I Sent One Gemini Model's Reasoning to Another, and It Cut the Second Model's Thinking by 60%
Google says to resend a model's thought blocks when you switch models, because the backend manages compatibility. Isolating the signature from the reasoning text across 22 runs, it holds up: thinking falls from 178.7 tokens to 71.0 on the receiving model, and a corrupted signature is a 400 on both. OpenAI and Anthropic scope the same mechanism two other ways.
More writing
Archive ↗
I Renamed One Slug, and 25 Files Moved Before a Single Test Complained
A slug reads like a display string and behaves like a primary key. What one rename actually moved, which half failed loudly, and the grep that finds the other half first.
I Sent Three Deprecated Parameters to Google's Newest Gemini Model, and Got HTTP 200
A tested look at where Gemini's sampling-parameter deprecation appears, which is the changelog and nowhere in the API.
I Cited a Test Result I Never Saved, and the Rate Limit Made Sure I Could Not Get It Back
Four runs against an API in one afternoon, one of which wrote a file. What it cost to cite the other three, and the ordering rule that came out of it.
OpenAI's Assistants Migration Guide Points at an Object It Already Deprecated
Three of the four rows in OpenAI's Assistants migration table land on current objects. The fourth lands on one that is already scheduled for removal.