Mythos Is Out of the Cage: How to Use Fable 5 Well — Two Free Weeks, Squeezed Dry

Myth or fable, either way — the Mythos generation of models is finally here. This model kept us waiting more than two months — word was that it was so capable that shipping it would open up all kinds of software security problems, which is why only large enterprises got early testing. And last night, the consumer version of the Mythos-class model line finally shipped, called Fable 5.

They built in safety guardrails: anything touching cyberattacks or biology-related content automatically switches to a lower-tier model, so you can’t use Fable 5 for that.

Officially, you can use it on subscription quota until June 22; after that, anyone who wants this model has to top up separately like fast mode and get billed like the API — and that gets very expensive. So in these two limited weeks, how do you burn through the subscription quota and put the strongest model where it actually belongs? I put real work into researching that today, and I’ll share what I found.

How expensive it is, how fast it burns quota

Start with the price, because it’s steep: $10 per million input tokens, $50 output, versus $30 for Opus 4.8 before it — roughly double the previous generation.

On the subscription plan: I used to exhaust my daily quota with about five hours of heavy use, and with this model it might all be gone in two and a half hours. Someone running multi-agent concurrency plus heavy tuning work could burn through five hours’ worth in a few minutes.

Which tasks deserve it

Get one thing straight first: Fable 5 isn’t a simple upgrade over the previous models. It’s an all-around model that crushes the previous generation and everything else across the board, and two things in particular: depth, and the more complex the task the better, plus tasks requiring long stretches of self-consistent, self-running work. So do not spend the quota on ordinary daily tasks.

Look at the gaps in the numbers: on production-grade software engineering it hits 80%, while other models including Opus 4.8 are just under 70% — GPT 5.5 and Gemini are left far behind. On the hardest production coding, it’s more than a 2x capability gain over Opus 4.8. On long-horizon autonomous tasks it reaches beyond the 16-hour mark.

Tasks that deserve it:

• Large project refactors, cross-repo migrations — it has to read large volumes of code on its own, run for long stretches against a task plan, and verify itself over and over. This is exactly what the model is for
• Coding problems the previous models couldn’t solve — pull Fable 5 out and have it look at that code
• Code audits with high logic risk — billing, payments, metering, authentication
• Refactoring a vibe-coded MVP demo to production grade — hand it a product demo built by someone who can’t code and let it re-pick the stack and rebuild from the framework up

Everyday copywriting, summaries, file cleanup, small edits, light research? Don’t waste this model’s quota on any of it.

One prompt: let it surface the high-leverage tasks itself

I wrote a prompt you can copy and use directly. The core is telling Claude Code: Fable 5 is a stronger model than Opus 4.8, our subscription covers it free until June 22, and to spend these two weeks’ quota where it counts —

1. Survey: spin up a multi-agent team, scan every project folder in all my working directories in parallel, plus memory and the registered to-do items, and tag each project with its status, its relationship to revenue (revenue-generating commercial projects are obviously prioritized), which step it’s currently stuck on, and how it compares to the remote repo
2. Priority filter: anything involving architecture design, large refactors, high-risk logic vulnerability audits, or adversarial review goes to the front of the priority queue; daily tasks get downweighted
3. Schedule: fill the quota as fully as possible every day of these two weeks, turn the tasks you’re handing it into a shift schedule, define the completion criteria for the output, and list them out to work through one by one
4. Exceptions: some projects have deadlines and need urgent completion — list those separately and handle them first

Given that, Claude Code goes through every project on my plate, finds the highest-leverage things that most deserve this model — refactors, audits, that category — then hands the to-do list to it and spends the quota where it counts, every day.

Small tricks for saving quota

I did some research here too. People are burning through quota fast, and the main reason is this: with Opus, we’re used to cranking thinking depth to Max or X-High. With this new model, you can start at High — it proactively spins up subagents and stacks workflows on top, and its autonomous concurrency is very strong, unlike Opus. Max it out and turn it loose, and the quota vanishes in an instant. Even for something like a code refactor, High or X-High is already plenty.

The safety review really is strict

Its safety review is extremely strict right now. Today I had it run a job downloading papers that touched on some pharmaceutical material, and it refused outright, downgrading to Opus 4.8 to run it. The official line is exactly what it looks like: they’re worried about a model this capable being used for unlawful purposes, and I get it. But a friend of mine is in a rough spot — he works at a biopharma company, and their day-to-day work and tool development are now all off-limits to the newest model. A real shame.

That’s it for today on Fable 5. I’ve posted two episodes today — I really am excited; this model is unbelievably capable. See you in the next video, tomorrow.

