← All transcripts

Did Anthropic just save its $1 trillion IPO? Transcript, AI Summary & Key Points

Fireship · yesterday · Science & Technology · 05:34 · EN

Watch on YouTube

Answer

Anthropic released Fable and Mythos 5.2 during its IPO roadshow after the expected valuation fell from $2 trillion to $1 trillion, giving investors new model results to focus on — though the video doesn't conclude the IPO is actually saved.

AI Summary

Anthropic released Fable and Mythos 5.2 in the middle of a rocky IPO roadshow, after US midterms turned into a referendum on data centers and Wall Street panic-sold AI stocks, dragging Anthropic's expected valuation from $2 trillion in September to $1 trillion. Fable and Mythos are the same model: Mythos is the unrestricted version reserved for large companies, while Fable has guardrails. Fable 5.2 is faster and cheaper than the versions released two months earlier, costing $8 in and $40 out (down from $10 and $50 for 5.1), though Anthropic recommends sticking with Opus 5.5 for most work. In testing it built a bootable operating system from an empty folder over 8 days of sub-agent work, completed a Lean formalization of Fermat's last theorem in 30 hours with 1,000 agents (a proof 7 times shorter, while Opus 5.5 crashed out at 26%), and designed hardware including a voice assistant hub, USB speaker phone and handheld virtual pet with circuit board company Diode. On benchmarks it beats GPT-6 Astra everywhere Anthropic listed, loses to Gemini 4 Carbon on business automation, and more than doubles Astra on a robotics benchmark Anthropic ran itself. On Thoroughbench — a realistic interactive 3D horse built with Three.js — Fable sculpted the horse from signed distance fields, self-critiqued and fixed its first attempt, and scored 7.77, a new state of the art; the horse, named TypeScript, lives at typescript.dev.

Key Points

  • Anthropic's IPO roadshow began as US midterm voters rejected data centers and Wall Street panic-sold AI stocks, dropping Anthropic's expected valuation from $2 trillion in September to $1 trillion
  • Fable and Mythos 5.2 are the same model: Mythos is unrestricted and reserved for large companies with cyber security insurance budgets; Fable has guardrails
  • Fable 5.2 pricing is $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for 5.1; Anthropic suggests sticking with Opus 5.5 for most work and swapping to Fable only when needed
  • Pointed at an empty folder, Fable's sub-agents spent 8 days building an operating system that boots on a real laptop with its own kernel and drivers
  • A swarm of 1,000 Fable agents formalized Fermat's last theorem in Lean code in 30 hours with a proof 7 times shorter than August's 11-day effort, while Opus 5.5 crashed out at 26% with the same setup
  • Working with Diode, Fable turned a one-paragraph prompt into designs for a voice assistant hub, a USB speaker phone and a handheld virtual pet
  • On benchmarks, Fable beats GPT-6 Astra everywhere Anthropic listed, loses to Gemini 4 Carbon on business automation, and more than doubles Astra on a robotics benchmark Anthropic ran itself
  • Thoroughbench test: asked to create a realistic interactive 3D horse with Three.js, Fable cooked for 43 minutes and sculpted the horse out of math using signed distance fields meshed and rigged to a skeleton on every page load, then screenshotted its own attempt, roasted the flaws (bluish mane, stick-like tail, thin neck and hind legs) and fixed them all

AI in practice

Used for

Agents

  • Build a complete operating system from scratch starting from an empty folder. 2 held 01:33
  • Complete a math project formalizing Fermat's Last Theorem in Lean that Opus failed to finish. 2 held 01:55

Links mentioned

🔒 Full analysis locked

Unlock more videos and the full analysis

A credit unlocks one video's full analysis for good — the build steps, the tools and how each was used, the methods behind every use case. Pro opens the whole library instead, and raises how many videos you can analyse a day.

Unlock full analysis — free

Transcript

Searchable transcript of Did Anthropic just save its $1 trillion IPO? — Fireship (05:34). Search for a phrase, then click its timestamp to jump straight to that moment in the video.

Captions sourced from the original video on YouTube, published by Fireship. The video, its captions and all related intellectual property remain the property of their respective owners; AINotes claims no ownership. Provided for research, accessibility and search — see the Transcript Notice and Copyright Policy.

00:00 While Honeypot Wong was busy falling head over leather jacket for a succulent Chinese spy. >> These China Chinese models are excellent. >> Daario was putting the final touches on his prep work for the anthropic IPO road show that started yesterday. Unfortunately, it's coming at a pretty rough time for Dr. Doomsday. The US midterms turned into a referendum on data centers where voters made it clear they hate them and Wall Street responded by panic selling every AI stock it could find.

00:24 that that dragged Anthropic's expected valuation down from two trillion in September to a mere 1 trillion today. So yesterday, dollar store Daario gave investors something else to look at by releasing Fable and Mythos 5.2, which you'll never believe are faster and cheaper than the versions that just came out 2 months ago. During testing, it built an entire operating system starting from an empty folder.

