When Anthropic shipped Fable 5, it was clearly the most capable model in the Claude lineup. It was also, at roughly five times Sonnet 5's per-token cost, the fastest way to burn through a usage budget if you routed every task through it. People noticed immediately, and Anthropic's own answer was not "write shorter prompts." It was a specific, documented pattern: use Fable 5 to plan, and hand execution to cheaper Sonnet 5 subagents. Done right, that gets you 92 to 96% of Fable 5's solo quality for roughly half the cost. That is what real token optimization looks like: not fewer words for the sake of fewer words, but the right model doing the right part of the job.
The problem with bloat
Tokens are the currency of AI. Every word you send and every word you get back costs money and time, but the bigger issue is that bloat degrades quality, not just cost. When people try to write "good" prompts, they often overcorrect into being verbose: repeating instructions, adding filler, including irrelevant context. That causes two real problems.
- The lost-in-the-middle effect. Models struggle to recall information buried in the middle of a long prompt. The more bloat you add, the more likely the model is to ignore something important.
- Latency and cost. More input tokens means slower responses and a higher bill, every single time you run the prompt.
Practical token optimization
How do you get more done with fewer words?
- Use structural markers. Markdown headers or XML tags parse far more efficiently than a paragraph of conversational prose saying the same thing.
- Drop the politeness. The model does not need "please" or "I would like you to." State the instruction directly.
- Use negative constraints where they are shorter. Sometimes "do not use class components" is more token-efficient than explaining the correct pattern in full.
- Route the task to the right model tier. The Fable 5 orchestrator pattern above is the clearest proof this matters at scale, not just per-prompt.
Why Kosmo saves tokens without trying to
Kosmo was not built to shrink prompts. It was built to remove ambiguity. But an unambiguous prompt happens to be a short one: no re-explaining, no hedging, no padding to cover a gap you were not sure the model would fill correctly. Kosmo compiles in only the documentation and constraints that are actually relevant to your specific request, structured the way your target tool parses fastest. The token savings are a side effect of precision, not a goal Kosmo is optimizing for directly, which is exactly why they hold up: you are not trading quality for brevity.
Frequently asked questions
Does writing a shorter prompt always save tokens overall? Not always. A prompt that is too short can cause the model to guess, get it wrong, and cost you a follow-up correction that uses more tokens than a precise prompt would have in the first place. Precision beats brevity.
Is the Fable 5 / Sonnet 5 pattern something I have to set up myself? If you are using Claude directly, yes, that routing is a manual decision. Kosmo applies the same underlying idea automatically: it only spends the more expensive reasoning where it actually changes the output.