AI Cost Calculator

Estimate the monthly cost of using AI to process your documents and spreadsheets - whether on your own GPU server or through a provider's API.

Reference pricing · Sep 28, 2026
Your workload
Tell us what you're processing. Every estimate below updates instantly as you change this.

~8 pages (e.g. a standard contract) · ~5,600 tokens/document

Pull structured fields out of a document (names, dates, amounts, clauses)

1

Increase this if your workflow calls the model more than once per document (e.g. extract, then validate, then summarize).

Usage pattern
When does the model need to be working? GPUs are sized so your volume gets done inside this window (not spread evenly across every hour of the month) - by default, business hours: 10 hours a day, 5 days a week.

Documents are processed in the background - sized to finish your monthly volume within the active hours below.

10h/day
5 days/wk
~217 active hours/month
Compliance regions
Select every region you need a separate deployment in for data-residency or compliance reasons. Your document volume is split evenly across the regions you pick, and each one is costed independently - compliance deployments don't share infrastructure across borders.
  • United States: AWS us-east-1 · Azure East US · GCP us-central1. Baseline region, used as the reference price for GPU rental.

21 open-source models · select one to size your server

Meta Llama

Llama 3.1 8B

$$$$

Small and cheap to run - good for high-volume, simple extraction tasks.

Intelligence
Capable

$734/mo

on its recommended GPU, with your current hosting settings

1x L4 24GB

recommended minimum

Mistral

Mistral 7B v0.3

$$$$

Fully open, permissively licensed, and inexpensive to serve.

Intelligence
Capable

$734/mo

on its recommended GPU, with your current hosting settings

1x L4 24GB

recommended minimum

Google Gemma

Gemma 2 9B

$$$$

Google's small open model - decent quality for routine document summarization.

Intelligence
Capable

$734/mo

on its recommended GPU, with your current hosting settings

1x L4 24GB

recommended minimum

OpenAI GPT-OSS

GPT-OSS 20B

$$$$

OpenAI's smallest open-weight model - runs on a single cheap GPU.

Intelligence
Strong

$734/mo

on its recommended GPU, with your current hosting settings

1x L4 24GB (native MXFP4)

recommended minimum

Google Gemma

Gemma 4 12B

$$$$

Unified multimodal Gemma model that still fits a single cheap GPU.

Intelligence
Strong

$734/mo

on its recommended GPU, with your current hosting settings

1x L4 24GB

recommended minimum

Mistral

Ministral 3 14B

$$$$

Mistral's compact edge-oriented model - cheap and fast for simple tasks.

Intelligence
Strong

$734/mo

on its recommended GPU, with your current hosting settings

1x L4 24GB

recommended minimum

Microsoft Phi

Phi-4 14B

$$$$

Small model punching above its weight on reasoning and structured extraction.

Intelligence
Strong

$918/mo

on its recommended GPU, with your current hosting settings

1x A10G 24GB (4-bit)

recommended minimum

Mistral

Mistral Small 3.x (24B)

$$$$

Good general business fit - handles most document workflows on a single GPU.

Intelligence
Strong

$2,505/mo

on its recommended GPU, with your current hosting settings

1x A100 40GB

recommended minimum

Google Gemma

Gemma 4 26B-A4B (MoE)

$$$$

Mixture-of-experts Gemma - strong throughput for its total size.

Intelligence
Strong

$2,505/mo

on its recommended GPU, with your current hosting settings

1x A100 40GB

recommended minimum

Mistral

Magistral Small 1.2

$$$$

Reasoning-tuned Mistral model - good for step-by-step document analysis.

Intelligence
Strong

$2,505/mo

on its recommended GPU, with your current hosting settings

1x A100 40GB

recommended minimum

DeepSeek

DeepSeek-R1-Distill-Qwen 32B

$$$$

Reasoning-distilled model - good for tasks needing careful, step-by-step analysis.

Intelligence
Strong

$6,278/mo

on its recommended GPU, with your current hosting settings

1x A100 80GB

recommended minimum

Meta Llama

Llama 3.3 70B

$$$$

The most widely deployed open 70B model - strong all-around document quality.

