Fast mode makes a coding agent's model write up to 2.5 times faster, but the agent does not finish 2.5 times sooner. Your transcript tells you how much sooner, before you pay double per token.
Engineer's view. This is the upgrade you made to a database that only cut the API's time by a tenth. Most of each request was spent elsewhere. A faster model only speeds up the part where it writes tokens. Tests, builds and searches run at the same speed, so you pay double for part of the time.
TL;DR. A faster LLM makes a coding agent faster only on the time the model spends writing tokens. Tests, builds and the wait for the first token stay the same, so measure that share in your own session before you pay for speed.
Explain Like I'm 5
You and a friend bake cookies. Your friend mixes the dough, then the cookies bake in the oven for ten minutes. Then your friend mixes the next batch.
Now your friend learns to mix twice as fast. The cookies still bake for ten minutes each time. So the whole afternoon gets a little shorter, not twice as short.
And if your faster friend makes a bigger mess and has to mix every batch twice, you save nothing at all.
The Problem
You have shipped this bug before, and it had nothing to do with AI.
An API endpoint was slow, so you moved its database to a bigger instance. The queries got twice as fast. The endpoint got 12% faster. Then you opened a profiler. Most of each request was spent waiting on a payment service, and you had paid for the wrong part.
A coding agent is that endpoint. Each turn, the model reads the conversation, waits for a first token, then writes its reply. Then a tool runs: a file read, a build, a test suite. Then the model goes again.
Faster models are now sold on that middle step. Claude Code's fast mode runs Opus 5.5 at up to 2.5 times the output speed for twice the price per token. Providers race on tokens per second. But a faster model shrinks only the time spent writing. Your tests run at the same speed.
There is a second catch. A fast model that writes more tokens can lose its lead. In a Latent Space test, a coding model ran near 1,000 tokens per second. Its median reply came three times sooner than its rivals' replies. On hard questions it wrote nine to eleven times more reasoning tokens, and it finished later.
The fix: measure how your agent's time splits between writing and everything else, and predict the saving before you pay for speed.
Read the full post on The AI Commit
How to Predict the Time a Faster Model Saves Your Coding Agent carries the rest: the fix and why it works, the code to change for each role, how to tell it worked, when it is the wrong tool, and an interactive model you can play with.
Originally published at theaicommit.com, a daily AI engineering lab for software engineers. The original has an interactive version you can play with, and the full runnable code is on GitHub.
Top comments (0)