Claude Fable 5 token limits, pricing, and context fit

What counts toward a Claude request?

A Claude request is more than the sentence typed into a chat box. The complete input can include system instructions, conversation history, retrieved documents, tool definitions, image or file content, and formatting added by an SDK. The response — and on Fable 5, the model's internal thinking — also needs room inside the model's request limits. Treating all of this as one shared budget is the safest way to plan a production workflow.

PromptCostLab gives you a fast local estimate for plain text, now backed by a real BPE tokenizer that runs in your browser. That is useful for comparing prompt versions and spotting context pressure without uploading sensitive text. For exact billing or requests containing tools and files, verify the final payload with Anthropic's official counting endpoint.

How many tokens does Fable 5 consume?

Two effects push Fable-era token counts above what older Claude models used. First, Anthropic's tokenizer refresh means the same text splits into roughly 30% more tokens on current-generation models than it did on Claude 4.5 and earlier. Second, Fable 5 always uses extended thinking, so every request also generates billable thinking tokens that arrive on the output side of the ledger.

For planning, count your input with a tokenizer-aware calculator, add your expected visible output, then add an allowance for thinking on complex tasks. If you migrate a workload from an older Claude model, re-measure rather than reusing old counts.

Fable 5 pricing at a glance

Input is $10 per million tokens, output is $50 per million, and cached input reads are $1 per million. Writing to the 5-minute cache costs $12.50 per million tokens (1.25x input); the 1-hour cache write costs $20 per million (2x input). Prompts that reuse a stable system prompt or long document prefix can cut effective input cost dramatically — repeated reads at $1 instead of $10 per million.

How to calculate a realistic token budget

Start with the stable instructions, add the average amount of retrieved context, include the recent conversation turns your application normally sends, and reserve a deliberate output allowance. Do not plan against the absolute published context limit. Leaving headroom makes retries, citations, tool results, and longer-than-average user messages less likely to break the request.

• Measure the full production prompt, not a shortened demo.
• Reserve output and thinking tokens before evaluating context fit.
• Test average and worst-case retrieval sizes.
• Record the model and pricing date used for the estimate.

Context fit and cost are different decisions

A prompt can fit comfortably and still be expensive at scale. Conversely, a very large context window does not mean every request should fill it. Calculate input and output cost separately, then multiply by expected daily request volume. This exposes whether the expensive part is repeated context, generated output, or the number of calls.

For RAG systems, the most effective optimization is often retrieval quality. Sending fewer, more relevant passages can improve answer quality while reducing token use. Long context is a safety net, not a substitute for selecting good evidence.

When local estimates are enough

Use a local estimate while drafting prompts, comparing versions, sizing documents, or selecting a likely model. Use provider-backed counting before setting hard production limits, showing a customer an exact bill, or sending structured requests with tools, images, and attachments. The label matters: a useful estimate should be transparent about what it includes.

• Drafting and prompt cleanup: local estimate.
• Model comparison and capacity planning: local estimate plus headroom.
• Billing reconciliation: official provider usage.
• Tool or multimodal payloads: provider-backed counting.

A practical Claude workflow

Paste the exact reusable prompt into the PromptCostLab Claude calculator, set a realistic output allowance, and note the context percentage. Next, repeat the calculation with the retrieved content your system usually adds. If the result is close to your operational ceiling, reduce repeated instructions, summarize old conversation turns, or retrieve fewer passages before changing models.