Skip to content
AI-Native PM
Field Notes

Field Note

Fable 5.1, First Impressions: Less Steering, Faster Spending

We gave Fable 5.1 a seven-chapter rewrite on its first day, with Claude Code's fleet mode on. This note covers what got better, why its estimate of its own run was short by a fifth, how a five-hour allowance ran out in 43 minutes, and how far the writing has come.

· 6 min read

Anthropic released Fable 5.1 with a launch note promising stronger long-running agentic work and cache reads at a quarter of the old price, and the Claude Code team's own post promised better writing with less of the flat house tone people call Claude-speak. We had a job waiting for exactly that: rewrite the first part of our course, seven chapters and seven slide decks, for a non-technical reader, in our voice and under our content rules. We ran it in Claude Code with Ultracode on, the mode that fans a task out to a fleet of agents instead of one. By the end of the first day the five-hour allowance on the top usage tier had run out in 43 minutes, and the week's allowance was gone before the job was.

Neither of us is an engineer, so this note is about the process rather than the code: how much steering the model needed, how well it estimated its own runs, how fast it spent, and how it wrote. The post we published that evening has the raw version. This is the version with the numbers checked and the story finished, because the run we were locked out of has since completed.

Fewer round trips, less babysitting

On the process, the launch note held up. Compared with Fable 5 on the same kind of work, it needed fewer corrections from us and finished more stages without a nudge. Here is what that looked like on this build. It wrote the checkers the rewrite needed (term ownership across chapters, reading-time honesty, cross-link text) before it wrote any prose, and it ran them itself after every edit. When a workflow crashed partway through an audit because some agents had returned nothing, it hardened the script against that case and resumed the run from its cache. The eighteen agents that had already finished were replayed instead of paid for again.

Its estimates of its own runs are still short

We ask every frontier model the same thing before it launches a fleet. In plan mode, before anything runs, it has to state what it will do, how, how many agents it plans to launch, and how many tokens that will take. Then it runs a pilot, so we can compare the estimate with the bill. In our experience the estimate is usually short by 30 percent or more. Fable 5.1's first paid run, an audit of the one chapter we had written as the exemplar, used 3.23 million tokens against its own estimate of 2.5 million, so the estimate covered 77 percent of the bill. The pilot chapter after it used roughly double its projection. The model then proposed a leaner audit and a second pilot to measure it, which is where the allowance ran out and the post went up. That second pilot finished after the reset: 1.75 million tokens per chapter against a projection of 1.5 million, so the estimate covered 86 percent.

The estimate against the bill, three runsThree pairs of bars in millions of tokens. The first paid run, the exemplar audit: estimate 2.5, bill 3.23, so the estimate covered 77 percent. The pilot chapter through the full pipeline: estimate 1.4, bill 2.72, 51 percent. The second pilot with the leaner audit, per chapter: estimate 1.5, bill 1.75, 86 percent. Caption: each pilot brought the next estimate closer to the bill.THE ESTIMATE AGAINST THE BILLits estimatewhat it used0M1M2M3M2.5M3.23Mfirst paid runthe exemplar auditcovered 77% of the bill1.4M2.72Mpilot chapterfull pipeline, lean auditcovered 51% of the bill1.5M1.75Msecond pilotleaner audit, per chaptercovered 86% of the billEach pilot brought the next estimate closer to the bill.

The routine that produced those numbers, and what we changed each time the estimate missed, is its own note: How We Budget an Agent Fan-Out Before We Run It. The short version is that the cost per agent held steady across every run, and what the estimates got wrong was how many agents the audits would spawn.

The speed is real, and so is the meter

With Ultracode on, Fable 5.1 fanned out more readily than any model we have run. On one session it picked back up after an interruption, and it launched six agents just to set up and check the environment before touching the task. We had not seen Opus 5 or Fable 5 do that. We do not know whether that is the mode or the model, and we will keep watching. What we do know is the arithmetic. A fleet finishes the job faster because six agents work at once, and it spends the allowance faster for the same reason. Our five-hour window on the top tier was empty after 43 minutes, and the work then waited more than four hours for the reset. Finishing in 43 minutes is not much use when the next 257 are spent waiting.

One five-hour window, spent in 43 minutesA horizontal track five hours long, marked at each hour. Above the first 43 minutes, six short agent lanes stacked in parallel, labeled the fleet: six agents at once, each lane one agent on its own slice of the job. The track's first 43 minutes are filled clay and labeled allowance spent; the remaining four hours and seventeen minutes are dashed and labeled waiting for the reset. Caption: a fleet finishes faster and spends faster, and a fixed allowance runs out on the rate, not the price.ONE FIVE-HOUR WINDOW, ON THE TOP TIERthe fleet: six agents at onceeach lane is one agent on its own slice of the job01 h2 h3 h4 h5 h43 min: allowance spent4 h 17 min: waiting for the resetA fleet finishes faster and spends faster; a fixed allowance runs out on the rate, not the price.

On price, Anthropic says the cheaper cache reads cut a typical workload by about a quarter and a heavily agentic one by up to about 45 percent. We cannot confirm that from one day, and it would not have changed the day. A subscription allowance runs out on the rate you consume, whatever the per-token price, and a fleet consumes at fleet rate. We will know what the new model actually costs us after a few weeks of usage, and we will write that down when we do.

Fable 5.1 does more of the work with less steering, and it spends a fixed allowance at fleet speed. On a subscription, the allowance is the number to plan around, not the price.

Claude Code now resumes an interrupted run by itself

We also met one feature we had not seen before, and it was the best part of the day. When the allowance ran out mid-run, Claude Code showed a panel that read "Out of usage credits, auto-resuming at 7:21 PM," with a box ticked for "Auto-continue when limits reset." Before this, one of us set an alarm for the reset and pinged the session from a phone with "start from where you were interrupted." Now it starts itself. One caution from our run: a run that stops on the limit can report on its way out, and that report is not a result. One of our workflows died on the limit and reported zero findings, because every agent inside it had failed, and a zero from a run that never worked means nothing. Whether a run resumes itself or you restart it, read what it returns before you trust it.

Writing: a marginal improvement, and not yet the voice

The text-heavy task was the part we were most excited about. We asked for the rewrite in our voice, under our content rules, against chapters that Opus 4.8 had written earlier. The chapters it returned were marginally better than the versions they replaced, and the run used more tokens and more agents to produce them. Here is chapter 1's one-line description, before and after, so you can judge the margin yourself.

Opus 4.8:

When an AI coding tool prints a plan full of names you do not recognize, work out what each line is really deciding, so you can approve it or push back.

Fable 5.1:

When an AI coding tool prints a plan full of names you do not recognize, work out what each line decides, so you can approve the plan for a reason you can say out loud.

The second is tighter and ends on something a reader can do. It is also one line, and the difference across a whole chapter was about that size. Every chapter still went through a line-by-line edit after the fleet finished, and that edit is where most of the voice came from. The house tone has improved by about that much, and it has not gone away.

What we do differently now

  • Plan around the allowance, not the price. On a subscription, a fleet's speed is the cost. Before a fan-out, we now ask how long the five-hour window will last at fleet rate, and we launch the long runs at the start of a window.
  • Ask for the estimate in agents, not tokens. The cost per agent held; the agent count did not. The pilot's job is to count agents.
  • Let it resume itself, then read what it returned. Auto-continue saves the alarm. It does not check the run.
  • Keep the edit. The fleet gets a draft to marginally better. The voice still comes from the pass after.

Sources