What is the cheapest open-weight LLM API?

Among the models we track, gpt-oss-20b is cheapest to run: Darkbloom charges $0.018 per 1M input tokens and $0.09 per 1M output tokens.

Why do providers charge different prices for the same model?

Open-weight models can be hosted by anyone. Providers run them on different hardware, at different precision (quantization), with different speed and reliability guarantees, and set their own prices.

Are quantized models worse?

Serving weights at lower precision, such as FP8 or FP4, cuts hosting costs and can reduce output quality, especially on harder tasks. The cheapest endpoints are often quantized, so test quality on your own prompts before switching.

Open-weight model API prices by provider

The same open-weight model can cost very different amounts depending on who hosts it. gpt-oss-120b runs on 22 providers, from $0.03 to $0.35 per 1M input tokens. The widest gap is on GLM 5.3 Flash, where the priciest host charges 56.9x the cheapest.

Provider prices from the OpenRouter API, retrieved October 26, 2026. Only endpoints reporting healthy status are included.

Model	Providers	Cheapest host	Cheapest (in / out)	Creator’s own API	Price spread	Context
gpt-oss-20bOpenAI	8	Darkbloomfp8	$0.018 / $0.09	—	3.6x	131K
GLM 5.3 FlashZ.ai	33	Relace	$0.0413 / $0.0758	$0.15 / $0.50	56.9x	1.05M
gpt-oss-120bOpenAI	22	Near AIfp4	$0.03 / $0.15	—	7.5x	131K
DeepSeek V4.1 FlashDeepSeek	32	Sail Researchfp4	$0.09 / $0.25	—	6.1x	1.05M
Qwen3.8 27BAlibaba (Qwen)	20	Tenstorrentfp4	$0.02 / $0.50	—	8.1x	1M
Llama 3.3 70B InstructMeta	9	DeepInfrafp8	$0.10 / $0.32	—	5.1x	131K
Gemma 4 31BGoogle	13	CoreWeavefp4	$0.10 / $0.34	—	5.1x	262K
Nemotron 3 SuperNVIDIA	3	DeepInfrabf16	$0.085 / $0.40	—	5.0x	262K
DeepSeek V4 Pro 0813DeepSeek	22	Baidufp8	$0.132 / $0.396	$0.66 / $1.98	12.5x	1.05M
Mistral Small 4Mistral AI	1	Mistral	$0.15 / $0.60	$0.15 / $0.60	1.1x	262K
Llama 4 MaverickMeta	3	DigitalOcean	$0.1875 / $0.6525	—	1.8x	524K
MiniMax M3MiniMax	11	DeepInfrafp8	$0.224 / $0.88	$0.30 / $1.20	2.7x	1.05M
GLM 5.3Z.ai	36	Waferfp4	$0.0699 / $1.40	$1.40 / $4.40	10.7x	1.05M
Kimi K3Moonshot AI	21	Relacefp4	$0.17 / $7.697	$3.00 / $15.00	4.4x	1.05M

USD per 1M tokens. Sorted by the cheapest host’s blended price (3 input : 1 output). Price spread compares the priciest and cheapest healthy host.

How to pick a provider

Price is only one axis. Hosts differ in speed, uptime, rate limits and supported features such as tool calling. Our tables show list prices only.
Check the precision. Many of the cheapest endpoints serve FP8 or FP4 weights. That is usually fine for chat and extraction but can matter for math, coding and long reasoning.
Compare with closed models. Some open-weight models now rival closed ones on benchmarks. See how they stack up on the benchmarks page, and compare first-party APIs on our API comparison pages.

Frequently asked questions

What is the cheapest open-weight LLM API?: Among the models we track, gpt-oss-20b is cheapest to run: Darkbloom charges $0.018 per 1M input tokens and $0.09 per 1M output tokens.
Why do providers charge different prices for the same model?: Open-weight models can be hosted by anyone. Providers run them on different hardware, at different precision (quantization), with different speed and reliability guarantees, and set their own prices.
Are quantized models worse?: Serving weights at lower precision, such as FP8 or FP4, cuts hosting costs and can reduce output quality, especially on harder tasks. The cheapest endpoints are often quantized, so test quality on your own prompts before switching.