Intelligence
Advanced

$6,278/mo

on its recommended GPU, with your current hosting settings

1x A100 80GB (4-bit)

recommended minimum

Alibaba Qwen

Qwen2.5 72B

$$$$

Excellent at structured documents and tables - a strong spreadsheet-heavy workload fit.

Intelligence
Advanced

$6,278/mo

on its recommended GPU, with your current hosting settings

1x H100 80GB (4-bit)

recommended minimum

Alibaba Qwen

Qwen3.8 27B

$$$$

Newer, high-intelligence-per-GPU option that still fits on a single card.

Intelligence
Advanced

$6,278/mo

on its recommended GPU, with your current hosting settings

1x A100 80GB

recommended minimum

Google Gemma

Gemma 4 (31B)

$$$$

Google's newest fully open model - top-tier reasoning that still fits one GPU.

Intelligence
Advanced

$6,278/mo

on its recommended GPU, with your current hosting settings

1x H100 80GB

recommended minimum

Meta Llama

Llama 4 Scout

$$$$

Frontier-class MoE model that still fits a single H100 - huge context window for bulk batches.

Intelligence
Advanced

$6,278/mo

on its recommended GPU, with your current hosting settings

1x H100 80GB (4-bit/fp8)

recommended minimum

NVIDIA Nemotron

Nemotron 3 Nano

$$$$

NVIDIA's efficient small model, tuned for cheap high-volume tasks.

Intelligence
Strong

$6,278/mo

on its recommended GPU, with your current hosting settings

1x A100 80GB (or 1x L4/A10G at 4-bit)

recommended minimum

OpenAI GPT-OSS

GPT-OSS 120B

$$$$

OpenAI's largest open-weight model - a tiny active-param MoE that's cheap to serve for its size.

Intelligence
Strong

$6,278/mo

on its recommended GPU, with your current hosting settings

1x H100 80GB (native MXFP4)

recommended minimum

NVIDIA Nemotron

Nemotron 3 Super

$$$$

NVIDIA's mid-size Nemotron - strong reasoning on a small 2-GPU node.

Intelligence
Advanced

$12,556/mo

on its recommended GPU, with your current hosting settings

2x H100 80GB (fp8)

recommended minimum

DeepSeek

DeepSeek V4 Flash

$$$$

The practical DeepSeek-class option - excellent reasoning on a small multi-GPU node.

Intelligence
Advanced

$25,112/mo

on its recommended GPU, with your current hosting settings

4x H100 80GB (int4)

recommended minimum

Mistral

Mistral Large 3 (open-weight)

$$$$

Mistral's open-weight flagship - frontier-class quality on a multi-GPU cluster.

Intelligence
Advanced

$50,224/mo

on its recommended GPU, with your current hosting settings

8x H100 80GB (cluster)

recommended minimum

Requires a multi-GPU cluster - a significant infrastructure commitment

Where will it run?
Choose a cloud GPU rental or your own server hardware. Llama 3.3 70B recommends 1x A100 80GB (4-bit) minimum.

GPUs stay on 24/7 (730 h/month), even outside your active hours.

Approx. discount for a 1-year commitment instead of on-demand rates.

25%
Smart model routing
Send simple requests to a smaller, cheaper self-hosted model and only run Llama 3.3 70B for the requests that actually need it. Routed by Laya - an open-source, self-hosted decision model with no external calls, consistent with a fully self-hosted deployment. Only open-source models cheaper than Llama 3.3 70B to self-host are offered below. Routing pays off once your volume is large enough that running Llama 3.3 70B alone would already need more than one GPU - at low volume, splitting into separate deployments can cost more, since each one needs its own minimum GPU.

Adds a Laya routing step in front of your self-hosted models. Optional - off by default.

Cost breakdown
Meta Llama
Llama 3.3 70B on AWS (H100-80GB)

$6,278

per month

Per document

$6.2780

Per year

$75,336

Server capacity used0%

1 GPU - minimum footprint for this model. Average load during your ~217 active hours/month, with 20% planning headroom kept free.

GPU rental

1 GPU x ~730 h/month

$5,022

Ops & maintenance

$1,256

Total

$6,278

Estimate based on public cloud list pricing and typical hardware/ops assumptions. Actual cost depends on region, negotiated discounts and real-world utilization.