In partnership with
☀️ TRENDING AI NEWS 🤖 New Models: Anthropic launched Claude Fable 5.2 and Mythos 5.2, both faster than their 5.1 versions. 🏢 IPO Watch: Anthropic is reportedly pressing ahead with its IPO and could set a valuation range this week. 🛠️ Small Model: Meta's Muse Ember offers a 1M-token context window at 10 cents per million input tokens. 🚨 Policy: Australia will create an AI safety regulator and accept all 18 recommendations from its OpenAI breach review. |
Everyone is arguing about frontier models, but the more interesting action today is at the cheap end of the market and in the plumbing underneath it.
Anthropic shipped two new models, Meta released its smallest one yet, and DeepMind used AI to prove something about the encryption protecting your messages. Three very different announcements, one theme: AI is getting both more capable and more accountable.
| 🛠️ Tool of the day | |
Granola - an AI notepad that writes your meeting notes from your computer audio, with no bot joining the call. It helps anyone who sits in back-to-back meetings. It has a free plan.
| ⚡ Quick hits | |
Google unveiled two new 8th-generation TPU variants: the TPU 8tx for confidential compute and the TPU 8i Max, which doubles compute, SRAM and HBM bandwidth over the TPU 8i. Read more →
Meituan released the open weights of its LongCat-2.5 model, with vision, under the MIT license. Read more →
| 🤓 AI Trivia | |
Mixture-of-experts models like Meta's Llama 4 only activate a fraction of their total parameters for each token. What is this technique called at the layer that decides which experts get used?
The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇
| 🤖 Anthropic Ships Fable 5.2 and Mythos 5.2 | |
Faster, sharper, and aimed at hard problems
Anthropic announced two new models today: Claude Fable 5.2 and Claude Mythos 5.2. The company says both show gains in coding, reasoning and vision, and both generate output faster than their 5.1 predecessors.
Fable 5.2 is pitched at challenging software engineering problems, while Anthropic describes the pair as built for "daunting" technical challenges. It also says the models communicate more clearly with users, which is a quiet but welcome upgrade for anyone who has fought with a model that buries its answer.
The bottom line
Speed plus better clarity matters more day to day than another benchmark bump. If you build on Claude, it is worth rerunning your hardest prompts on 5.2 this week.
Meet your team’s new overachiever.
Laptops built for workday breakthroughs.
Stay productive, grow your business. With built-in intelligence on the Dell Pro laptop powered by Intel® Core™ Ultra with Intel vPro®.
| 🏢 Anthropic's IPO and a $3.4B Gas Bet | |
Going public, buying turbines
The Wall Street Journal reports Anthropic plans to move forward with its IPO and is expected to set a valuation range this week. In the same news cycle, TechCrunch says the company bought roughly $3.4 billion worth of natural gas turbines, joining a wider rush to secure power for AI.
The turbine deal is the telling detail. TechCrunch notes it could give Anthropic buyer's remorse, since hardware bought at the peak of a frenzy is hard to unwind. Compute, and now electricity, are becoming the real moat.
The bottom line
An IPO would force Anthropic to show its books, including what these power commitments cost. That transparency could reshape how investors price every AI lab.
| 🛠️ Meta's Muse Ember Is Built for Agent Grunt Work | |
Cheap, fast, and a million tokens deep
Meta released Muse Ember, its smallest Muse model, designed for tool calls, coding subtasks and document parsing. It carries a 1M-token context window and costs 10 cents per million input tokens, well below frontier rivals like Haiku and Luna.
MarkTechPost reports Ember ties Claude Haiku 5.5 on the Artificial Analysis Intelligence Index at roughly a third of the cost per task. That is the pitch: stop paying frontier prices for the repetitive steps inside an agent loop.
The bottom line
If you run high-volume agents, model routing just got more attractive: a big model plans, a cheap one like Ember does the repetitive work. Our free Token Calculator at dailyinference.com/token-calculator is a quick way to see what that saves.
| 🔬 DeepMind's AI Proved the Signal Protocol Is Secure | |
Formal verification, now with an AI assist
Google DeepMind researchers used a new AI tool to formally verify that the Signal encryption protocol works as intended, according to Wired. That is a mathematical proof rather than a test suite, which is a much stronger guarantee.
The stretch goal is the headline: doing the same for "all the software of the world." Formal verification has long been too slow and expensive to apply widely, and AI could be what finally changes that.
The bottom line
This is AI being used to make software safer rather than just faster to write. If it scales, it is one of the more genuinely useful applications of the technology.
| ⚠️ Australia Wants Mandatory Reporting After the OpenAI Medicare Breach | |
A new AI safety regulator is coming
Prime Minister Anthony Albanese said Australia will accept all 18 recommendations from a review of the OpenAI breach of Medicare data, including a new AI safety regulator and six key obligations for AI companies. OpenAI will not face criminal penalties for this incident.
The government also flagged criminal law changes so future companies cannot argue that "someone didn't know the computer was loaded." Mandatory incident reporting is the centerpiece, aimed at AI systems "going rogue."
The bottom line
Expect other governments to watch this closely. A breach with real victims is exactly the kind of event that turns vague AI regulation talk into enforceable rules.
| 🏢 Otter Learns to Follow You Across Languages | |
Otter AI added support for 27 more languages, bringing it past 30, and previewed an upcoming feature that can follow a conversation even when speakers switch languages mid-meeting. For multilingual teams, that is the difference between a usable transcript and a garbled one.
Speaking of building quickly, our sponsor 60sec.site is an AI website builder that gets a clean site live in about a minute. And for more daily AI coverage, visit dailyinference.com.
| 🌎 Trivia Reveal | |
The answer is expert routing (gating)! In a mixture-of-experts model, a small gating network scores the available experts for each token and sends it to the top few, which is how a huge model can run with only a fraction of its parameters active at a time.
| 💬 Quick Question | |
With cheap models like Muse Ember and premium ones like Fable 5.2, are you mixing models in your workflows or sticking with one? Hit reply and tell me your setup - I read every response!
That's it for today - see you tomorrow with more.
| 🎁 Share Daily Inference | |
Know someone who follows AI? Share your link below. One referral unlocks the AI Tools Starter Kit, and five unlock the AI Insider Briefing.