Report
The cheapest API call, model by model.
Every model we hold a verified price for, ranked by what a million input tokens costs at its cheapest serving provider — and how much more the dearest one charges for the exact same weights.
Cheapest verified price per model
Prices in USD per million tokens, cheapest first. Each row shows the cheapest verified host for input price and what the same host charges for output tokens.
DeepSeek V4 Flash 0731
deepseek
OpenInference
$0.01
$0.58
−97% cheaper on input price
DeepSeek V4.1 Flash
deepseek
OpenInference
$0.02
$0.89
−95% cheaper on input price
Mistral Nemo
mistralai
Io Net
$0.02
$0.02
−89% cheaper on input price
Granite 4.0 Micro
ibm-granite
Cloudflare
$0.02
$0.11
—
gpt-oss-20b
openai
Darkbloom
$0.02
$0.09
−76% cheaper on input price
Llama 3.1 8B Instruct
meta-llama
DeepInfra
$0.02
$0.04
−91% cheaper on input price
Qwen3.8 27B
qwen
Tenstorrent
$0.02
$0.50
−98% cheaper on input price
Ling 3.0 Flash
inclusionai
Novita
$0.02
$0.06
−65% cheaper on input price
gpt-oss-20b (batch)
openai
DeepInfra
$0.02
$0.11
—
GPT-5 Nano
openai
OpenAI
$0.03
$0.20
−50% cheaper on input price
GPT-5 Nano (batch)
openai
OpenAI
$0.03
$0.20
—
Nemotron 3.5 Lightning
nvidia
Darkbloom
$0.03
$0.12
−64% cheaper on input price
Nex-N2.5-Mini
nex-agi
Nex AGI
$0.03
$0.10
—
Llama 3.2 1B Instruct
meta-llama
Cloudflare
$0.03
$0.20
—
gpt-oss-120b (batch)
openai
DeepInfra
$0.03
$0.14
—
gpt-oss-120b
openai
AkashML
$0.03
$0.19
−91% cheaper on input price
Qwen3.7 Flash
qwen
Alibaba
$0.03
$0.13
—
Schematron V2 Turbo
inference-net
InferenceNet
$0.03
$0.15
—
Nova Micro 1.0
amazon
Amazon Bedrock
$0.04
$0.14
—
GLM 5.3 Flash
z-ai
OpenInference
$0.04
$0.90
−76% cheaper on input price
Command R7B (12-2024)
cohere
Cohere
$0.04
$0.15
—
GLM 5.2
z-ai
Relace
$0.04
$4.00
−98% cheaper on input price
Extract
scaledown
ScaleDown
$0.04
$0.00
—
Gemma 3 4B
Parasail
$0.04
$0.05
−20% cheaper on input price
Llama 3 8B Lunaris
sao10k
DeepInfra
$0.04
$0.05
—
Mercury 2.5
inception
Inception
$0.04
$0.15
—
Mercury 2.5 Preview
inception
Inception
$0.04
$0.15
—
Summarize
scaledown
ScaleDown
$0.04
$0.00
—
Gemma 4 26B A4B
Darkbloom
$0.04
$0.22
−72% cheaper on input price
Hy-MT2-1.8B
tencent
Tencent
$0.04
$0.18
—
DeepSeek V4 Flash 0423
deepseek
OpenInference
$0.04
$1.15
−90% cheaper on input price
Claude Haiku 5.5 (batch)
anthropic
Anthropic
$0.05
$0.25
—
Gemini 2.5 Flash Lite
Google AI Studio
$0.05
$0.20
−50% cheaper on input price
Gemini 2.5 Flash Lite (batch)
$0.05
$0.20
—
Gemma 3 12B
DeepInfra
$0.05
$0.15
—
GPT-4.1 Nano (batch)
openai
OpenAI
$0.05
$0.20
—
GPT-6 Luna
openai
OpenAI
$0.05
$0.25
−55% cheaper on input price
GPT-6 Luna (batch)
openai
OpenAI
$0.05
$0.25
—
GPT-6 Luna Pro
openai
OpenAI
$0.05
$0.25
−50% cheaper on input price
GPT-6 Luna Pro (batch)
openai
OpenAI
$0.05
$0.25
—
Showing the 40 cheapest of 491 tracked models. Browse the full catalog to filter by vendor, quality score or host spread.
Which provider is cheapest, host by host
For every model sold by more than one provider, exactly one host has the lowest input rate. This counts those wins. A provider that carries many models but wins few is convenient rather than cheap.
OpenAI
88
2326%
$1.00
DeepInfra
75
2229%
$0.15
Amazon Bedrock
38
1437%
$2.20
54
1120%
$0.56
Azure
56
1120%
$2.00
Darkbloom
9
778%
$0.05
Novita
52
713%
$0.41
GMICloud
26
623%
$0.30
Google AI Studio
17
635%
$0.38
OpenInference
4
4100%
$0.04
Parasail
39
410%
$0.15
Alibaba
33
412%
$0.35
Relace
10
330%
$0.06
StreamLake
19
316%
$0.52
DekaLLM
8
225%
$0.06
SiliconFlow
40
25%
$0.27
Minimax
8
225%
$0.30
Venice
38
25%
$0.33
Inceptron
6
233%
$0.55
Tencent
5
120%
$0.07
Want this priced against your own volumes? Use the LLM pricing comparison calculator.
How to read a cheap API call
A price is per token, not per call. An API call costs input tokens plus output tokens. Two providers can advertise the same headline rate and still bill differently once the output side is counted, which is why the blended column exists above.
The same model has more than one price. Open weight models are sold by several serving providers at once. Where that gap is wide, the cheapest call is the same model on a different host, not a weaker model.
Cheapest is not automatically correct. A model that is cheap per token but needs two attempts, or emits three times the output, is not cheap per finished task. We publish the benchmark scores next to every price on the model catalog so a quality claim has something to clear.
Prices come from the same catalog our engine prices against, and aggregator listings are excluded from provider-to-provider gaps. The full sourcing rules are in the methodology.
Price it against your own traffic.
Compare reads this same catalog and shows what your current models would cost on the cheapest verified host. Free, forever.