In partnership with
Four words into my prompt, Ultracode was already forking itself: "clean up the checkout flow, it's a mess." No file list, no scope, nothing else. A few minutes later my monthly quota dashboard read a flat 0%, no warning screen, just gone. I'll tell you what actually happened, and where the same setting earned its price back, at the end.
This week
Claude Fable 5 came back online July 1st, three weeks after a US export-control suspension pulled it offline. Within 48 hours the reaction split into two camps: people pointing to a demo that reportedly cost $173 in tokens for a playable Rocket League clone, and people trying to turn the same model into revenue without burning a month of credits before lunch. I meant to land in the second camp. Instead I flipped on Ultracode, gave it that one vague line about the checkout flow, and watched it fork subagents into payment logic, the shipping calculator, and a settings page I hadn't touched in a year.
Ultracode isn't a reasoning-depth setting from the API. It's specific to Claude Code: it pushes every request in the session to xhigh effort and gives Claude standing permission to spin up parallel subagents on anything that looks substantial. On the API, Fable 5 runs $10 per million input tokens and $50 per million output, roughly double Opus 4.8. That premium is fine when one model does the work. It stops being fine when Ultracode quietly turns one request into twelve running in parallel, each burning its own share. I wasn't the only one this happened to: a Max subscriber reported burning 20% of a week's quota in a day, a Pro user hit their cap in ten minutes, one builder launched 62 subagents on a single task and hit the five-hour cap in 18 minutes (scattered reports, not a measured average, but the pattern repeats).
The part that actually changed how I write prompts: Fable 5 wants less from you, not more. Every prior Claude model rewarded longer, more detailed instructions. Load Fable 5 with hyper-specific step-by-step constraints and edge cases spelled out in advance, and the output often gets worse, not better. Anthropic's own documentation backs this up: Fable 5 at low or medium effort frequently beats older models running at their maximum setting, and the advice is to start at "high" (the default) and escalate only when a specific task measurably needs it. Each subagent Ultracode spins up inherits whatever effort level the session is set to, so cranking to xhigh by default means paying that premium once per subagent, in parallel, for however many Ultracode decides to launch.
Tomorrow morning, before you touch Ultracode again: write down every file the task should touch and the ones it shouldn't, in the prompt, before you turn it on. That single line of scope is what separated my 75% quota burn from the one Ultracode run that actually paid for itself: an audit of a partner API integration, fully scoped in advance, that shipped three real fixes before I'd finished my coffee.
Scale AI support on AWS, see how July 9
Customer expectations keep rising. Support budgets don't. On July 9, Fin and AWS are hosting a live executive session on how leading enterprises close that gap: scaling AI-powered support while simplifying how they buy it.
You'll see how to resolve an average 76% of conversations with Fin on AWS enterprise-grade infrastructure, procure through AWS Marketplace to put committed cloud spend to work, and turn the Fin and AWS collaboration into lower support costs. Register for the live session to see how.
In the news
Anthropic's own announcement confirms Fable 5's return comes with a catch: on Pro, Max, Team, and select Enterprise plans, the model is included at no added cost for up to 50% of your normal weekly usage limit through July 7. After that, it moves to metered usage credits billed at standard API rates, on top of whatever plan you're paying for. Anthropic says the restriction is temporary and capacity-driven, not a permanent removal, though no restoration date has been given. Check your own usage dashboard this week rather than assuming Fable 5 stays inside your included quota.
This week on the blog
- I Found Claude Fable 5's Best Use Case. It Cost 75% of My Monthly Quota to Get There.: the full story behind this week's edito, including the one job where Ultracode actually paid for itself.
- The Claude Code YouTube channels worth your time in 2026: where to actually learn the tool instead of burning quota on trial and error.
- WordPress Is Dead. The Replacement Has No Mechanic.: what happens to maintainability when every site is a one-off, AI-coded build.
Revenue Stack: July 2026
The lesson that survives whether or not Fable 5 stays in your subscription: scope the task before you hand it to a model that can fork itself twelve ways. That discipline is the whole difference between a tool that pays for itself and one that quietly eats your budget while you're not looking.
Have you tried Claude Code's Ultracode mode yet?
Phil
PS: reply and tell me what happened the first time you gave a model more autonomy than you meant to grant it.
Back to that flat 0%. Ultracode didn't fail by delegating to twelve subagents. I failed by never telling it what it was allowed to delegate to. Same setting, same model, and the only thing that changed between the burn and the payoff was one sentence of scope written before I hit enter.
You're receiving this because you signed up at rentierdigital.beehiiv.com.