1. Specify the length of the answer
Output tokens are priced higher than input tokens on most models (e.g. Haiku 4.5 is $1 input vs $5 output). Adding a short instruction like "be concise," "in 3 bullet points," or "under 100 words" cuts unnecessary preamble and repetition, directly reducing output tokens.
2. Don't resend the entire chat history
In chat-style interactions, the full prior conversation is resent as context on every turn. Irrelevant tangents and trial-and-error exchanges pile up and get billed every single time. Start a fresh chat when the topic changes, or periodically summarize the history instead of carrying every message forward.
3. Summarize long documents before pasting them
Pasting an entire long document counts the whole thing as input tokens. Extract only the relevant section first, or ask the model to summarize the key points, then continue the conversation from that summary — this keeps total token usage far lower.
4. Match the model to the task difficulty
Simple classification, casual chat, or light rewriting is usually well handled by a cheaper model (Haiku 4.5). Save the higher-tier models (Sonnet 5, Opus 5) for complex design decisions, precise long-form translation, or multi-step reasoning. This single choice is the biggest cost lever available. TOKENSE automatically routes each question to the right model.
5. Cache repeated system prompts
Long, unchanging system prompts ("act as an expert in...") or reference documents you reuse across turns benefit from Anthropic's prompt caching: reused portions cost just 10% of the standard input price on a cache hit. The more a workflow reuses the same context, the bigger the payoff.
FAQ
FAQ
- Q. Does Japanese use more tokens than English?
- A. Yes — Japanese text tends to consume more tokens per character than English for equivalent content, which is a widely known characteristic of most tokenizers.
- Q. Does shortening a prompt hurt answer quality?
- A. Trimming unnecessary preamble and repetition rarely hurts quality — clarifying your ask can even improve it.
- Q. Does TOKENSE do any of this automatically?
- A. Yes — automatic chat history summarization/compression, prompt caching, and per-message model routing are all handled automatically. The two things worth doing manually are specifying output length and summarizing long documents.