Detailed Analysis
A Reddit user working on a project identified as "Fable/Cowork" documented a striking pattern of escalating token consumption when using Claude across multiple code or file generation sessions, culminating in a single request for a 4kb file consuming 44% of the platform's daily usage allowance. The progression revealed in the post tells a consistent story: an initial context-feeding session plus six files (~40kb) consumed 40% of daily limits; a second batch of comparable size took 35%; a third batch of 36kb across five files exhausted the remaining 25% and hit the 100% cap. After a five-hour pause — presumably waiting for the limit to reset — the user issued a "resume your work from where you left" prompt to generate only the final file, triggering the disproportionate 44% charge.
The explanation for this behavior lies in how large language models like Claude handle conversational context. Claude does not retain memory between sessions natively; instead, it relies on the full text of the conversation being passed back into its context window with every new prompt. When a user invokes "resume your work," the entire prior exchange — including all the original context-feeding text, all previously generated file content, all instructions, and all model responses — must be re-ingested as input tokens before Claude can produce even a single line of new output. The file being *generated* may only be 4kb, but the *input context* driving that generation could represent many tens or hundreds of kilobytes of accumulated conversation history. Token-based billing and usage metering counts both input and output tokens, so a tiny output atop a massive input context produces exactly this kind of lopsided consumption.
This phenomenon highlights a fundamental architectural tension in using Claude or similar LLMs for long, multi-session, multi-file project work. Each resumed session does not start "fresh" from a checkpoint — it restarts from the beginning of the accumulated context, meaning costs and usage scale with the *total length of the conversation*, not with the size of the incremental task. The user's workflow of batching files across multiple prompts actually compounded this problem: by the time the final file was requested, the context window contained not just the original instructions but also the full generated content of the previous 17 files across three prior prompts.
To avoid this outcome in future workflows, the most effective strategy is conversation segmentation — starting a new, clean conversation thread for each distinct phase of work rather than attempting to resume a single long thread. Users can manually copy forward only the minimal necessary context (e.g., a brief project summary and a specification for the remaining task) rather than relying on the model to reconstruct state from a full conversation history. Structured project documents, external file management, or tools like Claude's Projects feature (which maintains summarized memory more efficiently than raw conversation replay) can also reduce the input token overhead substantially. Essentially, the goal is to keep the *input context per prompt* as lean as possible, decoupling it from the cumulative length of all prior work.
The broader trend this episode reflects is the growing awareness among power users of Claude that token economics matter enormously in agentic or multi-step workflows. As Anthropic positions Claude for longer-horizon tasks — code generation, document drafting, multi-file projects — the gap between user intuition ("I just need one more small file") and underlying model mechanics ("but I must re-read everything first") becomes a significant friction point. This is driving demand across the AI industry for better stateful memory architectures, caching mechanisms for repeated context, and clearer usage transparency so users can anticipate costs before issuing a prompt rather than after depleting their daily allotment.
Read original article →