00:43 It finished a math project in 30 hours that Opus failed to finish, and it even designed a working Tamagotchi. In today's video, we'll look at everything these new models can do and answer the question I know you've all been wondering. It can it create an interactive thoroughbred with an appropriate amount of realness. It is November 10th, 2026 and you're watching the code report.

01:02 If you don't follow the clawed cinematic universe closely, the Fable and Mythos are the same model, but Mythos is the unrestricted one reserved for large companies with budgets for cyber security insurance, and Fable is the one with guardrails that won't let you research how to cook up smallox. Now, we'll get to the fun stuff in a second, but let's get the boring stuff out of the way first.

01:20 The price is now $8 in and $40 out, which is down from 10 and 50 from 5.1. But Anthropic suggests that you should stick with Opus 5.5 for most work and only swap to Fable if you really need it. Before the operating system demo, an engineer pointed Fable at an empty folder and let sub agents have a Jensen inspired Freakfest for 8 days. What comes out is a fairly legit OS that boots on a real laptop with its own kernel and drivers.

01:45 And you may remember back in August when an unreleased Claude spent 11 days turning Form's last theorem into lean code that a computer could verify. This time a swarm of 1,000 Fable agents did it in 30 hours and the proof was 7 times shorter while Opus 5.5 crashed out at 26% with the same setup. And apparently it's decent at hardware, too. Working with a circuit board company called Diode, Fable took a one paragraph prompt and used it to design a voice assistant hub, a USB speaker phone and a handheld virtual pet.

02:15 On the TMBBBS, it beats GPT6 Astra everywhere Anthropic listed in OpenAI score. It loses to Gemini 4 Carbon on business automation and it more than doubles Astra on a robotics benchmark that Anthropic ran itself. But enough glazing. As you know, there's only one benchmark I trust. a benchmark that's so good it's become the industry standard by which humanity measures its new god in a box toys thoroughbench.

02:36 So I open Claude Code and type those five famous words to create a realistic interactive 3D horse with 3JS. Make no mistakes. Fable cooked for 43 minutes and instead of downloading a horse model like a normal developer, it sculpted the entire animal out of math using signed distance fields that get meshed up and rigged to a skeleton every time the page loads.

02:58 It even screenshotted its own first attempt and roasted it. The mane is bluish, the tail is stick-like, and the neck and hind legs are thin. Then it went back and fixed every one of them. The end result is the most realistic horse Thoroughbench has ever seen. The skin texture and lighting look great. Its head follows your mouse cursor, and you can steer it with the arrow keys or orbit around it in 3D.

03:16 It's very cute and friendly, but gets extremely shy if you get too close. The transitions between gates are buttery smooth from a standstill all the way to a gallop, although it's a little more awkward when coming to a full stop. Under the hood, Opus 5.5's horse had 95,000 vertices and 35 bones, while Fables has over 120,000 vertices and 38 bones. And it plants each hoove with inverse kinematics.

03:37 So unlike Kimmy, it's not just warping the mesh. If Fable 5.2 didn't hallucinate a giant low poly fallus like Gemini, but it equipped our horse with those rear leg shock absorbers, which probably come in very handy when the horse decides to rear, as you can see here. And the man is a little lopsided unless this was a custom request for his barber. And even though I told it to make no mistakes, it admitted the horse is not photoreal.

03:59 And the eyes are halflitted. So the final score for Fable 5.2 on Thoroughbench is a 7.77, making it the new state-of-the-art in horse. And this really got me thinking. The models are starting to get so good that we should really be naming the horses they create to keep track of them. So, we've decided to call this one TypeScript in honor of Secretariat, the greatest racehorse of all time, who I can only assume would have loved to type if it had fingers.

04:23 And if you want to take him for a ride yourself, he now lives at typescript.dev, where you can walk, trot, and gallop him right in your browser. And if you ever need to run a video in your browser, you can do so with MX, the sponsor of today's video. Their highly customizable API is by far the easiest way to add video features to your application. We've used it for years to host our courses, but it does a lot more than just infrastructure.

04:43 When you upload a video to MX, you automatically get transcripts, storyboards, thumbnails, and clips. From there, MX Robots gives you AI workflows that can find key moments, or pick the best thumbnail without you needing to host a model or maintain a pipeline. You can automate all of this with directives where you define a workflow once, and it runs on every new upload.

05:02 And robots keep up with new AI models automatically, so you can safely ignore the next video about whichever model just changed everything. Perplexity, Patreon, and many other prestigious companies all trust MX, and you can try robots on your own video in the dashboard. The free plan includes 100,000 robots, units, and delivery minutes every month, and you get an extra $50 credit at mx.com/fireship or at the link below. This has been the Code Report. Thanks for watching, and I will see you in the next one.