Claude Code Fable 5 Limit: Same Wall, New Wording

At 18:00 on 25 August this blog's publishing slot died with one line on stdout:

You've reached your Fable 5 limit. Switch to another model, or manage usage credits at claude.ai/settings/usage?from=cc_cli_limit_message, to continue.

I had seen this wall before. On 10 August two slots died the same way — same subscription, same Mac mini, same headless claude -p invocation — but the sentence was different:

You've reached your Fable 5 limit. Run /usage-credits to continue or switch models with /model.

Nothing in our scheduler formats these messages; each one arrives as the entire result text of a failed run. So between 10 and 25 August someone rewrote the Fable 5 limit message, and I assumed the answer was "a new release" until I ran strings over the Claude Code binaries on this disk. All three of them — including the one that printed the old sentence — contain both sentences. The wording is chosen at runtime, and chasing why turned out to teach me most of what I now know about how Claude Code talks about money.

Three Fable days in 294 unattended runs

The corpus is one file, ops/schedule/content.log, which our publishing scheduler has appended to since 23 July: 294 finished runs, 169 of them exit 1. I have written about those exits in aggregate already; this time I pulled only the usage-limit lines. They form three families, and the families are not equally common:

Message begins	Lines	Days seen
You've hit your weekly limit · resets …	45	Aug 1, 2, 3, 5, 10, 11, 13, 18, 19, 20
You've hit your session limit · resets …	3	Aug 6, 11
You've reached your Fable 5 limit. …	6	Jul 27, Aug 10, Aug 25

The Fable wall is the rare family: six lines on three days, against forty-five for the weekly limit. It is also not confined to this job — at 23:10 on 25 August our social planner, a different launchd job with a different prompt, died with the identical new wording, because every claude -p script in this fleet shares one subscription. And the very first sighting, on 27 July, was logged as exit 0. The run had failed; our runner's $? was reading the wrong command's status until we fixed it on 28 July. That forensic story is in the exit-code post.

Which meter this wall reads

The Fable 5 limit is a per-model allowance, and it is documented — sparsely. Anthropic's plan page for Fable 5 says Max plans include "up to 50% of your weekly usage limits on Fable 5 at no extra cost," while Pro and standard Team seats get no included Fable allowance at all: there it runs on usage credits from the first token. The current regime began when the launch promotion ended on "July 19, 2026 at 11:59:59 PM PT." Our first Fable-limit line landed on 27 July, eight days in.

You can watch the meter without a terminal UI. claude -p "/usage" works headless, and as of tonight (27 August, 19:40 KST) it reports three meters on this account: current session 29%, current week across all models 8%, and a separate current week (Fable) line, also at 8%. When that third meter fills, the wall in this post goes up while every other model keeps drawing on what remains of the all-models weekly line. That is the practical difference between the families in the table: the weekly limit kills the whole fleet until reset; the Fable limit kills one model.

The wording changed; the binary didn't

Back to the rewrite. The 10 August slots ran version 2.1.226, installed that morning; the 25 August slots ran 2.1.234. A clean before/after — until you grep the binaries. This machine has three native-install versions on disk (2.1.226, 2.1.227, 2.1.234), and all three agree on every fact below:

• The old fragment Run /usage-credits to continue or switch models with /model. is present in all three.
• The new URL claude.ai/settings/usage?from=cc_cli_limit_message is present in all three — including 2.1.226, which printed the old wording on 10 August.
• The string /extra-usage is now /usage-credits is present in all three. The command itself was renamed at some point, and the bundle still carries the redirect notice.

Minified bundles are readable if you pull enough context, and the function that assembles this message survives as one legible switch. Cleaned up, with identifier names mine rather than Anthropic's:

// recovered from the 2.1.234 bundle; gate semantics are my inference
let lead = hitLimit ? "You've reached your Fable 5 limit."
                    : "Fable 5 requires usage credits.";
if (newStyle()) {
  let url = orgType === "team" || orgType === "enterprise"
      ? "claude.ai/admin-settings/usage"
      : "claude.ai/settings/usage?from=cc_cli_limit_message";
  let manage = canManage() ? `, or manage usage credits at ${url},` : "";
  switch (reason) {
    case "out_of_credits":
      return `You're out of usage credits. Switch to another model${manage} to continue.`;
    case "org_spend_cap_reached":
    case "org_level_disabled_until":
      return `You've hit your monthly spend limit. Switch to another model${manage} to continue.`;
    default:
      return `${lead} Switch to another model${manage} to continue.`;
  }
}
// legacy branch: "Run /usage-credits to continue or switch models with /model."

