"What's the cheapest LLM right now?" is the wrong question. There isn't one cheapest model — there's a cheapest model for your workload, and it changes depending on what you're actually doing.
The reason is simple once you see it: your bill is input_tokens × input_price + output_tokens × output_price. A chatbot, a RAG pipeline, a coding agent and a summarizer have wildly different input-to-output ratios, and some jobs won't even run on the cheapest model because it can't fit the context. So the model that wins on one workload loses on the next.
Here's what that looks like across five common jobs, priced against each model's current, official API rates. Every figure is computed from AI Model Watch, which tracks these prices daily from provider pricing pages — nothing here is invented.
The cheapest pick flips with the job
| Use case | Monthly workload (in / out) | Hard requirement | Cheapest GA pick | Est. cost |
|---|---|---|---|---|
| Chatbot | 10M / 3M | — | Llama 3.1 8B (Meta) | ~$0.29/mo |
| RAG | 50M / 5M | ≥128K context | Llama 3.1 8B (Meta) | ~$1.15/mo |
| Summarization | 80M / 4M | ≥200K context | Qwen-Flash (Alibaba) | ~$5.60/mo |
| Coding agent | 90M / 25M | ≥200K context | Amazon Nova Lite | ~$11.40/mo |
| Vision | 20M / 5M | image input | Ministral 3 3B (Mistral) | ~$2.50/mo |
Five jobs, four different models from four different providers. No single model is the answer.
Why the winner keeps changing
1. The context-window floor eliminates the cheapest option. Llama 3.1 8B is the cheapest text model in the catalog on raw price ($0.02 in / $0.03 out per million), and it wins the chatbot and RAG rows outright. But it caps out at 128K tokens — so the moment your job needs ≥200K of context (coding, long-document summarization), it's disqualified before price even matters. The cheapest model you can't use isn't cheap.
2. Once it's a fair fight, the input:output ratio decides. Both summarization and coding require ≥200K context, so they draw from the same eligible pool — yet they pick different winners:
- Summarization is lopsided input (80M in / 4M out, ~20:1). That rewards the lowest input price, and Qwen-Flash ($0.05 in) takes it at ~$5.60/mo.
- Coding writes far more back (90M in / 25M out, ~3.6:1). Now output price carries real weight, and Amazon Nova Lite ($0.06 in / $0.24 out) beats Qwen-Flash ($0.40 out) — ~$11.40 vs ~$14.50 — purely because it's cheaper on those 25M output tokens.
Same context requirement, same candidate list, opposite winner — decided entirely by the shape of the workload.
3. A capability requirement changes the eligible pool entirely. Vision needs image input, which knocks out every text-only winner above and surfaces a different provider (Mistral) at the top. Eligibility filters run before price ranking, so a hard requirement can matter more than the price table.
The honest caveats
- These are token-price estimates, not your invoice. Real cost depends on how many tokens your app actually burns, retry rate, and whether you use prompt caching (reused input bills at roughly 10× less — worth enabling for a static system prompt).
- Cheapest-per-token ≠ best-per-task. A weaker model that loops more, or answers worse, can cost more end to end and hurt your product. Price is the floor of the decision, not the whole of it — benchmark the shortlist on your data.
- Preview tiers and vision tags carry uncertainty. Rankings here cover generally-available, priced models; "vision" is filtered by the catalog's image-input tag, so treat it as a starting shortlist, not a quality verdict.
Compute it for your own numbers
The picks above assume specific monthly token volumes. Your ratio is probably different — so the useful move isn't memorizing a winner, it's re-running the math on your input:output mix and your context floor.
Two ways to do that:
- The ranked tables, one per use case (with the full methodology and the next ~10 models, not just the winner): aimodelwatch.dev/guides
- The raw data, if you'd rather compute it yourself — a free, no-key, CORS-open JSON feed of every model's price, context window and lifecycle status: aimodelwatch.dev/api/models.json
Wire the feed into your own cost script and you never have to trust someone else's ranking — including this one.
Full per-use-case cost tables and methodology: aimodelwatch.dev/guides
Top comments (0)