Anthropic's Fable 5 Is Back, Etched Pockets $800M, and Gemini's Image Tool Goes Free
July started with a bang. The biggest headline this morning — and I mean the kind that made me refresh my feeds three times to make sure I wasn't reading old news — is that Anthropic's Claude Fable 5 is coming back online today. And that's just the appetizer.
The Fable 5 Rollercoaster Is Over (For Now)
If you've been following this saga, here's the short version: on June 12, the Trump administration slapped export controls on Anthropic's Fable 5 and Mythos 5 models, citing national security concerns about the models falling into foreign hands. Anthropic responded by… pulling the plug globally. Not just limiting access in certain regions — they straight-up disabled Fable 5 for everyone, everywhere, because they couldn't reliably filter users by geography in real time.
That's two and a half weeks of silence from one of the most capable publicly available models out there. And yesterday evening, the Commerce Department finally lifted those controls on both Fable 5 and Mythos 5. Anthropic says the models will be back online today.
A lot of people are wondering what changed behind closed doors. The public narrative from Reuters, Politico, and WSJ all point to White House-level negotiations that wrapped up late Tuesday. My guess? Anthropic probably agreed to some monitoring or usage auditing framework that satisfied the administration's concerns without requiring a full geo-block. Pure speculation on my part, but the speed of the reversal suggests a deal was struck fast once both sides got serious.
Now, the honest take: this whole episode raises a question that's not going away. If the US government can shut down a frontier model overnight over export rules, what happens when the next administration has a different set of concerns? The fragility of "available to everyone" suddenly feels very real. Fable 5 users — especially developers who'd built workflows around it — got a brutal reminder that cloud models are ultimately someone else's infrastructure, subject to someone else's rules.
Etched Comes Out of Stealth: $800M, $1B in Contracts, Working Chip
On the hardware side, Etched — the AI chip startup that's been building in relative quiet — just dropped a massive announcement. They raised $800 million from investors including Jane Street and a TSMC-linked venture fund. They've also signed over $1 billion in customer contracts and have a working chip ready for summer shipments.
The pitch is straightforward but aggressive: instead of building a general-purpose AI accelerator like Nvidia's H100 or B200, Etched's chip is hardwired for the transformer architecture — the "T" in GPT. By locking the silicon to one architecture, they claim much better inference economics per watt compared to Nvidia's offerings.
To be fair, this is a bet-the-company strategy. If the industry moves beyond transformers — say, if state-space models or some other architecture gains traction — Etched's custom silicon becomes very expensive scrap. But for right now, transformers run everything from Claude to Gemini to Llama, so the bet isn't unreasonable.
The $1 billion in pre-revenue contracts suggests serious customers are already convinced. We don't know who those customers are yet, but at that scale, we're probably looking at hyperscalers or major AI labs. Summer shipments means we'll have real benchmarks in a few months — that's when we'll know if the economics actually deliver.
Hardware folks running inference workloads should keep an eye on this. If Etched's numbers hold up, it could meaningfully change the cost equation for hosting open-source models.
Google Makes Gemini's Personalized Image Gen Free
On the consumer side, Google dropped a genuinely user-friendly update yesterday: Gemini's personalized image generation — previously a paid feature — is now free for all eligible US users. The system uses what Google calls "Personal Intelligence" to pull context from your connected Google apps (Gmail, Photos, YouTube, Search) and generate images that actually reflect your interests and data.
The rendering engine is called Nano Banana, which is honestly one of the better AI feature names I've heard in a while. But the privacy angle is worth a closer look. Google says this is opt-in and permission-based, which is good. Still, I'd be lying if I said I wasn't a little uneasy about an AI that comb through my Photos library to generate art. The feature is neat on paper — imagine asking Gemini to "make a birthday card using photos from last summer's trip" and getting something decent back — but the data exposure tradeoff is real.
This move also tells us something about Google's AI strategy right now. They're making consumer-facing gen AI free to capture usage data and habits, even if the direct revenue isn't there yet. The personalized angle is a differentiator that OpenAI and Anthropic can't easily match right now because they don't have the same personal data ecosystem.
Trajeckt: An Open-Source Firewall for AI Agents
Shifting gears to the security side — there's a new open-source tool on GitHub called Trajeckt that's getting some attention on HN. It's a runtime enforcement gateway for AI agents, written in Rust, that blocks multi-step tool-call exploits in about 1.6 milliseconds.
The key insight behind Trajeckt is that most per-action security checks are vulnerable to what they call "multi-step data exfiltration sequences." If each individual tool call looks harmless, but the sequence of calls across multiple steps leaks sensitive data, conventional guards don't catch it. Trajeckt enforces sealed pre-session commitments — basically, the agent commits to a policy upfront, and the runtime checks the full call chain against that policy before any action executes.
I think we're going to see a lot more tools like this in the coming months. Agentic workflows are exploding in popularity, and security tooling hasn't kept pace. Trajeckt being open-source and lightweight (~1.6ms overhead) is exactly the kind of thing that makes sense to layer into existing agent pipelines rather than a heavy proprietary solution.
Energy and Infrastructure: The Real Bottlenecks
Two stories from yesterday underline a theme that's been building all year.
First, DBS's CIO put out a note saying energy is the "chokepoint" for AI's next growth phase. The recommendation is straightforward: investors should add energy infrastructure exposure alongside pure tech plays, because AI power demand is growing faster than grid capacity can handle. This isn't a controversial take anymore — it's becoming conventional wisdom.
Second, Texas Governor Greg Abbott just called for a ban on new data center construction in rural parts of the state. This is a stunning reversal from someone who previously marketed Texas as the data center capital of the US. Rural communities have been pushing back hard against the noise, water usage, and grid strain that come with massive data centers, and Abbott's pivot signals that the NIMBY backlash is real and spreading.
If you're running any kind of AI workload at scale, these stories matter. Energy costs and data center availability are going to become competitive differentiators in the next 12-18 months. The days of "just spin up more instances" without thinking about where the power comes from are numbered.
Quick Bits
- The Register reports security researchers tricked LLMs into generating cocaine recipes through prompt injection that abused "role model" personas. The researchers' take: models can't reliably distinguish authorized from unauthorized input, and prompt injection isn't going away. Hard to argue with that.
- A self-hosted LLM user on XDA shared their experience running a local model on an RTX 5090 and watching it degrade over hours of continuous use — token drift, confusion, eventually nonsensical outputs. A reminder that local inference has thermal and memory management challenges that cloud APIs abstract away.
If you're running cost calculations for your next project, check out Decision Calculator.
Disclaimer: Some information in this post is based on preliminary reports and may be updated as more details emerge. Unverified industry rumors are marked accordingly. This is not financial or investment advice.
Cover image: AI-generated illustration for representational purposes.
Top comments (0)