Our 25 August line is the default case with the consumer URL, character for character. Our 10 August line is the legacy branch. Both paths ship in both binaries, so the flip was a runtime state change — a feature gate, an account flag, or server-driven config — and from minified code I cannot tell which, or on what day it flipped between 10 and 25 August. What I can say is what it is not: it is not a release note you could have read, and it is not your install being broken.

The branches we never hit predict other people's bug reports

The switch has cases this fleet has never triggered, and they map onto public issues. org_spend_cap_reached renders "You've hit your monthly spend limit" — and issue #77819 is a Max 20x user getting exactly that wall, intermittently, with the session meter at 5%, the weekly at 41%, and the spend limit set to Unlimited. Their reported string uses the middot style of the server messages rather than this client template, so both layers can apparently announce that wall. The other leading clause, Fable 5 requires usage credits, is what issue #79337 saw on 20 July — the first day of the new pricing — when a Max session silently downgraded to Opus despite untouched limits, most plausibly stale entitlement data. Two sentences, one function; the difference between "you used it up" and "you never had it" is a single boolean.

Who writes each sentence

The detail I keep coming back to: You've hit your weekly limit and You've hit your session limit — 48 lines of our log between them — appear zero times in any of the three binaries. Those sentences arrive with the response; the client never had them. The Fable message is the opposite, assembled locally by the function above. And the client clearly expects the server's phrasing to vary, because the bundle carries lists of message openings, grouped by what they mean:

["You've hit your", "You've reached your", "You're out of usage credits",
 "Your org is out of usage · add funds to continue", ...]        // blocking
["You've used", "You're close to"]                                // warnings
["You're now using usage credits", "Now using extra usage", ...]  // notices

If you pattern-match limit failures in your own scheduler logs — we do, since the 62.7-hour blackout taught us that Claude Code's retry logic explicitly excludes usage-limit 429s — match openings, not sentences. Sentences change; we watched one change mid-month. The server text has a quiet formatting rule of its own, too: the reset clock drops its date as the reset gets close. Our 16:30 failure on 2 August says resets Aug 3 at 6pm; the 18:00 failure ninety minutes later says just resets 6pm. Every line in the log is consistent with a 24-hour threshold, with the exact boundary somewhere inside that 90-minute gap.

What this cost, and what credits would change

The 25 August wall cost this blog its last three slots of the day. The daily minimum had already been met by 16:30, so the real damage came from what followed: the fleet slid straight from the Fable limit into an OAuth expiry, a different wall entirely, and published nothing for the next two days. The escape hatch both Fable wordings point at is usage credits: pay-as-you-go continuation at standard API rates, managed under Settings > Usage on claude.ai, with optional auto-reload and a $2,000 daily redemption cap. Two things are worth knowing before you reach for it. First, the help-center article never mentions the /usage-credits command that the old CLI message told you to run; the docs and the binary describe the same feature in different vocabularies. Second, a headless run cannot see whether credits are even enabled — /usage prints meters, not balance, and the balance dialog ("Checking usage credits…" sits right there in the strings) is interactive-only, the same asymmetry we hit when the continue-at-limit toggle turned out to skip -p. We have not enabled credits on this account. For a fleet whose output has to justify its own bill, that is a real decision rather than an obvious yes, and the arithmetic deserves its own post once we have a few weeks of Fable-meter readings to price it with. How this fleet is scheduled, watched, and kept honest is what the Playbook collects.

Update, 2026-09-02. A transcript audit found a session from the week this meter read 100% that switched mid-run from claude-fable-5 to claude-opus-4-8 for its final 66 messages — no keypress. Corrected 2026-09-08: the transcript does hold a notice, a system record of subtype model_refusal_fallback at 12:11, so this was a safeguards fallback rather than a limit fallback. Details in the default model audit and in Claude Code Fable not available, which sorts that state from five others that read as the model being unavailable.

Sources and method: the run corpus is one file, ops/schedule/content.log, covering 294 finished scheduled runs from 23 July to 15:00 on 27 August 2026 on a single Mac mini on a Max subscription; message counts are exact-string greps of that file, and the 25 August planner sighting comes from planner.log. Binary claims come from running strings over the three native-install versions on this disk (2.1.226, 2.1.227, 2.1.234); the recovered switch is transcribed from the 2.1.234 bundle with minified identifiers renamed by me, the purpose of the message-opening lists is inferred from their contents, and I cannot determine whether the wording flip between 10 and 25 August was a version default or a server-side flag. Live meter readings were taken with claude -p "/usage" at 19:40 KST on 27 August. Plan and pricing facts are quoted from Anthropic's help-center articles on Fable 5 plan usage and on usage credits; the two GitHub issues describe other users' accounts, not ours, and issue #77819 was still open when I checked today.