Best AI model for answering support questions.
Prices checked
Llama 3.1 8B Instruct at Helicone costs you 1.3¢ per 1,000 replies, the cheapest of 325 models that can do this job, checked .
Cheapest 10
all 392 models- 1
Llama 3.1 8B Instruct
Runs it themselves
Helicone · $0.02 in · $0.03 out1.3¢per 1,000 replies - 2
Nemotron 3.5 Lightning
Resells
NanoGPT · $0.05 in · $0.01 out1.3¢per 1,000 replies - 3
Mistral Nemo
Lower precision (fp8)
Eden AI · $0.02 in · $0.03 out1.4¢per 1,000 replies - 4
gpt-oss-20b
Resells
TrustedRouter · $0.01 in · $0.05 out1.6¢per 1,000 replies - 5
LFM2.5-24B-A2B
Same as makerNanoGPT · $0.05 in · $0.03 out1.7¢per 1,000 replies - 6
Ministral 3 3B
Runs it themselves
Azure Cognitive Services · $0.04 in · $0.04 out2¢per 1,000 replies - 7
Muse Spark 1.3 Contributor
Same as makerNanoGPT · $0.10 in · 0.2¢ out2.1¢per 1,000 replies - 8
Muse Spark 1.2 Contributor
Same as makerNanoGPT · $0.10 in · 0.2¢ out2.1¢per 1,000 replies - 9
GPT-6 Luna
Same as makerNanoGPT · $0.10 in · $0.01 out2.3¢per 1,000 replies - 10
Mercury 3
Same as makerNanoGPT · $0.10 in · $0.01 out2.3¢per 1,000 replies
Other ways to pick
What matters
- Tool calling required
- The answer lives in your help centre, not in the model. It has to be able to call a search function and write from what comes back.
- Output price decides it
- A reply is longer than the question, so this is one of the few jobs where the output price is the one to watch.
- 32k context is enough
- A question, a few retrieved articles and the reply fit comfortably. Paying for a million tokens of context you never fill is waste.
How we count
A customer asks something, the model looks it up in your help centre and writes the reply. Short in, longer out, and it has to be able to call your search. We count only models with 32k of context or more and tool calling, and 325 of the 392 we track qualify. The cheapest of them costs 1.3¢ per 1,000 replies at Helicone. They are ranked by what the work costs, not by the price per million tokens, because a model that is cheap to prompt and dear to answer looks different once you know the shape of the job.
A support reply is about 200 tokens in and 300 out. A cost per 1,000 replies is that multiplied out at the row's own input and output rate, so it says what the work costs; the price per million tokens under each name says what the model costs.
A price below the maker's is listed only when the provider says why.
- Runs it themselves
- The provider runs the open weights on hardware of its own, so the price is its own rather than a resale.
- Resells
- The provider sells another company's endpoint, at a margin of its own or at the maker's price.
- Lower precision
- The open weights, served at a lower precision than the maker's, so not quite the same product.
- Same as maker
- Sold by someone else at the price the maker charges.
Frequently asked
- What is the cheapest model for answering support questions?
- Llama 3.1 8B Instruct at Helicone, 1.3¢ per 1,000 replies at $0.02 in and $0.03 out per million tokens, read on 31 Oct 2026.
- How is the cost per unit worked out?
- A support reply is about 200 tokens in and 300 out. Multiply that by 1,000 and price it at each row's own input and output rate. Nothing else is counted: no cache discount, no batch rate, no free tier.
- Why do only 325 models qualify?
- The job needs 32k of context or more, tool calling. Models that fall short are still in the index, they just cannot do this job as described.
- How current are these prices?
- Prices are read every 3 days from 237 providers' public catalogs and this ranking recomputes with them. Last refresh: 31 Oct 2026.