Choosing a model with your eyes open

The dearest model isn't always the right one; sometimes you pay ten times as much for the same job.

6 minLesson 3 of 80 of 3 steps

The eternal temptation is to reach for the strongest model. But to detect the language of a sentence or classify a comment, a small model gives the same answer for a fraction of the price.

Start from the job, not the name

JobStart with
Classifying, extracting, short summariesA small, cheap model
Answering customers, writing, translatingMid-range
Hard code, multi-step reasoningA strong or reasoning model
Reading an image or documentA model with image or file input

What each model's page tells you

Open a model on the models page. These sections matter:

  • At a glance: context length, maximum output and speed.
  • Price & cost: the price per million tokens in rials, and any conditional prices.
  • Performance: time to first token and generation speed, each with its source.
  • Capabilities: image input, tools, reasoning and so on.
  • Peers: similar models to compare.

Context length is how many tokens fit into the model at once; it matters for long conversations and agents. For how the market compares, see Rankings.

Don't guess when comparing: run the same samples through two models and read each one's cost in the log.

If you'd rather not choose

Packs do this for you: one name for several models, such as iranrouter/auto:cheap, iranrouter/auto:balanced and iranrouter/auto:best, plus the others on the models page. Each is a list of models we update over time, and if one cannot answer the next takes over. You always pay the price of the model that actually answered, and its name appears in the log.

For work that must behave the same every time, a fixed model id is better, since an auto chain can change over time.

Compare two models0 of 3 steps

  1. Write one fixed question.
  2. Send it once to a cheap model and once to a dearer one.
  3. Compare the quality and the Cost column in the logs.

What you learned

  • Decide the job first, then the model.
  • Each model's page gives price, performance and capabilities.
  • Next: reading the logs together.
Reading Usage & logsEvery request leaves a trace here: which model, how many tokens, what it cost and, if it failed, whose side it was on.Done, next lesson