What one run costs
A typical run of this task sends about 25K tokens and gets back 2.5K, over 8 calls — 220K 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 | $0.0180 | not stated | Vendor page |
| Gemini 2.5 Flash Lite | $0.0280 | not stated | Vendor page | |
| Claude Haiku 5.5 | Anthropic | $0.0300 | confirmed | Vendor page |
| GPT-6 Luna | OpenAI | $0.0300 | not stated | Vendor page |
| GPT-5.6 Luna | OpenAI | $0.0640 | not stated | Vendor page |
| GPT-5.4 Nano | OpenAI | $0.0650 | not stated | Vendor page |
| Gemini 3.1 Flash Lite | $0.0800 | not stated | Vendor page | |
| GPT-5 Mini | OpenAI | $0.0900 | not stated | Vendor page |
Do not ask a model what it knows. Ask it to find out, and make every claim carry a link you can click.
Invented citations are the characteristic failure of this task. A fabricated reference reads exactly like a real one — correct-looking authors, plausible title, a year that makes sense — and the only defence is opening it.
The knowledge cutoff (The date after which a model saw no training data. It knows nothing later unless you tell it or it can search.) bites here too. Anything after the model’s training data is outside what it knows, and it will not tell you that unaided.
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.