Good morning, folks. Lots are moving today, money changing hands in AI, cheap AI getting expensive, and a machine that's now doing 82% of what human researchers took months to do. Let's get into it.
Today's Top 5
Stripe Wants a Piece of Every AI Model Call: Stripe's reportedly buying OpenRouter, the gateway devs use to switch between hundreds of models, for $7B+. Three months ago it was worth $1.3B.
DeepSeek's Biggest Selling Point Just Took a Hit: Launched V4 Pro, then hiked API prices up to 1,100% starting August 17. The first real test of whether people stick around once cheap isn't guaranteed anymore.
Prime Intellect's Fable 5 Pushes AI Closer to Human-Level Research: 100+ autonomous runs, no internet, just raw compute and time. Fable 5 closed 82% of the gap to a record that took humans months, and it ran the whole loop itself, not just suggested ideas.
Zuckerberg's Biggest Bet Isn't Building the AI. It's Getting You to Trust It.: Wants billions to have one AI that knows their goals across every device. The hard part was never the tech; it's getting people to hand over that much access.
Google Wants AI to Work Without Seeing Your Data: Open-sourced HEIR, which lets AI process encrypted data without ever decrypting it. Great for health or financial data, still expensive to actually run.
Stripe is reportedly buying OpenRouter for more than $7 billion, according to Bloomberg. OpenRouter is the AI gateway that lets developers switch between hundreds of models without building around one provider.
The deal would be a huge jump from OpenRouter's $1.3 billion valuation just three months ago. The Wall Street Journal first reported the talks last month, while Stripe still says it doesn't comment on rumors or speculation. Stripe has been building deeper into the AI stack, and buying the layer developers use to access models would give it a pretty interesting position.
DeepSeek's Biggest Selling Point Just Took a Hit
DeepSeek launched V4 Pro, a serious upgrade for coding agents and complex tasks, built on a 1.6-trillion-parameter architecture with a 1 million-token context window. The same day, they announced API prices jumping 50% to 1,100%, starting August 17, with new peak and off-peak pricing, so timing now actually matters.
This isn't just DeepSeek getting greedy, though; they're racing to keep up with compute demand while chasing an $8 billion funding round at a $74 billion valuation. Their whole reputation was built on being the cheap alternative to OpenAI and Anthropic. This is the first real test of whether developers stick around once cheap isn't guaranteed anymore.
Prime Intellect ran 100+ autonomous research runs across 10+ frontier models, no internet, just 8xH200s and up to eight days each. The best result, from Fable 5, closed about 82% of the gap to a record humans took months to build. The models weren't just suggesting ideas either; they picked what to test, ran it, checked their own results, and tried again.
It's a narrow benchmark, tightly boxed in on purpose. But it's a real glimpse of AI actually running parts of the research process, not just assisting with it.
Zuckerberg's Biggest Bet Isn't Building the AI. It's Getting You to Trust It.
Mark Zuckerberg wants billions of people to have one AI that knows their goals and follows them across every device. In an essay this month, he compared the privacy bar to WhatsApp's encryption, even floating a future where Meta itself can't see your data. Meta AI's already at 1 billion monthly users, and he's betting WhatsApp and Messenger become where these agents actually live.
The hard part isn't building it, though; it's trust. Handing an AI your finances and relationships is a much bigger ask than a chatbot answering questions, and no demo can fake that.
Google Wants AI to Work Without Seeing Your Data
Google just open-sourced HEIR, a toolkit for building AI systems that can work with encrypted data without decrypting it first. It uses homomorphic encryption, which means a cloud service can process private data while the underlying information stays locked.
The idea is useful for things like health or financial data, where sending raw information to an AI system is a non-starter. The catch is speed: homomorphic encryption is still expensive to run, and HEIR is Google's attempt to make it easier and more practical.
Other AI Signals For Today:
AMD now supports Alibaba’s Qwen3.8 27B on Ryzen AI Max+ PCs and Radeon GPUs, with speeds reaching 24.5 tokens/sec on the Ryzen AI Max+ 395 and 51.8 on the Radeon AI PRO R9700. Larger models are becoming easier to run locally instead of relying entirely on the cloud.
China could have another major export on its hands: open AI models. Cheaper models that can be downloaded, modified, and run locally are spreading quickly, especially in emerging markets. The bigger play is influence; the more countries build on Chinese models, the more China can shape the standards and rules around AI.
Anthropic CEO Dario Amodei thinks the AI backlash comes down to trust. People have heard plenty of big promises, but they want to see what AI actually delivers. His point: results will build more trust than another prediction about what AI might do.
Vietnam has launched a global network connecting Vietnamese AI researchers, experts, and entrepreneurs at home and abroad. The goal is to turn that talent into research, training, and commercial projects across areas like healthcare, education, cybersecurity, and smart cities.
Today’s AI Tools to Try:
DeepSeek Harness: DeepSeek’s open-source agent runtime where models, tools, memory, sandboxes, and workflows can all be swapped as plugins.
OpenMotion: A free Mac motion-design studio that turns prompts into editable product videos, with a real canvas and timeline for refining the final animation.
Click: An MCP that brings live data from social, finance, marketplaces, and other sources into ChatGPT and Claude.
Nube.sh: Rent and deploy affordable cloud servers for websites, apps, APIs, and AI workloads.
