Ethan Mollick's early test of Anthropic's Mythos-class model shows it tackling a roughly 19-page design spec with almost no ongoing guidance, running for 9.5 hours while spinning up sub-agents, pulling real flight and rail data, writing code, and verifying its own outputs to deliver working software for statistical analysis.
I've had access to Fable for a bit. A genuine jump in capability, I could feed it a 15 page design document for a project and it would work for 9+ hours and deliver terrific results.
But working with it is weird & weirder is coming
Lots of examples: https://open.substack.com/pub/oneusefulthing/p/what-it-feels-like-to-work-with-mythos?r=i5f7&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true
Instead of constant back-and-forth, the model handles hundreds of micro-decisions on its own and only surfaces occasional updates, which feels novel for anyone used to steering AI step by step.
At double the per-token rate of earlier Opus models, the nine-hour runs could get expensive unless cheaper sub-agents offset much of the spend, though exact totals from the tests remain unclear.
Many users praised Fable AI's nine-hour project results and snake game demo as impressive evidence of real-world AI progress and AGI potential.
@emollick Thanks for testing it!
I've had access to Fable for a bit. A genuine jump in capability, I could feed it a 15 page design document for a project and it would work for 9+ hours and deliver terrific results.
But working with it is weird & weirder is coming
Lots of examples: https://open.substack.com/pub/oneusefulthing/p/what-it-feels-like-to-work-with-mythos?r=i5f7&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true
Some fun examples, I just gave basic prompts and the AI executed: Balatro, but for coin flipping (all the design and ideas was Fable): https://play-flipside.netlify.app/ The best self-aware snake game: https://snake-stable-build.netlify.app/ An isochronic map using real data: https://isochronic-passage-chart.netlify.app/#syd
@emollick Can you share the prompts you used for the games? The snake one (http://snake-stable-build.netlify.app) is way more creative than anything I've seen from other models, so I'm curious how much of that was you vs the model
@emollick In case anyone is interested, had ChatGPT 5.5 Pro review the paper https://chatgpt.com/s/t_6a284f7783b4819191f96c0c5c5d638e. Briefly put: A solid paper, and the "flaws" are not major errors but elements that are typically modified during a review process. Note, I haven't read the paper fully myself.
@emollick Does it talk like Opus 4.8 ?
@emollick the security cost hides in "9+ hours." approval used to be per action. a model that runs unattended that long collapses it to one decision at hour zero. you're not approving a task anymore, you're approving nine hours of judgment you never see.
@emollick ye similiar experience so far, its very careful model and verifies its own work a lot better than pervious model. its also very slow at least so far today
@emollick
@bribiotech @emollick Have you tried it in cowork? It's done amazingly well for PPT documents in cowork for me. Clear and design forward.
@emollick the 9-hour run is wild but the thing that stuck with me is the bottleneck just moved onto the doc you hand it. fed a sloppy spec once and got back beautifully executed nonsense, fast. writing the spec is quietly becoming the whole job.
@emollick 9+ hours on a 15-page doc is exactly what we see with Claude Code on Mato. It runs unsupervised and makes judgment calls you didn't ask for. Some are right. Some aren't. What category of weirdness showed up first for you?
@emollick Oh my god the self-aware snake game is... genuinely funny and surprising?!
@emollick Ethan, stop everything you’re doing, drop every engagement you have, and open a game company with just this snake game as the product. It’s awesome!
@emollick Checked your article, most of it at least.
Still did not find anything surprising or innovative, have you tried to test it in what apparently is its niche? (Security and attack vectors eval).
@emollick Sounds like GPT-Codex-5.5
@emollick nine hours is past the point where it feels like your project. that's the weird part.
@emollick Nice I don't get is why Anthropic can't sort out presentations. Claude in PowerPoint produced slides read like someone had a case verbal diarrhea, theres is no balance all ideas are dumped into the pack
@emollick 15 pages in, 9 hours of autonomous work, terrific results. the part that's hard to adjust to isn't the output. it's not being present for any of the decisions.
@emollick Application on distance calculation using Isochromic map was really impressive
@emollick I've wondered how Fable would approach subtasks. From your write-up, it's clear it continues the trend of smarter models producing better outcomes by better managing subagents.
Frontier models managing small open models (via model routing) is the path to abundant intelligence
Ethan Mollick's early test of Anthropic's Mythos-class model shows it tackling a roughly 19-page design spec with almost no ongoing guidance, running for 9.5 hours while spinning up sub-agents, pulling real flight and rail data, writing code, and verifying its own outputs to deliver working software for statistical analysis.
I've had access to Fable for a bit. A genuine jump in capability, I could feed it a 15 page design document for a project and it would work for 9+ hours and deliver terrific results.
But working with it is weird & weirder is coming
Lots of examples: https://open.substack.com/pub/oneusefulthing/p/what-it-feels-like-to-work-with-mythos?r=i5f7&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true