The lowest-cost LLM APIs, ranked by real per-token rate
Most "cheapest LLM" lists rank models by their published input price and stop there. Two problems with that: output tokens usually cost several times more than input, and the model you actually run is rarely the one with the lowest sticker. The table below ranks by blended cost at a 1/3-output workload.
Twelve lowest blended rates
| Model | Vendor | Rate in / out |
Blended at 1/3 output | Official list |
|---|---|---|---|---|
| gemma2-9b-it | $0.01 / $0.01 | $0.01 | — | |
| jev-1.14.0 | Other | $0.021 / $0 | $0.014 | — |
| gemma-3-1b-it | $0.01 / $0.03 | $0.0167 | — | |
| glm-4.5-flash | Zhipu | $0.01 / $0.04 | $0.02 | — |
| gemma-2-2b-it | $0.0125 / $0.0375 | $0.0208 | — | |
| yi-spark | Other | $0.05 / $0.05 | $0.05 | — |
| qwen-turbo | Alibaba | $0.025 / $0.1 | $0.05 | — |
| glm-4 | Zhipu | $0.0265 / $0.106 | $0.053 | — |
| doubao-seed-2-0-mini-260215 | ByteDance | $0.015 / $0.15 | $0.06 | — |
| llama-3.3-70b | Meta | $0.05 / $0.148 | $0.0827 | — |
| gpt-5-nano | OpenAI | $0.025 / $0.2 | $0.0833 | — |
| qwen-flash | Alibaba | $0.025 / $0.2 | $0.0833 | — |
USD per 1M tokens, checked 2026-11-01. Blended assumes one third of tokens are output — that is a typical chat or coding-agent mix. "Official list" is the vendor's published US rate where one exists; a dash means the vendor does not publish a comparable single number (tiered or regional pricing).
The cheapest model from each vendor
| Model | Vendor | Rate in / out |
Blended at 1/3 output | Official list |
|---|---|---|---|---|
| gemma2-9b-it | $0.01 / $0.01 | $0.01 | — | |
| jev-1.14.0 | Other | $0.021 / $0 | $0.014 | — |
| glm-4.5-flash | Zhipu | $0.01 / $0.04 | $0.02 | — |
| qwen-turbo | Alibaba | $0.025 / $0.1 | $0.05 | — |
| doubao-seed-2-0-mini-260215 | ByteDance | $0.015 / $0.15 | $0.06 | — |
| llama-3.3-70b | Meta | $0.05 / $0.148 | $0.0827 | — |
| gpt-5-nano | OpenAI | $0.025 / $0.2 | $0.0833 | — |
| ernie-3.5-8k | Baidu | $0.059 / $0.1475 | $0.0885 | — |
| deepseek-ocr | DeepSeek | $0.108 / $0.108 | $0.108 | — |
| Claude-haiku-5-5 | Anthropic | $0.05 / $0.25 | $0.1167 | — |
| MiniMax-M3.1-Flash | MiniMax | $0.075 / $0.3 | $0.15 | — |
| grok-4-fast | xAI | $0.1 / $0.25 | $0.15 | — |
| hunyuan-a13b | Tencent | $0.25 / $1 | $0.5 | — |
| kimi-k2 | Moonshot | $0.3 / $1.2 | $0.6 | — |
| Dolphin3.0-R1-Mistral-24B | Mistral | $0.7286 / $1.4572 | $0.9715 | — |
The lowest-priced model available from each vendor. Useful once you have already picked a vendor, for tooling, data-residency or existing-integration reasons, and want the floor of that vendor's range rather than the cheapest model overall. Prices checked 2026-11-01.
How to price a high-volume workload honestly
Cheap per-token rates are where budgets are won and lost, because volume multiplies every small difference. Four things matter, in this order:
- The input:output split. Output tokens are priced several times higher than input on nearly every model. A workload that looks cheap at the input rate can turn expensive the moment it generates.
- Retries. A model that fails 10% of the time and needs a re-run is priced at 1.1x its nominal rate. Cheap models are not exempt from this (test accuracy before you commit volume).
- Context resends. In an agent loop, the whole conversation is re-sent on every turn, so a long thread multiplies your input cost even if each individual call looks small.
- Escalation. If 20% of requests go to a frontier model, your blended cost sits much closer to the cheap tier than the frontier rate, which is why routing usually beats picking one model.
Run your own split through the cost calculator rather than using the table above as a budget: the ranking is real, but the totals depend entirely on your mix.
Cheap models compared with the mainstream
For context on what "cheap" means relative to what most teams run: Claude vs Gemini · GPT vs Gemini · DeepSeek vs Claude · all 852 models
Related guides
- How to reduce API cost — the three levers that move the number most
- How to monitor API spend — know your real mix before you optimise it
- OpenAI-compatible endpoints — what breaks when you switch
FAQ
What is the cheapest LLM API available?
Of the models with usable per-token pricing in our index, gemma2-9b-it is the cheapest at $0.01 per 1M blended tokens, with jev-1.14.0 and gemma-3-1b-it close behind. The table lists the twelve lowest; the full catalogue of 852 model IDs is on the pricing page.
Is the cheapest model the cheapest option overall?
Not necessarily. A cheaper per-token rate can cost more in total if it needs retries, a verification pass, or a second attempt on harder inputs. Price the whole task (including retries) rather than the per-token rate alone.
Why do some models not appear in this list?
We exclude models without usable per-token pricing, plus embeddings, rerankers, image, audio and speech models, because they are not billed on the same basis as chat models. Preview and test identifiers are also excluded so the list reflects models you can depend on.
Do these prices include any discount?
These are the current gateway rates, read from the provider's live rate configuration on 2026-11-01. They are not a permanent discount and they move with upstream promotions, which is why we publish dated rates per model rather than a headline percentage.
How do I start using the cheapest model?
Create an account, generate a key, and point your SDK at the endpoint (the model ID is the only thing that changes). The SDK guide has the one-line version.
How much does gemma2-9b-it cost per month?
At $0.01 per 1M blended tokens: about $0.1 for 10M tokens a month, $1 for 100M, and $10 for 1B. Those are token costs only, there is no subscription. The caveat that matters more than the number: this is the cheapest per token, not necessarily per task. A model that needs fewer tokens to reach the same answer can cost less overall.
Are open-weight models cheaper than closed models?
Per token, usually by a wide margin — the gap between the lowest-priced model here and a frontier model is roughly 583x on a blended rate. Two things complicate it: the same open-weight model is hosted by many providers at different margins, so identical weights can cost several times more on one host than another; and self-hosting replaces a per-token bill with GPU capacity you pay for whether or not it is busy. Cheap per token is not the same as cheap to run.