|
Read Online | Sign Up | Advertise
|
|
|
|
|
|
Good morning, AI enthusiasts, and welcome to our 5,334 new readers. Anthropic spent the fall telling the AI industry to pace itself. The company’s first look at what that means just arrived, and the result still sits comfortably at the top of the leaderboards.
|
|
Fable 5.2 is cheaper, faster, and strong enough to bump Google’s Gemini 4 Carbon before the public even got to try it. If this is Anthropic’s slow lane, it still looks faster than the rest of the industry can keep up with.
|
|
|
|
In today’s AI rundown:
|
-
Anthropic’s Fable 5.2 leads the frontier with a price cut
-
Meta undercuts Anthropic, OpenAI on the cost of agents
-
Clean up a messy Excel report with Claude
-
Google adds two more AI chips to its Nvidia fight
|
|
|
|
|
|
|
|
|
ANTHROPIC
|
|
|
 |
|
Image source: Artificial Analysis
|
|
|
The Rundown: Anthropic just released Claude Fable 5.2 and Mythos 5.2, two cheaper, faster upgrades to its most powerful AI family that take the top spot on Artificial Analysis’ AI leaderboard — and show how CEO Dario Amodei’s call to “pace the frontier” will work.
|
|
The details:
|
-
Fable debuted at No. 1 on Artificial Analysis’ Intelligence Index leaderboard, edging out Google’s still unreleased Gemini 4 Carbon model by 5 points.
-
Mythos is limited to vetted security and life sciences groups, while Fable is the same model with stronger cyber and biology guardrails.
-
Teams of 1,000 Fable agents built an operating system in 8 days and formalized the full proof of a famous math problem in 30 hours, with Opus finishing just 26%.
-
Fable was held near Opus level in certain skills that speed up development, with exec Dianne Penn saying of the slowdown, “We called for pacing, not pausing.”
-
5.2 is 20% cheaper and as much as 30% faster than 5.1, though at $8 in / $40 out per million tokens, it still costs twice as much as Opus 5.5.
|
|
Why it matters: Google’s run at the top of the leaderboard ended before Carbon was even released, and Anthropic’s pacing certainly doesn’t feel like a slowdown for users without its internal super-models. With OpenAI also expected to release a new model, the bar just moved even higher — and anything less will likely come as a disappointment.
|
|
|
|
|
TOGETHER WITH VANTA
|
|
|
|
|
|
The Rundown: Vanta’s AI governance webinar series is six live and on-demand sessions for security and GRC leaders on how to govern AI as fast as you adopt it — covering frameworks, expert Q&A, AI risk, and a live product demo.
|
|
In the series, you’ll explore:
|
-
Frameworks for governing AI from day one
-
Real-time Q&A with Vanta’s GRC experts
-
AI risk, workshops, and a live demo
|
|
Register for a session.
|
|
|
|
|
META
|
|
|
 |
|
Image source: Meta
|
|
|
The Rundown: Meta released Muse Ember, the smallest AI model in its lineup, built to handle routine, high-volume steps inside AI agents — with the company saying it matches Anthropic’s budget Claude Haiku 5.5 at roughly one-third of the per-task cost.
|
|
The details:
|
-
Ember tied Haiku 5.5 at 36 on AA’s Intelligence Index, while beating out OpenAI’s comparable GPT-6 Luna (28) by a wide margin.
-
Meta AI chief Alexandr Wang said Ember is meant to pair with the larger Muse Wonder, writing that “wonder can plan, ember can handle the volume.”
-
Ember’s price matches both rivals, with savings coming from using fewer tokens per job and an average test task running $0.14 on Ember and $0.41 on Haiku.
-
Meta trained Ember using its larger Muse Wonder Full, the first model built that way, and said “even more efficient models” are in training.
|
|
Why it matters: Meta shipped its first Muse model in April, and it now has answers for every tier Anthropic and OpenAI sell, from the near-frontier down to the bargain bin. Add in a new report of a sales push through the consulting giants those two already use, and Meta’s 2026 consumer success might soon touch the enterprise end of the AI pool.
|
|
|
|
|
AI TRAINING
|
|
|
|
|
|
The Rundown: In this guide, you’ll use Claude inside Excel to clean up a messy sales export, fix a wrong total, and build a report. You’ll check the numbers and test whether a new order updates the chart.
|
|
Step-by-step:
|
-
Install the official Claude add-in and open a copy of your workbook. Pro members can use our practice file and keep the reference workbook closed until they finish
-
Tell Claude what each row means and which orders to count. Ask it to find duplicate rows, missing details, and formula mistakes before changing anything
-
Approve the fixes on a separate sheet. Keep the original data and ask for a record of each change, including any rows left out
-
Build a monthly report with formulas and a chart. Check the order count, refunds, and sales total against the source rows
-
Add a test order and check that the report and chart update. Remove it and confirm the original totals return before saving your finished workbook
|
|
Pro tip: A spreadsheet can show a wrong total without showing an error. Ask Claude to explain which rows it counted, then check a few yourself.
