The most common and most expensive mistake is running the flagship on everything. Here is how to choose deliberately.
The rule in one line
The flagship is for holding many relationships at once. Everything else is work for a smaller model.
If the task can be described as "take this and turn it into that", you do not need the flagship. If the task cannot be solved without understanding how the whole system fits together, you do.
Where Opus earns its price
- reading an unfamiliar repository to understand the architecture
- refactoring that touches several files coherently
- debugging races and intermittent failures
- design work that weighs trade-offs
- long agentic runs where the model chooses its own next step
Where Opus is overpaying
- parsing logs and traces
- rewriting and proofreading text
- generating boilerplate from a pattern
- classifying inbound messages
- format conversion, parsing, data normalisation
- simple questions about documentation
A smaller model handles these just as well for a fraction of the bill.
The two-model trick
A pattern that works: let the flagship think and a smaller model do the work.
Opus reads the task and produces a plan. A Sonnet or gpt-5.6-luna then executes the mechanics against that plan. You pay for expensive reasoning once rather than on every step.
There is no Haiku tier in our lineup. The cheapest model is gpt-5.6-luna, but note it is a different family. The request format is compatible, behaviour and prompts may differ, so check it against your own task.
It shows up most on batch work: define the approach once, apply it to a thousand items cheaply.
About input and output
One more thing people miss. Output costs several times input on every model.
So a long context in the prompt is almost always cheaper than a long answer. Do not fear giving context, fear asking for a verbose reply.
Checking against your own numbers
Abstract reasoning translates poorly into money. The site has a calculator: pick a model, enter volumes, see the number. Run the same scenario across tiers, the gap usually argues better than any prose.
In short
Flagship for tying context together, smaller model for transformations. Thinking is expensive and doing is cheap, so split those two jobs across different models.