Which model to choose when they all share one key

When every model is available through a single key, the choice stops being a technical question and becomes an economic one. Here is how to pick without overpaying.

Three tiers

Flagships. Claude Opus 5.5, GPT-5.6 Sol. They take the hard work: architectural decisions, long agent chains, reading an unfamiliar codebase, tasks where a mistake is expensive. At Zukko they cost $0.30/$0.30 per million tokens — more than the light models, and that is justified only where quality genuinely decides.

Mid tier. Claude Sonnet 5 ($0.30/$0.30), GLM 5.3 ($0.10/$0.10, 128K window, no vision). The workhorse: ordinary code, refactoring, tests, documentation. Whether you notice a difference from a flagship on your everyday work — test it on your own tasks.

Fast and cheap. GPT-5.6 Luna ($0.10/$0.10), GLM 5.3 Flash ($0.04/$0.04, 128K window, no vision). Classification, data extraction, short answers, bulk processing. Where you need volume rather than depth.

A practical rule

Start at the mid tier. Move up to a flagship only when the mid tier actually failed on a specific task — not in advance, "just in case".

That sounds obvious, but this is exactly where money leaks. Keeping Opus on autocomplete or on generating commit messages is the most common and most expensive habit.

What matters in agent work

An agent like Claude Code or Codex makes dozens of calls in sequence, resending the accumulated context each time. So what matters is the input token price rather than output, and the total is driven by context size rather than by how long an answer you get.

The second factor is the context window. On a large project, a model with a million-token window (Claude and GPT-5.6; GLM has 128K) saves you from constantly slicing tasks by hand. And a pricing caveat: vendors charge ~10% of the input price for prompt-cache reads, while Zukko bills cached tokens at the regular price, so for agents with a large context the real saving is smaller than a straight price-list comparison suggests.

What matters in chat

Here it is the reverse: few input tokens, many output tokens. Look at the output price and at how the model writes — that part is a matter of taste, and better tested on your own texts than read about in comparisons.

About benchmarks

Public leaderboards are useful as a rough orientation and useless as a basis for a decision. A couple of percent on a synthetic test says nothing about your task, and models are deliberately tuned for popular benchmarks.

A small comparison of your own is more reliable: take three to five real tasks from your work, run them on two or three models, and compare both result and cost. That is half an hour, and the answer is about you rather than about an average.

Switching

When every model sits behind one key, switching is a single line. In code you change the model field; in Claude Code, the slots in settings.json; in Cursor you pick the model in the interface.

Model prices and limits are served by /v1/models/info (or the Models page in the dashboard):

curl https://api.zukko.pro/v1/models/info -H "Authorization: Bearer fetch_your_key"

Zukko↗ currently has six enabled models: claude-opus-5-5, claude-sonnet-5, gpt-5.6-sol, gpt-5.6-luna, glm-5.3 and glm-5.3-flash. The set changes; the current list is GET https://api.zukko.pro/v1/models. All on one key and one balance, so experimenting does not mean opening new accounts.

The $0.50 bonus for subscribing to @zukkopro↗ is meant for exactly this kind of comparison.