|
|
|
|
|
PRESENTED BY CUBE.DEV
|
|
|
|
|
|
The Rundown: Cube is the semantic layer for AI. Define metrics and access rules once, let AI build your model, and agents, dashboards and apps get consistent answers. BetterUp’s team uses it to query governed data right in Claude, saying it delivers “reliable results, even if you don’t know exactly the right question to ask.”
|
|
With Cube, you can:
|
-
Model metrics and joins in code, with access policies enforced at query time
-
Connect Claude or ChatGPT over MCP to chat with your governed data
-
Build dashboards without code: a non-technical BetterUp manager built one of its most-used dashboards
|
|
See BetterUp’s story, then get started with no card required.
|
|
|
|
|
GOOGLE
|
|
|
 |
|
Image source: Google
|
|
|
The Rundown: Google announced the TPU 8tx and 8i Max, two specialized versions of its custom AI chips arriving next year, with one built to keep sensitive data safe and the other to handle jobs needing an AI to hold huge amounts of information at once.
|
|
The details:
|
-
Google began selling TPUs to other companies this year and has nearly 100 customers, including Anthropic and Apple, with every available chip spoken for.
-
Data stays encrypted on the 8tx even while an AI model is working on it, a step where it normally sits exposed, with governments and regulated firms as targets.
-
The 8i Max joins two chips as one unit to double processing power and memory speed, for AI agents that must keep an entire codebase or long video in view.
-
“The future is going to be about specialization,” Amin Vahdat, who runs Google’s chip work, told Bloomberg, expecting both to reach customers next year.
|
|
Why it matters: Google couldn’t have asked for better timing, rolling out two Nvidia challengers the day after Jensen Huang’s apology and his board’s probe. While Google still seems constrained in how much it can build, the demand is clearly there — making it interesting both as a chip supplier and as an AI lab that owns its entire stack.
|
|
|
|
|
|
|
|
|
COMMUNITY AI WORKFLOW OF THE DAY
|
|
▸ A Rundown reader built an AI running coach for a friend
|
|
Today’s workflow comes from an anonymous Rundown reader:
|
|
“My friend is a competitive runner who wears a Garmin watch for every session. Each run produces a spreadsheet’s worth of data: pace, heart rate, cadence, ground contact time, vertical oscillation, and power, broken down lap by lap. Garmin Connect shows the numbers but can't explain them. It can't tell him whether a late pace drop came from fatigue, wind, or a hill, or whether his fitness is really improving. I built Stride to add that meaning for runners who already have the data.
|
|
Stride runs sports-science-style analysis. Aerobic fitness tracks heart rate within each pace band week over week, so a lower heart rate at the same pace signals improving fitness. The Environment layer adds weather and elevation, so a slow lap into a headwind is judged fairly.
|
|
AI Coach sends a structured 30-day summary to Claude, which returns a plain-English brief on trends, environmental factors, and things to watch. What-If answers questions grounded in his own data, such as his likely 5K time on a flat, windless course.”
|
|
See the workflow on our Community Workflow Hub →
|
|
|
|
How do you use AI? Tell us for a chance to be featured in the newsletter.
|
|
|
|
|
|
|
-
🏆 Claude Fable 5.2 - Anthropic’s top-ranked model for the hardest challenges
-
🎨 Seedream 5.5 - ByteDance’s new image AI with strong editing skills
-
🔥 Muse Ember - Meta’s low-cost AI model for coding and agent tasks
-
🖋️ Inkling 2 Small - Thinking Machines Lab’s new coding and vision model
|
|
|
|
|
|
|
|
Nvidia’s top lawyer reportedly emailed agents in August 2025 as they investigated whether Singapore’s Megaspeed smuggled its chips to China, with the U.S. Commerce Department then halting the case and handing it to the Justice Department.
|
|
Anthropic is preparing to proceed with its IPO despite current fear in the AI sector, with shares likely launching for trading as soon as Thursday next week, The WSJ reports.
|
|
Adobe signed a multi-year deal with Disney to build custom models trained on Disney’s characters and visual styles, with work underway for theme parks, film, and ESPN.
|
|
Instagram launched Effect Builder, a free web tool that lets U.S. users turn text prompts into AI Stories effects, which others can use after passing a safety review.
|
|
Open-source AI hosting startup Fireworks confirmed a hack that exposed billing details and parts of some customers’ models and datasets, days before it planned to start raising at a $30B+ valuation.
|
|
|
|
|
|
|
|
|
That's it for today!
Before you go we’d love to know what you thought of today's newsletter to help us improve The Rundown experience for you.
|
|
|
|
See you soon,
|
|
Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown
|
|
|
|
|
|
|
|