Source: EP0036_audio.mp3 · ASR model gemini-2.5-pro (chunked parallel) · full text of the original recording

[00:00] Myth or fable, whichever you prefer — we finally have the Mythos-generation model. As you all know, Mythos has been dangling in front of us for more than two months. The story was that the model is extremely capable, and that once it shipped there’d be all sorts of software security problems, so it only went to large enterprises for early testing. And then just last night, they finally released the civilian version of the Mythos Class model family, called Fable 5. This model actually has some safety guardrails built in — limits, basically. Whenever we get into anything involving

[00:26] network attacks, or anything biology-related — because they’re worried someone will use it to develop biological viruses, that kind of thing. So any content touching those two topics, it automatically switches to a lower-tier model and won’t let you use Fable 5 for it. Right now the official line is that before June 22nd you can use your subscription quota on this model. Which means after June 22nd, anyone who wants to use it has to top up separately, like with fast

[00:51] mode — billed per use like the API. And that gets very expensive. So within these limited two weeks, how do we burn through our subscription quota and put the strongest model where it actually belongs? I put real work into researching this today, so let me share it with you. First let me go over what’s different about this model. Point one: it’s very expensive. Its per-million-token input and output pricing is almost —

[01:17] think about it, input is $10 and output is $50. Opus 4.8 before it was $30. That’s almost double the price of the previous generation. And on the subscription side — say normally I’d use it heavily for about five hours a day to burn through my quota. Now with this model it might be 2.5 hours, two and a half hours, and your whole subscription quota is gone. Some people turn on, turn on multi-agent concurrency and then do

[01:42] a lot of heavy tuning work on top of it, and they can burn that whole five hours’ worth in a few minutes. So let’s get one thing straight first. This new Fable 5 model is not a simple upgrade of the previous one. It is an all-round model in the sense that it crushes the last generation and every other model on every dimension. But there are two things it’s especially good at. The first is depth —

[02:07] the more complex the task, and tasks that have to stay self-consistent and self-running for long stretches, it does really well on those. So whatever you do, don’t spend this quota on ordinary everyday tasks. Look at the gap in the numbers. First, on production-grade software engineering it hits 80%. Look at the others — even Opus 4.8 is only close to 70%. And the rest, GPT 5.5, the Gemini models, it far outruns all of them on

[02:33] that benchmark. And on the hardest thing, production coding, it’s more than double the capability jump over Opus 4.8. On long-horizon autonomous tasks it also reaches a level of more than 16 hours. So what kind of task deserves a model like this? Number one — say a complex project, a big project’s

[02:58] refactor, a cross-repo migration. The kind of thing where it has to read a huge amount of code on its own and run a long time against a task plan, and keep verifying itself over and over. That’s where this model most belongs. Another one: coding problems we ran into during development that earlier models simply couldn’t solve.

[03:23] You can pull Fable 5 out to go look at that old code. Also code audits on high-logic-risk things — charging, payments, billing, authentication — you can hand those to Fable 5 too. And those of you who don’t know how to code, the vibe-coding MVP product demos you built — you can have it do a framework-level

[03:50] re-selection of the stack and rebuild it into a production-grade product. That kind of refactor also suits this model. But our day-to-day output — copywriting, summaries, file cleanup, small edits, a bit of research — don’t waste this model’s quota on any of that. I’ve written a prompt here that you can just copy and use. I tell Claude Code that Fable 5 is a stronger model than Opus 4.8,

[04:16] that our subscription can use it free until June 22nd, and that to spend these two weeks’ quota where it cuts, I want it to spin up a multi-agent team and scan every project folder under all my working directories in parallel, plus memory, plus the items logged in my to-dos, and tag each project with its status and its relationship to my revenue — commercial projects that make money obviously come first — plus which step it’s currently stuck at, and a comparison against the remote repo.

[04:41] So first, a survey pass. Second, a priority screen: anything involving architecture design, big refactors, or audits and adversarial review of high-risk logic holes gets pushed up the priority list, while everyday tasks get weighted down. Then it does scheduling — within this week I want to max out that quota every single day, so it builds a shift plan of the tasks I’m handing it, and, and it also

[05:08] designs a done-standard for each deliverable, then lays out a checklist to work through one by one. And last, there’s an exception clause — some projects actually have deadlines and need finishing urgently. For those, some you can choose to run on Fable, some you can leave as they are, but either way they get handled first. So once you give it this skill, our Claude Code

[05:33] goes through every project on your plate and finds the places this model most deserves to be used — the high-leverage things, the refactors and reviews I just mentioned. Then we hand it the to-do list, and every day our quota goes where it cuts. Finally, a few small tips, because I did some digging too — some people burned through their quota very fast.

[05:58] The main reason is that when we use Opus we often crank thinking depth to Max, or to X-High, right? With the new model I’d actually start at High, because it proactively spins up subagents and stacks workflows on top of that. Its autonomous concurrency is very strong — not like Opus. If you really max it out

[06:23] and let it go at something, it’ll burn your quota to zero in an instant. So even for, even for a task like a code refactor, High or X-High is honestly enough. It’s enough. The other thing you’ll notice is that its safety review is still relatively very strict right now. Today I had it run a job downloading research papers, and it involved some

[06:48] pharma-related material — it refused, just refused, and downgraded to Opus 4.8 to run the job. So it really is like they said: they’re still worried a model this capable gets used for so-called illegal purposes. I get it. But I have a friend who’s in a rough spot, because his company is a biopharma company. Which means his, his company’s everyday —

[07:14] even their everyday work, or tool development, none of it can be done on this newest model anymore. Which is a real shame. That’s it for today’s share on the new model, Fable 5. I put out two videos today — I’m genuinely excited, this model is just too strong. See you in the next video. All right, see you tomorrow, bye.