Pick of the Week
Anthropic#Research
Anthropic used an unreleased research build of Claude to push a Riemann Hypothesis-related lower bound from 41.6% to 67.2%
This isn't the "we solved the Riemann Hypothesis" kind of overclaim, it's incremental progress on a specific, known mathematical lower bound, backed by a formally verifiable proof rather than the model's say-so. The same week Anthropic was busy with watermarking and earnings, the kind of conservative stuff, its frontier research line clearly hasn't slowed down.
Full story
Anthropic's research team had an unreleased research version of Claude work on the known lower bound for the proportion of Riemann zeta function zeros satisfying the Riemann Hypothesis, and pushed it from 41.6% to 67.2%. The work ran across two Claude Code sessions totaling 31 million output tokens: the first attempt produced 650 failed approaches before being told to start over, the second coordinated roughly 60 Claude subagents, ran 2,400 shell commands and hundreds of Python scripts, ran thousands of numerical verifications, and pulled from 54 arXiv papers. The method builds on existing results from several mathematicians, and the resulting proof can be formally verified in Lean. It was also reviewed by two Anthropic mathematicians and two external mathematicians, Brian Conrey and Dan Goldston.
Anthropic#Product#Security
Claude's text output now ships with a built-in watermark, to comply with EU rules
This is a rare case of one of the big three proactively building "detectability" into the product, and the timing lines up with EU AI Act requirements. Not everyone's buying it: Daring Fireball called it "a perversion of writing," worth reading both sides.
Full story
Anthropic announced that Claude's text output will carry a built-in watermark going forward, built on Google DeepMind's SynthID-Text: among several "equally reasonable" candidate words, a key plus the preceding context decides which one gets picked, making the text statistically detectable as AI-generated while looking identical to the human eye. Anthropic says internal testing found no impact on content quality or readability, no added token count, and no way to trace text back to a specific user. The limits are real, though: short, fact-heavy passages (where word choice is already narrow) detect poorly, code gets almost no watermarking, and text that's been human-edited may carry too weak a signal.
Qwen#OpenSource#Engineering
Qwen3.8-27B goes open source claiming it beats Claude Opus 4.6 Max, but hands-on testing finds the default reasoning "overthinks"
A good official benchmark score is one thing, whether the default settings actually make sense in practice is another. This gap between "great scores, questionable defaults" is worth noting more than the raw scores themselves, and it's a reminder that when a Chinese open-source model's leaderboard numbers show up, it's worth waiting to see if anyone's actually put it through real use first.
Full story
Alibaba's Qwen team open-sourced Qwen3.8-27B, 27 billion parameters, with native support for vision, PDF, chart, and video understanding, a native context window of 262K tokens expandable to 1 million. In Qwen's own benchmark comparisons, it leads Claude Opus 4.6 Max by 8.3 points on SWE-bench Pro and 2.5 points on CoWorkBench, and it runs on a consumer 24GB GPU. But when Simon Willison tested the same SVG generation task, the default "xhigh" reasoning setting took 21 minutes and burned through over 22,000 reasoning tokens; with reasoning turned off, the same task took 137 seconds.
DeepSeek#OpenSource#Business
DeepSeek V4-Pro goes GA, and the API splits pricing into peak and off-peak
Going from "we're raising prices" to "GA plus a peak/off-peak split" within a week shows DeepSeek is moving fast on revenue, not just talking about it. Peak/off-peak pricing is a familiar cloud-computing move, but it's not common yet for LLM APIs. Worth watching whether others follow.
Full story
DeepSeek moved V4-Pro (version V4-Pro-0813), the model it warned about a price hike for just last week, from preview to general availability across app, web, and API. The update strengthens coding agents, tool calling, and long task chains, and adds an OpenAI-compatible Responses API, with context up to 1 million tokens. Starting 8/17, API pricing splits into peak hours (9am-12pm and 2pm-6pm Beijing time) and off-peak, with peak output pricing hitting 27 RMB per million tokens, half that off-peak.
Google#Research
Google research: LLMs actually "know" most facts, they just fail to retrieve them
This is a pretty practical clue for how to fix AI hallucination. If the problem is mostly retrieval and not knowledge volume, then throwing more pretraining data at it may not be the most effective fix. Figuring out how to actually surface what's already been learned might matter more.
Full story
Google's research team tested frontier models including Gemini-3-Pro and GPT-5, and found that 95-98% of facts are already encoded in the model's parameters, but direct recall fails 26-34% of the time, and even with thinking mode on, 11-12% still can't be pulled out. In other words, most factual errors look more like "can't find the key" than "the book isn't on the shelf at all." Chain-of-thought reasoning recovers 40-65% of these "encoded but unretrievable" cases. The team also released a new benchmark, WikiProfile, containing 2,150 facts sourced from Wikipedia.
Nvidia · OpenAI#Business
Nvidia scales back its financial guarantee on OpenAI's data centers, Stratechery draws a parallel to the railroad bond boom
When even Nvidia is tightening the financial guarantees it gives downstream customers, and analysts are starting to reach for historical financing-bubble comparisons, the "AI infrastructure built on debt" narrative has looked shakier these past few weeks. Whether that's a false alarm, I honestly don't have a firm view yet.
Full story
Reuters reports Nvidia significantly scaled back a roughly $250 billion financial guarantee it had committed toward OpenAI's data center buildout (the report doesn't give an exact figure for the reduced amount). The same week, a Stratechery analysis compared the entire AI infrastructure investment boom to the 1870s railroad bond boom: projected 2026 AI spending sits around $600 billion, big tech has already issued $194 billion in debt this year, and Nvidia has lined up financial partners like Apollo and BlackRock to mobilize $500 billion for infrastructure financing. The piece worries this is shifting risk onto long-horizon money like insurance float and pension funds.
Anthropic#Business
Leaked Anthropic earnings: quarterly revenue up 1,400% year over year, possible October IPO at a $2 trillion valuation
Private company financials are opaque by nature, so how much to trust these numbers is a fair question. But the growth rate and IPO valuation rumors together give a rough sense of just how high the market's expectations for Anthropic are right now, and it's an interesting pairing with this week's "AI backlash is a trust crisis" story.
Full story
According to a report cited by 量子位, Anthropic's Q2 revenue hit $11.5 billion, up 1,400% from $787 million in the same quarter last year and up 143% from Q1's $4.73 billion. This quarter also marked the first time "adjusted operating profit" turned positive (no amount disclosed), though gross margin guidance was revised down from 50% to 40%. Separately, per an FT report, multiple investors expect Anthropic's October IPO to value the company above $2 trillion, potentially the largest IPO ever, with annualized revenue projected to reach $100-120 billion by year-end (versus the $47 billion figure disclosed in May). All of these numbers are leaked, not confirmed by official Anthropic financials.
Stripe · OpenRouter#Business
Stripe reportedly acquiring AI gateway startup OpenRouter for over $7 billion
A payments company buying a model-routing company signals that the "AI gateway" layer of infrastructure has gotten valuable enough that a player at Stripe's scale wants to just buy it outright, not merely plug into the APIs themselves. Going from $1.3 billion to $7 billion in three months is itself a valuation jump worth noting.
Full story
According to TechCrunch and Bloomberg, Stripe is about to acquire OpenRouter for over $7 billion. The startup lets developers switch between 400-plus models based on need and budget, and has 8 million users globally. OpenRouter's Series B in May valued it at just $1.3 billion on a $113 million raise, with investors including Sequoia, a16z, Menlo Ventures, and Alphabet's CapitalG. Acquisition rumors first surfaced in July, and by August word was it's a done deal.