By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

I was a Claude hater. Fable 5.2 changed my mind.

Anthropic just released Fable 5.2, their new frontier model, and I spent two days actually building things with it before recording this.

Claire Vohost

Nov 9, 202624mWatch on YouTube

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

CHAPTERS

Jump to publisher timestamps

Chapter summaries

From “Claude hater” to shipping with Claude Code again
Claire opens by admitting she’s been openly frustrated with earlier Claude models, especially Fable and Opus 5, which she found intelligent but hard to work with. She explains why Codex briefly became her preferred environment for coding, then reveals she’s now back in Claude Code—largely due to Fable 5.2.
- •Earlier Fable felt like a big intelligence leap but was difficult to communicate with
- •Opus 5 was effective but not enjoyable to use
- •Codex won on developer experience, including browser/computer use
- •Claire is now mostly coding in Claude Code again and feels strongly positive about it
What launched: Fable 5.2, Mythos 5.2, and where each fits
Claire lays out what Anthropic released and how the models are positioned. She describes Mythos 5.2 as the same underlying model offered for cybersecurity and life sciences purposes, and frames Fable 5.2 as complementary to Opus 5.5 rather than a replacement.
- •Anthropic released Fable 5.2 alongside Claude Mythos 5.2
- •Mythos 5.2 is the same underlying model but aimed at cybersecurity and life sciences
- •Fable 5.2 is described as a frontier model
- •Claire says Anthropic frames Opus 5.5 as handling most daily coding and knowledge tasks
- •Fable 5.2 is positioned for scope, complexity, and creativity
Big claims and benchmarks: black holes and coding strength
Claire highlights some of the headline-grabbing claims around Fable 5.2, including a notable science anecdote she shared with her kids. She also summarizes what Anthropic says it’s good at for engineering work, and notes strong benchmark performance while emphasizing she’ll run her own full bench later.
- •Claire cites a headline claim that the model solved two open problems in black hole physics
- •Anthropic says it’s strong on stubborn bugs, performance optimization, architecture, and long-running tasks
- •She mentions large migrations and ambitious builds as claimed capabilities
- •She says benchmarks show it outpacing other frontier models at high effort settings
- •She plans a full run against her How I AI bench but shares only impressions in this episode
Guardrails and anti-distillation: what to expect (and what she hit)
Claire notes that the model includes cyber and bio guardrails and additional anti-distillation safeguards. She adds that she personally didn’t run into these guardrails during her tests, and underscores that the experience felt smoother and less “annoying” than prior versions.
- •Cyber and bio guardrails remain unless someone is in the Mythos program
- •There are anti-distillation safeguards in the model
- •Claire says she didn’t encounter guardrails in her testing
- •She emphasizes the model felt notably less annoying to work with
Pros and cons: speed, creativity, autonomy—and a design taste problem
Claire gives her high-level evaluation after a few days of hands-on use. She praises speed, creative output, and helpful autonomy, while calling out a key weakness: web/front-end design when left without strong style direction, plus limitations around image generation compared with tools she pairs alongside it.
- •She describes the model as extremely fast across turns and in autonomous work
- •She says it alerts her when it needs something but doesn’t over-interrupt
- •She praises creative work and detailed video generation (with one exception she later explains)
- •She calls out poor web design and “concerning taste” without a strong brand/style guide
- •She often invoked OpenAI’s Image Gen 2.5 because Fable wasn’t strong at image generation
Video generation demo: turning a tweet into a cinematic animated story
Claire shows how she used Fable 5.2 to transform a tweet about “software factories” into a fully animated video using her brand guidelines. She contrasts the result with slide-like AI videos she previously made, saying this felt more like real animation and was largely self-directed by the model.
- •She previously made videos using HyperFrames and Opus 5.5 but felt they resembled animated slides
- •She fed Fable a tweet plus cxo.dev brand guidelines and asked it to make a video
- •She highlights the output as a cohesive animated sequence with multiple scenes and visuals
- •She notes she did not storyboard and gave minimal creative guidance beyond the brand book
Animating the invader bots: sprites, movement, and a new benchmark task
Claire zooms in on a sub-problem inside the video project: making her brand’s “invader bot” characters animate believably. She describes how Fable 5.2 generated sprite-like motion behaviors without being explicitly prompted, and she adds this as a new evaluation item in her How I AI bench.
- •The invader bots existed previously as static website images made by Jamie Gannon
- •Claire wanted real animation (not simple wiggling across the screen)
- •She says Fable created sprite-like movements such as cheering and jumping without being prompted
