Detailed Analysis
A user on the Anthropic subreddit has documented a frustrating experience with Claude's behavior on Fable Extra, a platform that integrates Anthropic's Claude model, in which a large-scale prompt consumed 100% of the platform's usage allocation while delivering only a fraction of the expected output. The user submitted an 18KB, 440-line prompt that required Claude to read nine files totaling approximately 350KB in order to conduct a multi-section audit. Claude generated only one of eleven requested sections before an orange error banner appeared reading "Claude wasn't able to finish this answer." When the user attempted to resume the task with a follow-up instruction, the remaining 21% of their usage limit was exhausted without producing meaningful additional output. The end result was a complete consumption of available usage credits with roughly 9% of the intended deliverable completed.
The incident raises important questions about how usage limits interact with long-context tasks in third-party integrations of Claude. The user's frustration centers on a perceived gap between what the platform appeared capable of handling — given that it accepted the prompt and began processing a substantial volume of context — and what it actually delivered. The error appears to stem from Claude hitting an output token limit or some platform-imposed constraint mid-generation, rather than a failure to process the input. This distinction matters because the user was charged for context ingestion and partial generation without receiving a usable artifact. When the user sought redress through Fable's chat support, they were told that refunds are prohibited under the platform's terms of service, even in cases involving technical errors or instability, and were advised to restructure the prompt into smaller segments going forward.
The situation reflects a broader challenge in deploying large language models for complex, multi-part tasks: the gap between context window capacity and reliable output generation. While Claude's context windows have expanded significantly — allowing models to ingest hundreds of thousands of tokens — the ability to sustain coherent, complete output across very long generation sequences remains a distinct and sometimes separate limitation. Users who conflate input capacity with output reliability often encounter exactly this type of failure. The Fable platform's lack of transparent communication about output length constraints, combined with its apparent acceptance of the full prompt before failure, likely contributed to the user's expectation mismatch.
The refund dispute highlights an underexplored commercial tension in the AI platform ecosystem: when a model or integration fails mid-task due to a technical constraint rather than user error, the question of who bears the cost of that failure remains largely unresolved. Most platforms, including those built on Anthropic's API, structure their terms of service to treat consumed tokens as non-refundable regardless of whether the output was useful. This policy may have been designed with straightforward misuse scenarios in mind but becomes harder to justify in cases where a platform-side limitation — rather than a flawed prompt — causes the failure. As AI tools take on increasingly complex, long-horizon tasks, the industry will likely face growing pressure to develop more nuanced policies around partial delivery failures.
For users attempting to replicate this type of large-scale audit task, the practical implication is clear: breaking complex prompts into modular, independently executable segments is currently the most reliable mitigation strategy, even if it introduces significant friction. The user's question about whether the existing partial output can serve as a foundation to recover costs through continued prompting is technically reasonable in principle — Claude can in many cases resume from a prior context — but whether Fable's platform and usage accounting would allow that without consuming a new full allocation is uncertain. The incident ultimately underscores that the usability ceiling for Claude-powered platforms on complex professional tasks is still significantly shaped by platform-specific implementation choices, not solely by the underlying model's capabilities.
Read original article →