What one run costs
A typical run of this task sends about 30K tokens and gets back 2K, over 15 calls — 480K tokens in total. Those are the numbers below, against the published prices.
| Model | Vendor | Per run | Can do what this needs? | Price evidence |
|---|---|---|---|---|
| GPT-5 Nano | OpenAI | $0.0345 | not stated | Vendor page |
| Gemini 2.5 Flash Lite | $0.0570 | not stated | Vendor page | |
| GPT-4.1 nano | OpenAI | $0.0570 | not stated | Corroborated |
| Claude Haiku 5.5 | Anthropic | $0.0600 | not stated | Vendor page |
| GPT-6 Luna | OpenAI | $0.0600 | not stated | Vendor page |
| GPT-4o mini | OpenAI | $0.0855 | not stated | Corroborated |
| Jamba Mini | AI21 | $0.1020 | not stated | Vendor page |
| GPT-5.6 Luna | OpenAI | $0.1260 | not stated | Vendor page |
This task needs tool use, computer use.Most vendors do not state their models’ capabilities on the page we read, and a blank is not a “no”. So a model is only removed from this list when a vendor explicitly says it cannot do something — everything else stays, marked “not stated”. Check the vendor’s own documentation before committing.
An agent loop (Plan, call a tool, read the result, decide again. The cycle an agent repeats until it finishes or is stopped.), which means the context grows on every turn and the cache is rewritten as it does. Price fifteen turns, not one: the last turn can cost several times the first.
Start with a dry run that only reports what it would do. An agent that misunderstands the goal pursues the wrong thing energetically, and you want to find that out before it has authority over anything.
Where there is an API, use tools (Letting a model call functions you define — search, read a file, send a request — and read back what they return.) rather than computer use (A model driving a screen directly — looking at pixels, moving a cursor, typing — instead of calling an API.). Driving a screen is slower, dearer in image tokens, and considerably less reliable than calling the thing directly.
What we would pick
This section is our judgement, not a figure read off a page. Everything above is arithmetic on published prices; this is an opinion, and it is labelled as one.