Guides & pitfalls
Practical material on AI API relay services: how rates are calculated, how to verify a model has not been downgraded, what enterprises should check before buying, and the pitfalls we ran into while measuring these services ourselves.
Browse by topic
All articles
Pinned API Compass is live: why we keep a relay service's claims apart from our own measurements
The relay industry's problem is not a shortage of directories. It is that nearly every directory copies its data straigh…
How to read the Q4 2026 model list: naming, tiers and three pricing traps
This wave of model names mixes generations, product codenames, tier suffixes and task markers. Read them as version numb…
Model retirement is not an accident: turning deprecation into a scheduled task
You open your laptop, get 404 model_not_found, and production is down. It is almost always an upstream retirement, not a…
RPM, TPM and concurrency: the four real shapes of relay rate limiting, and how to find the true ceiling
Rate limiting does not always return 429. Silent queueing, silent downgrades and mid-stream cutoffs report no error at a…
Video generation behind a relay: three API shapes, four billing traps, one acceptance checklist
Video generation has no shared protocol; every vendor ships its own asynchronous task flow. The shape decides how you in…
MCP through a relay: what works, what never will, and how to test it in ten minutes
MCP never touches the model API, so “does this relay support MCP” is the wrong question. Three tool-use behaviours decid…
How the relay industry may move in the last three months of 2026
Based on changes already visible, four possible directions for the next three months. This article works only from obser…
Three traps in choosing a site from a leaderboard alone
The leaderboard is not the problem; treating it as the only basis is. This article covers three specific traps and how t…
We probed 194 sites: what this data can and cannot answer
Our data has its own methodology and its own limits. This article sets them out, including which judgements it supports…
A relay selection decision tree: working back from your scenario to a site
Selection is not about which site is best, it is about what your scenario demands. This article turns it into a path you…
Evening peak jitter: upstream, network path, or your own concurrency
When the evening peak slows you down, do not blame the relay immediately. This article gives a three-step diagnosis that…
Logging and reconciliation: attributing token spend to the project level
If you cannot see where the money goes, optimisation is meaningless. This article gives a minimum viable call log design…
Supporting data & tools
The articles explain the method; the pages below give you measured data and prices to check it against. Reading them together is easier — once you know how to tell whether a rate is real, you can look the official price up in the price list.