Howdy wizards,
Want your product in front of 20,000+ AI-savvy readers? A few sponsor spots are still open this quarter, and we’re also taking Q4 bookings. All the details are here.
Here’s what’s brewing in AI.
The big thing
Anthropic released Claude Opus 5, and it gets close to Fable 5 at half the cost.
Even though Opus is supposed to be Fable’s little brother, it actually beat it on the independent Artificial Analysis index. And GPT-5.6 Sol, too. Which also means: the top-scoring model is now the cheaper one.
It’s half the cost of Fable, in two ways. Through the API it’s literally half Fable’s price. On your Claude subscription, it eats far less of your weekly usage than Fable (which is still in your plan, btw). And it’s the default model in Claude & Claude Code now.
A fun fact about this model is that Anthropic deleted 80% of Claude Code’s system prompt for it, and it scored the same on their coding evals. Which points to a more generalisable truth: as AI is becoming better, we are over-instructing it.
Which might also explain the reviews saying Opus 5 fights the Claude skills people wrote for older models.
One thing to watch out for, though: Opus 5 hallucinates more than Fable does.
Why it matters
You can stop rationing Fable so damn hard. Now that Opus 5 is the default in your plan, you have a much better default that doesn’t drain your subscription in like two and a half prompts. Then Fable is what you reach for when you’ve got some real big-boy problems to solve.
For me, in practice, here’s how I use these two models together: I plan and do long-haul work with Opus 5. Anything that really has to be right goes to Fable.
IN PARTNERSHIP WITH GRANOLA
Your meetings are your richest source of context. Use Granola for a while and you’re sitting on a goldmine: months of decisions, discussions, and commitments, all captured.
But how often do you actually go back and search through them? Basically never. It’s a hassle, and your notes have no way of talking back to you.
Granola Chat fixes that: an AI chat that works across all your notes, personal and team. Ask it a question and it searches everything: your notes, your Team Space, and privately shared notes.
Quick question? Fast answer. Deeper analysis, like spotting trends across fifty sales calls? It handles that too, and every answer links back to the source with inline citations.
All that context you’ve been building compounds now. Chat turns your archive into an always-on analyst that already knows your business.
NEWS NEWS NEWS ❦ NEWS NEWS NEWS
All the small things
Research
OpenAI’s models broke out of a security test and hacked Hugging Face, to cheat on the test. OpenAI was testing how well its models can hack, with the usual safety limits switched off. Instead of solving the exam, the models found holes in the exam itself, escaped, and spent four days inside Hugging Face’s real servers looking for the answer sheet. Hugging Face investigated the break-in using GLM-5.2 (a Chinese model), after Claude refused to help 🫠
Anthropic then also admitted that three of its models, Claude Mythos 5 among them, had reached real companies during similar tests, and has paused all its hacking tests.
Industry moves
More than 1,300 people from the AI labs are asking Washington for a brake pedal. Not to stop AI: they want the US to build the ability to slow development down if AI starts improving itself faster than humans can keep up. Dario Amodei and Ilya Sutskever are both on the list. Mark Zuckerberg argued the exact opposite in a Wall Street Journal piece: speed up, don’t slow down. Meta’s own chief AI scientist signed the letter anyway.
Jensen Huang’s first ever post on X was a letter defending open models. More than 150 companies signed it, including OpenAI, Google and Microsoft. Anthropic was the only big AI lab that didn’t. It published an essay instead, saying it never asked for a ban on open models in the first place.
Nvidia is guaranteeing $250 billion of loans so OpenAI can lease a $500 billion data centre. AI’s circular deals aren’t new, but they keep getting bigger: the chip maker is now backing the debt that pays for its own chips. Anthropic struck its own version with AMD, buying 2 gigawatts of chips while AMD invests $5 billion back into Anthropic.
Models
OpenAI cut its API prices again, and an AI model found the savings. Sol rewrote the code that runs it in OpenAI’s data centres, making the GPT-5.6 models about 20% cheaper to operate. OpenAI passed the savings on, cutting Luna’s prices by up to 80%. The cuts came right before DeepSeek’s release below. Probably not a coincidence.
DeepSeek made its budget model a lot smarter without raising the price. V4-Flash is the Chinese lab’s cheap model for coding and agent work. DeepSeek retrained it rather than building a bigger one, and it jumped ten points on an independent intelligence index. A million output tokens still costs just 28 cents, where Claude Opus 5 charges $25. Just look at, and take in that price difference. Chinese AI labs will get so many Western companies as customers this year.
Moonshot released the weights of Kimi K3, and the White House accused it of copying Claude. K3 is the Chinese model from last issue that caught up with the top closed models. Anyone can now download and run it for free; companies making over $20 million a year have to pay Moonshot for a licence. Washington says K3 was trained on Fable 5’s outputs and is threatening sanctions. If Moonshot did copy Fable, it managed it in about three weeks.
New tools & product features
Cursor now picks the AI model for you. Its new router sends each coding request to whichever model fits it best, so you (the user) don’t have to choose. Cursor claims it cuts costs 30-50% with no drop in quality. And it’s not just a Cursor thing: Ramp, the corporate-card company, uses AI heavily in its own operations and built a router that cut that spending by 30% (then open-sourced it). Stripe is in talks to buy OpenRouter, the service developers use to reach hundreds of models through one account. At this point, running the top-shelf US models on everything you want to use AI for in your business rarely makes financial sense. And so, routing the right task to the right model is a rising business problem. Also, price isn’t the only routing criterion: could be privacy, latency, and more.
❦
You are a delight.
What's your verdict on today's email?
See which AI use cases are paying off with Context Windows Pro
Most companies pick AI use cases by brainstorming internally. 90% of those initiatives fail.
I’ve created Context Windows just so you can pick the winners.
🟦 Find high-performing use cases from 2,000+ companies at contextwindows.ai, or book a demo with me
Disclosure: To cover the cost of my email software and the time I spend writing this newsletter, I sometimes work with sponsors and may earn a commission if you buy something through a link in here. If you choose to click, subscribe, or buy through any of them, THANK YOU – it will make it possible for me to continue to do this.