The cheapest AI models are not the ones everyone talks about

Everyone compares flagship models. Your invoices don't need one.

Most AI comparisons pit the biggest models against each other. But the jobs a small business actually hands to AI are short and repetitive: tidy up a client email, pull the amount and due date out of a bill, turn rough notes into invoice line items, write a polite payment reminder. A small, cheap model does these well, and the price gap is enormous.

A token is a piece of a word; 1,000 tokens is roughly 750 English words. Providers charge separately for what you send (input) and what the model writes back (output), per million tokens. Output usually costs four to eight times more, so short answers save more money than short prompts.

The price list, checked on the providers' own pages

Models that "think" before answering bill those thinking tokens as output. On reasoning-heavy tasks your real bill can be several times the table. Turn thinking down (or off, where the model allows it) for simple jobs.

Five underrated picks, and what each is quietly good at

1. GPT-5 nano: the lowest sticker price

At $0.05 in and $0.40 out, it's the cheapest model on OpenAI's own price list, cheaper than the newer GPT-6 Luna. It's an older generation, so test it on your task, but for sorting, tagging and pulling fields out of text, older small models are often all you need.

2. Gemini 2.5 Flash-Lite: the old one is the cheap one

Google's newer Gemini 3.1 Flash-Lite costs 2.5 times as much for input and 3.75 times as much for output. The 2.5 version is still listed, has a free tier for trying things out, and its batch price halves the cost again to $0.05 in and $0.20 out for jobs that can wait.

3. DeepSeek Flash after 3:30 pm: a discount built for India's evenings

DeepSeek charges double during peak hours, which are 01:00–04:00 and 06:00–10:00 UTC on weekdays. In India that's 6:30–9:30 am and 11:30 am–3:30 pm. Everything else, including every evening and the whole weekend, is half price. Run your overnight batch jobs then. Repeated text it has already seen (a cache hit) costs just $0.003 per million tokens off-peak.

4. Claude Haiku 5.5: cheap if you keep prompts short

At $0.10 in and $0.50 out it matches GPT-6 Luna, and repeated context read from its cache costs $0.01 per million tokens. Its batch price is $0.05 in and $0.25 out. The catch is in the small print: prompts over 100,000 tokens cost five times more, so it suits short, frequent jobs rather than whole-book summaries.

5. GPT-6 Luna: newest generation, budget price

OpenAI's newest small model costs the same as Haiku. If you want a current-generation model without paying mid-range prices, start here and compare it with Haiku on the same 20 tasks.

How to choose in an afternoon

1. 1Collect 20 real examples of the job: emails, bills, notes. Remove names, phone numbers, PAN and bank details first.
2. 2Write down what a correct answer looks like for each, before you try any model.
3. 3Run the same prompt through two or three cheap models from the table.
4. 4Count how many of the 20 each got right. Keep the cheapest one that passes, and test the next size up only if none do.
5. 5Re-check prices every few months: in this market they often fall.

If you don't want to pay anything at all, open models can run on your own laptop for free. We compared them in our post on free AI models by task.