Compare AI models side by side

How to read a model comparison without being misled

Input price is the least useful number in the table

Providers advertise input cost because it is the smaller of the two, and output tokens are typically priced three to five times higher. What you actually pay depends on your ratio: a summariser sends a lot and returns a little, so input dominates; an agent that writes code inverts that completely. Read both columns together, weighted the way your own work is weighted, and treat any single headline figure as marketing.

A context window is a ceiling, not a working range

A million-token window means a million tokens fit, not that quality holds across them. Retrieval accuracy degrades well before the limit on every model, and you pay for every token you send on every turn — a long conversation re-bills its own history. Treat the window as headroom that saves you from chunking, and assume real quality lives comfortably inside it.

Benchmark gaps under a few points are noise

Published scores come from different harnesses, different prompt scaffolds and often different attempt budgets. A two-point difference on a coding benchmark tells you far less than a five-times price difference or a latency difference you can feel. Use scores to rule models out, not to rank the survivors.

Speed labels and real latency are different things

Throughput — tokens per second once a model starts — is not the same as time to first token, and reasoning models can sit silent for a long time before either. If the model sits in an interactive loop, the wait before the first character is the number that decides whether people keep using it.

A shortcut for the common case

Compare three models, not ten: the one you use now, the obvious upgrade, and the obvious cheaper option. If the upgrade is not clearly better on the axis you care about, stay put. If the cheaper option is within a few points on that axis, run a week of real work through it before assuming you need the expensive one. Every comparison here is shareable — the model selection travels in the URL.