A busy day in the world of AI as we take a step towards the future of off-world data centers. And what about those Claude watermarks? Will your report be flagged as AI after you turn it in? Check out the news today and if you need hourly AI News updates check out the SkyNet Tribune.
Claude’s new Scarlet Letter watermark is invisible—for now
Anthropic has revealed that it will soon watermark content that is processed (not just generated!) by any of its models. In a support article, Anthropic explained that it was rolling out machine-readable watermarks to comply with the European Union’s AI Act, which requires all AI system providers to watermark AI-generated or manipulated audio, image, text, and video outputs. The law applies to any AI model released after August 2 and provides a grace period until December 2026 for providers to update previously released models.
Anthropic confirmed that moving forward, all new models offered globally—not just in the EU—will mark AI-generated content “from day one.” Text outputs will “carry embedded watermarks,” invisible to the user, and other “generated files will include digitally signed provenance metadata where supported,” Anthropic said. [ArsTechina]
The big story: Gemini 3.7 Flash arrives at half price. Gemini 3.5 Pro still doesn’t have a date.
Google released Gemini 3.7 Flash yesterday, calling it “our most intelligent workhorse model yet for coding and agents.” It landed just three weeks after Gemini 3.6 Flash — a cadence that would be impressive in isolation, and is impossible to read in isolation.
Because the model Google didn’t talk about is Gemini 3.5 Pro, its flagship, which has now slipped past three release deadlines and still has no announced date (Bloomberg via The Star). Sundar Pichai told investors on the July earnings call that Google was aiming to ship faster and was already burning serious compute training Gemini 4. Shipping your fourth Flash variant of the summer while the frontier model stays dark is a particular kind of answer to that.
The numbers
The gains over 3.6 Flash are real and concentrated in code:
| Benchmark | 3.6 Flash | 3.7 Flash |
|---|---|---|
| DeepSWE v1.1 (long-horizon SWE) | 49.0% | 65.3% |
| FrontierCode 1.1 Main (production code) | 34.4% | 43.6% |
| WebDev Arena (Elo) | 1538 | 1588 |
| GDP.pdf (complex documents) | 22.0% | 34.0% |
| AutomationBench (business workflows) | 17.0% | 30.4% |
The DeepSWE jump — 49 to 65 in three weeks, on a mid-tier model — is the one worth staring at. Google attributes it to developer feedback plus “algorithmic innovations” it plans to carry into future models, which is as close as a launch post gets to saying this is a preview of what Pro will do.
The price is the strategy
$0.75 per million input tokens and $3.75 per million output — half what 3.6 Flash originally cost. Note the asterisk: that’s introductory pricing through December 31, 2026. On January 1 it doubles to $1.50 and $7.50.
So the pitch is a cheap, competent brain for agents that run long and call tools often, priced to win developer habits during the exact window when Google’s flagship can’t compete on capability. Whether developers who build on 3.7 Flash at 75 cents stick around at $1.50 is the actual question.
Google also folded 3.7 Flash into Gemini Spark, its 24/7 personal agent, for AI Pro and Ultra subscribers across 160+ countries — though notably not the EEA, UK, Switzerland, or Nigeria. Spark gets better Workspace tool use: consolidating files, drafting emails, updating status docs.
Safety notes
3.7 Flash ships with updated Frontier Safety safeguards against CBRN and cyber-offense misuse, consistent with DeepMind’s bioresilience approach and its cyber program. There’s a model card. Worth reading alongside the next item.
Also today
Data centers in space could be a new frontier for insurers — if they can price the risk. The race to build data centers is reshaping power grids, construction markets and corporate capital spending on Earth. Now, some of the biggest names in technology and space want to move part of that infrastructure into orbit.
This would launch a new frontier for insurers — if they can figure out how to price a risk that has never existed at scale.
SpaceX has laid out the most aggressive vision. In January, the company filed with the Federal Communications Commission for a constellation of up to 1 million satellites that could form an orbital artificial intelligence data center. CEO Elon Musk has argued that solar-powered computing in space could become cheaper than terrestrial data centers within two to three years as launch costs fall and the cost of adding power on Earth rises. [CNBC]
Hassabis is lobbying Washington for a FINRA for AI. In his new role as Alphabet chief scientist, Demis Hassabis has personally pitched Treasury Secretary Scott Bessent and OSTP Director Michael Kratsios on an independent standards body — industry-funded, government-overseen, staffed by technical experts — that frontier labs would submit models to up to 30 days before public release, for probing on cybersecurity, biology, and autonomous deception. Target: operational by end of 2026. The wrinkle is that Treasury has been developing its own version of roughly the same institution in parallel (Fortune, Axios).
DeepMind’s identity crisis is now the story, not the reshuffle. Fortune reported yesterday that engineers see Koray Kavukcuoglu’s elevation as accelerating a slow-motion transfer of power from London to Mountain View, and worry the firewall between DeepMind and Google is dissolving. Google disputes the framing, saying London remains central and research autonomy is intact. Add the Pentagon contract revolt — resignations plus a union drive — and the reorg looks less like a leadership change than a symptom. Kavukcuoglu now owns Gemini development, frontier research, and the app and developer teams, reporting to Pichai (CNBC).
Sign language translation shipped to phones on Tuesday. DeepMind’s SL2T model now powers sign-to-text dictation in Gboard and Live Transcribe on the Pixel 11, ASL to English. Trained on 100,000+ hours across 50+ sign languages; it sees pose landmarks rather than video, so the camera feed never leaves the device. Scores 70 BLEURT zero-shot on FLEURS-ASL. First time sign language AI has shipped in a consumer product at this scale.
WeatherNext got open-sourced. A Nature paper on Aug 6 showed WeatherNext Cyclones hitting state-of-the-art on cyclone track, intensity, and wind structure at once — three-day forecasts as good as prior two-day ones, roughly a decade of meteorological progress. Code and weights are on GitHub.
Sources linked inline. Corrections welcome.
Leave a Reply