak); V4-Pro output goes from $0.87/M to $3.96/M at peak. Re-check your batch jobs.

Anthropic reportedly cleared $11.5B+ in Q2 revenue — a 14-fold jump year over year and its first profitable quarter on an adjusted basis — ahead of a fall IPO. It also raised its catastrophic-misalignment risk rating from "very low" to "low" (citing broader uncertainty, not new evidence of danger) and disclosed it is shelving a more capable internal "Model 2" with no release plans. OpenAI told investors enterprise revenue has overtaken consumer ChatGPT for the first time, driven heavily by Codex; its public S-1 is expected within weeks at a reported $852B–$1T valuation.

Grok 4.6 is now live in GitHub Copilot and Cursor. And a Connecticut judge sanctioned a self-represented litigant for hiding white-on-white prompt-injection text in court filings — instructions telling any AI reading 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.

1. Thirty seconds: Gemini → Settings → Media Watermark. Set it the way your work needs, and add one line to your client contracts saying how AI content is labeled.

2. Tonight (deadline is real): Using DeepSeek? Re-quote your batch jobs at the new rates and move what you can to off-peak before the Aug 16 (16:00 UTC) hike.

3. This weekend: Run 10 samples of your most expensive recurring API task through Qwen 3.8 27B locally (LM Studio or Ollama). Find out whether that bill is still worth paying.

Ongoing: Scope the internal tool you keep wishing existed while Gemini 3.7 Flash is half price (through Dec 31), and keep a list of every "spinner moment" in your workflow for the day ultra-fast inference goes GA.