Fable users report $130 prompts and quota drain

• yacineMTB reported one Fable prompt had already cost more than $130, then estimated the same usage pattern would run about $300 per day.
• Peter Gostev's effort table showed the same Three.js prompt moving from 73.2K tokens and 12 minutes on low effort to 367.1K tokens and 1h 50m on max.
• A ClaudeAI Reddit user said one day of partial Fable use consumed 96% of a weekly Fable quota, while bridgemindai said serious use pushed them to a third $200 Max subscription.
• The Fable safeguard notice says broad checks can flag routine coding, cybersecurity, or biology work and switch the session to Opus 4.8.
• BridgeBench's rerun reported debugging falling from 86.2 to 25.9, while Arena's preview said Fable's pre- and post-redeployment scores were mostly consistent across text, document, vision, and frontend arenas.

trq212's availability post points back to Anthropic's Fable 5 and Mythos 5 announcement, where Anthropic framed the rollout as capacity-constrained and safeguard-heavy. The weird bits are all in the receipts: a Fable fallback screen can appear on safe-looking prompts like “raspberry”, developers are packing context into PNGs through pxpipe, and one Reddit analysis found Claude Code subagents paying extra because static prompt blocks miss cache reuse in long fanout sessions.

The first cost story came from yacineMTB, who said a single Fable prompt had cost more than $130, then later said 12 hours of heavy use implied about $300 per day at the same intensity.

In replies, yacineMTB said the run had lasted about eight hours, with much of that time spent running training jobs on a server, and the same thread said the prompt was on high effort. A later reply described the workload as continuing a robotics project from a phone over tmux.

That is Fable's sharpest cost lesson: long agentic runs turn model pricing into a wall-clock tax.

Gostev ran the same Manhattan Three.js prompt through Fable at different reasoning levels and published the run table. The spread was not subtle:

1. Low: 73.2K tokens, 27 tool calls, 12m 0s, 711 lines of code.
2. Medium: 190.3K tokens, 58 tool calls, 44m 0s, 1,702 lines of code.
3. High: 296.3K tokens, 88 tool calls, 1h 15m 42s, 1,660 lines of code.
4. XHigh: 261.7K tokens, 165 tool calls, 1h 35m 30s, 2,291 lines of code.
5. Max: 367.1K tokens, 122 tool calls, 1h 50m 40s, 2,425 lines of code.

Gostev's thread linked the prompt and hosted comparison page, and the prompt itself asked for a complete procedural Manhattan with nested LOD, traffic, boats, landmarks, dusk lighting, and smooth laptop performance.

The run explains why Fable discourse keeps splitting in half. The model can build spectacle, and the meter keeps running while it does.

A Reddit user said Fable used only 11% of the tokens Haiku needed for one summary task, then still pushed the account to 96% of the weekly Fable quota after moving about half of one day's work to Fable.

The subscription math got stranger in user reports. bridgemindai's usage screenshot showed 56% of weekly Fable usage gone while the all-model weekly bar sat at 39%, and bridgemindai said they were rotating three separate $200 Max subscriptions because the weekly Fable limit died in about a day.

Anthropic's public availability message also made the cap explicit. Rohan Paul's summary of the current access window said Fable was included for Pro, Max, Team, and seat-based Enterprise subscribers only until July 7, with up to 50% of weekly subscription limits usable on Fable at no extra cost.

The fallback screen says Fable 5 runs automated safety checks on every user request and can fall back to a non-Mythos model such as Opus 4.8. The categories shown in the linked guideline screenshot are:

• Offensive cybersecurity techniques, including exploits, malware, or attack tooling.
• Most biology, chemistry, and life sciences queries, including lab methods or molecular mechanisms.
• Distillation attacks on Fable 5.
• A narrow set of frontier LLM development tasks, including distributed training infrastructure, accelerator design, and kernel work for some non-standard chips.

Users also found older routing behavior in logs. aibuilderclub_'s screenshot quoted a prompt being downgraded with the internal-looking label TOO_DUMB_TO_NEED_FABLE, while Sam McAllister's reply said that behavior had been rolled back and replaced by a visible fallback to Opus 4.8.

Sholto Douglas gave the same correction in a reply, saying Anthropic had reverted the earlier behavior and would “only ever fall back loudly.”

BridgeBench reported a large drop after redeployment:

• Debugging: 86.2 to 25.9, down 60.3 points.
• Refactoring: 73.6 to 38.4, down 35.2 points.
• Hallucination handling: 75.9 to 61.7, down 14.2 points.

bridgemindai's follow-up said the underlying model behaved like the June version when tasks reached Fable, and that the issue was the guardrail layer sending work to Opus 4.8. Another BridgeBench reply said only 3 of 12 debugging tasks went through when fallback scored as zero.

Arena saw a different shape. Arena's early score preview said thousands of votes across Text, Vision, Document, Code, and Agent arenas showed mostly consistent scores before and after redeployment, with the roughly 20-point frontend drop still inside the confidence interval.

The practical measurement problem is obvious from Wes Roth's formulation: once a classifier can hand work from Fable to Opus, a benchmark can measure a routed product rather than one raw model.

The dominant workaround is to spend Fable tokens on planning, judgment, and review, then push implementation into cheaper models.

The pattern showed up in several forms:

• Matthew Berman framed it as “Fable -> Planning” and “GPT-5.5 -> Execution.”
• Daniel Mac described “Fable advisor,” with Sonnet or Opus writing code while Fable handles high-leverage judgment.
• The fable-advisor plugin added GPT-5.5 as an implementer subagent through Codex CLI and included a directive for Fable to minimize its own token use.
• dzhng's skill post used Fable as planner, orchestrator, and reviewer, with Codex as implementer.
• Simon Willison said letting Fable choose lower-power subagents for coding tasks appeared to save tokens, with more notes in his Fable's judgement post.

This is the cleanest emerging Fable workflow: expensive model as foreman, cheaper models as crew.

Fable's cost pressure immediately produced compression infrastructure.

Rohan Paul's pxpipe explanation says the tool renders dense text into PNG pages, then sends those pages as image blocks. The cited numbers are concrete: a 1928×1928 image costs about 4,761 vision tokens and can hold roughly 92K characters, with the caveat that exact strings, hashes, IDs, and code details can be misread.

Condense attacks a different part of the bill. TestingCatalog's Condense summary says its proxy sits between a coding agent and the model, with Helene 1 stripping unnecessary tokens and Adeline 1 compacting settled loops to about 9% of their size. Condense's token-bill writeup reported roughly two-thirds bill reduction in tested sessions.

The cache layer has its own leak. A ClaudeAI Reddit analysis parsed about 95 local Claude Code sessions with 1,800 subagents and 6.8B input tokens, then modeled subagent prompt cost as about 14% too high because static context was billed at cache-write rates instead of read rates. The same post said a naive one-hour cache for subagents would make the modeled bill 8.6% worse, because 98% of cache reuse happened within about 34 seconds.