Price is only half the question
An agent needs more than a cheap model — it needs a model that follows multi-step instructions, uses tools correctly and recovers from mistakes. A model that fails those tasks costs more in retries than a slightly pricier one that gets them right.
The usual candidates
| Model class | Price feel | Fit for agent work |
|---|---|---|
| deepseek-chat | Very low | Good general agent; the demos on this site run on it — see the maps build |
| GPT-4o-mini-class | Low | Solid tool use; the default choice in many frameworks |
| Local models via Ollama | $0 per token | Private and free, but slower and weaker — see running Ollama |
| Frontier models | High | Hardest reasoning; reserve for the steps that need it |
The split-model pattern
The cheapest real answer is usually two models: a low-cost one for planning, summaries and routine tool calls, and a stronger one for the hard step. Most of the cost levers apply regardless of model, but the split does the heavy lifting.
Prices change; measure instead
Provider prices move monthly, so do not take any table as permanent. Run your own three tasks on each candidate, compare the per-task token spend, and pick by cost per finished task — with budgets and alerts in place first.