In July, OpenAI released GPT-5.6 as three separate models: Sol, the flagship; Terra, for everyday work; and Luna, the fastest and cheapest. Anthropic sells Claude in tiers too, and Google does the same with Gemini. The message from every provider is the same: not every task needs the biggest model.
Most people ignore this and use the most powerful model for everything. That's the AI equivalent of taking a taxi to the mailbox.
How Big the Price Gaps Are
The differences are not small. At launch, GPT-5.6 Sol cost $5 per million input tokens and $30 per million output tokens; Luna cost $1 and $6. On July 30, OpenAI cut Luna's price by a further 80% and Terra's by 20%, while Sol stayed the same.
There's a trap, too: a plain GPT-5.6 API call goes to Sol, the most expensive tier, unless you pick another one. If you never choose, you always pay top price.
A Simple Way to Match Model to Task
Use the small model for mechanical work. Reformatting text, extracting fields, tagging, translating short passages, cleaning up lists. The answer is either right or wrong, and small models get these right.
Use the mid-tier for everyday writing and analysis. Drafting emails, summarizing documents, answering questions about a report, first drafts of most content. This is where most daily work lives.
Use the flagship for hard reasoning. Multi-step problems, complex code, strategy questions with many moving parts, anything where a subtle mistake is costly. This is what you're paying the premium for.
A quick test: if you could check the answer in seconds, a smaller model will probably do. If checking it would take real expertise, go bigger.
Better Prompts Make Smaller Models Viable
Here's what most people miss: much of the gap between a small and a large model is about how much the model has to guess. Flagship models are better at working out what you meant from a vague request. Smaller models need you to be explicit.
That means a well-structured prompt often lets a cheaper model match the flagship's result on routine tasks. Give it the role, the context, the format and the constraints, and there is far less left to guess.
This also protects you when things change. In June, access to Anthropic's newest flagship models was suspended for three weeks because of US export controls. Teams with clear, model-independent prompts could switch to another model and keep working. Teams whose workflows depended on one model's quirks had to start over.
The Takeaway
Pick the model on purpose instead of by default. Small for mechanical, mid-tier for everyday, flagship for the hard problems — and write prompts clear enough to work on any of them.
IacuWise optimizes your prompt for the model you choose, so you can move routine work to cheaper tiers without losing quality.