- •She highlights speed and a strong sense of geometry/movement
- •She adds “animate the invader bots” to the How I AI evaluation bench
Real-world geometry win: assembling instruction-less flat-pack furniture
Claire recounts a practical test involving a flat-pack furniture delivery with no instructions. She says multiple models failed to produce workable assembly steps, while Fable 5.2 generated both a diagram-like breakdown and an order of operations that worked in practice during assembly.
- •She received furniture with many pieces, loose screws, and no instructions
- •She says every other model she tried failed to create usable directions
- •Fable 5.2 produced a diagram and step-by-step assembly instructions
- •She emphasizes the importance of correct assembly order and says it got the sequence right
- •She notes it succeeded even from low-resolution images and non-English labels
A “bug factory” shooter game: humor, concepts, and an auto-made trailer
Claire shares a playful project where she extended the brand aesthetic into a Space-Invader-style shooter about stopping bugs before production. She praises the model for being funny, generating game concepts tied to business themes, and creating a trailer with a narrative about agent count, token costs, and a “monorepo” boss.
- •She prompted it to make a bug-fighting game using the brand’s space invaders
- •The game includes choices like more agents, protected time, and tests as tradeoffs
- •She notes comedic elements like “vibe coding” causing many bugs
- •It generated a trailer storyline involving finance reacting to token bills and switching to cheaper models
- •She highlights how much more enjoyable this felt compared with earlier “not a good hang” model behavior
Serious engineering: re-architecting ChatPRD’s product intelligence pipeline
Switching to practical work, Claire describes using Fable 5.2 on a complex product intelligence pipeline at ChatPRD. She reports that the model proposed sensible organization that aligned with her own roadmap, and that it worked autonomously while stopping at appropriate moments for review and pull requests.
- •She uses Claude Projects and Opus 5.5 for many coding tasks already
- •She applied Fable 5.2 to ChatPRD, describing it as a complex ingest/organize/summarize/auto-wiki system
- •She asked it to improve efficiency and maintainability and to do “product thinking” on the pipeline
- •She says the restructuring was sensibly organized and matched her mental roadmap
- •She notes it stayed autonomous but paused appropriately when it needed a PR; she’s mid-migration but happy so far
Where it falls down: ugly web design and weak default aesthetics
Claire explicitly demonstrates the model’s weaknesses in web design and visual taste when not constrained. She shows examples like a browser homepage concept and a storybook project, describing them as unattractive and “slop,” and suggests pairing with other models or approaches for design-heavy work.
- •She calls the model “real bad” at web design
- •A browser homepage concept meant to help her slow down looks very ugly to her
- •A ghostly storybook project is “okay” but still pretty ugly in her view
- •She reiterates it can follow a good brand/style guide, but defaults are concerning
- •She hints she may pair Opus 5.5 with Fable for certain design needs
The “AGI moment”: building Claire’s fully interactive 3D typewriter
Claire presents her standout project: a functioning 3D rendered typewriter experience inspired by a 3D record player project she saw from Guillermo at Vercel. She walks through features like clickable keys, keyboard input, camera zoom, white-out corrections, ambient audio, writing timers, prompts, multiple scenes, and printing/saving output—emphasizing how much was created in essentially one shot.
- •She was inspired by a 3D record player project she saw from Guillermo at Vercel
- •She introduces “Claire’s typewriter,” a 3D rendered interactive typing environment
- •Features include typing via clicking keys or using the keyboard, zooming in, and white-out corrections
- •She adds ambience (music, rain), writing timers via sandglasses, and writing prompts
- •Multiple scenes include evening lighting, train travel, oil lamp, handwritten page mode, and a Y2K computer with green/amber screen plus dot-matrix printing and saved stacks
- •She describes it as largely one-shot from Fable and says it sparked strong excitement and creativity
Bench preview, availability, and final take
Claire closes with a quick preview of her How I AI bench results emphasizing speed, and says she’ll run full comparisons later. She explains where the model will be available and mentions upcoming “fast mode” and subscriber resets, ending with her overall verdict that Anthropic won her back.
- •She shares a bench preview emphasizing fast runtimes for complex tasks like animating the invader bots
- •She plans a full How I AI benchmark against other new models in the coming weeks
- •She says it will be available via API and the Claude platform, plus AWS, Google Cloud, and Microsoft Azure
- •She notes it will appear in Claude Apps and Claude Code, and mentions an upcoming fast mode
- •She mentions subscribers will get Fable resets and concludes she’s excited to build more consumer apps and videos

Publisher timestamps

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.