Watch on YouTube · subscribe for the daily brief

The cheap ride is over and the invisibility is ending. DeepSeek just raised prices up to 1,100%, and Anthropic is watermarking every word Claude writes. The two things that defined AI's summer both reversed this weekend.

Saturday's list is short and sleepy — a Grok integration, some bird photos, a couple of security explainers. But hidden in it are two stories that mark a genuine turning point, because they're the exact reversal of the two forces that defined AI's entire summer: the race to the bottom on price, and the race to make AI output invisible. Both flipped this weekend.

Start with price, because the number is staggering. DeepSeek — the company whose whole identity was being the cheapest frontier-adjacent model on earth — is raising API prices by up to 1,100% with its V4 Pro launch, effective August 16-17. The specifics matter: peak-hour output tokens jump from $0.87 to $3.96 per million, more than four times; some cache-hit input tokens rise twelve-fold. And it comes with something the industry hasn't seen before — peak and off-peak pricing, 01:00-10:00 UTC billed at double the off-peak rate. Read what that actually is: AI inference is being priced like electricity now, by time of day, to smooth demand on strained infrastructure. DeepSeek even framed it as allocating resources "more reasonably." The company that started the price war just ended its own participation in it.

Why this is bigger than one vendor's rate card: for over a year, every roadmap, every business case, every "we'll just scale it later" assumption rested on one belief — AI gets cheaper over time. DeepSeek was the proof and the pressure that kept everyone else cutting. This weekend that belief cracked. And the picture is now genuinely split: the same weekend DeepSeek quadrupled prices, OpenAI cut its GPT-5.6 Luna rates. So the simple "AI gets cheaper forever" story is dead, replaced by something messier and more real — prices moving in both directions as each lab confronts the actual economics of serving these models at scale. This is the Tokenpocalypse maturing into a market, and the practical takeaway is blunt: re-baseline your model costs now, and separate your token types. A blended estimate will hide the difference between a 1.5x increase and a 12x one. If you run anything time-insensitive on DeepSeek, move it to off-peak. The era of not thinking about the meter is over.

The second reversal is quieter but philosophically sharper. Anthropic explained Claude's new text watermarking in detail (TechCrunch), and Sebastian Raschka published "Building an AI Text Detector From Scratch" the same day. Put those two together and notice the whiplash. All summer, the entire industry optimized to make AI-generated output indistinguishable from human work — that was the product, the selling point, the benchmark. Now the same industry is racing to make that output detectable again — watermarking every word Claude writes (C2PA provenance metadata, EU AI Act-driven), and independent engineers building detectors from scratch. We spent a year erasing the line between human and machine text, and now we're spending the next year frantically trying to redraw it. That's not a contradiction the industry stumbled into — it's the predictable consequence of succeeding too well at the first goal. Raschka's piece is the one to actually read, because it shows you how detection works under the hood, which means it also shows you how fragile it is: watermarks blur under editing, detectors have false positives, and no method is airtight. If you build anything that ingests or publishes text — content platforms, plagiarism tools, trust-and-safety systems — this is your field now, and it's an arms race, not a solved problem.

The two security explainers on your list fit the weekend's practical mood. TechCrunch's "how to spot and stop unauthorized access to your AI accounts" is a genuinely useful read after a month where the risk stories were all about breaches — this one's about your account, your API keys, your exposure, and it's worth sending to your whole team. And the Wired retrospective on OpenAI's Hugging Face crisis is the definitive long-form account of the incident that started this whole saga — if anyone on your team still hasn't grasped why "agent security" became the phrase of the summer, that's the piece to forward.

The builder's bench, light for a Saturday: Grok 4.6 landed in GitHub Copilot (xAI continuing its fast push into developer tooling — worth a look if you're a Copilot shop, given 4.6 matches GPT-5.6 Sol on the intelligence index at a fraction of the price). And Simon Willison, in his characteristic weekend mode, shipped "CORS Chat" (a small practical web-dev tool) and posted a Northern Gannet sighting — the reminder that even the people closest to this relentless news cycle still close the laptop and go look at birds. Not nothing, on a Saturday.

The through-line, one sentence: on August 15 the two forces that defined AI's summer both reversed — DeepSeek ended the price war by quadrupling its own rates and pricing tokens like electricity, while Anthropic and independent engineers raced to make AI text detectable again after a year spent making it invisible — which together mark the moment the industry stopped chasing "cheaper and indistinguishable" and started reckoning with the bill and the consequences.

One thing to do today: open your AI cost model and do two things. First, if DeepSeek is anywhere in your stack, re-run your numbers against the new peak/off-peak rates that hit this weekend — and while you're in there, separate cache-hit, cache-miss, and output tokens, because the increases range from 1.5x to 12x depending on type, and a single blended number will lie to you. Second, take the broader lesson: pick your most-used model and ask "what happens to my unit economics if this price doubles?" For a year that question was hypothetical. This weekend DeepSeek made it real. The teams that model cost per task by token type — and have a fallback provider ready — will absorb the next price move. The ones running on "AI keeps getting cheaper" just got their warning.

Sources covered

1 story · 8 sources (4 news, 1 company, 3 voices)

1. Anthropic explains Claude's new text watermarking in detail
   📰TechCrunch↗🎙️Simon Willison↗📰TechCrunch↗🎙️Sebastian Raschka↗🎙️Simon Willison↗📰Engadget↗📰Wired↗🏢xAI↗