veqen
Sign in Run a free audit

Product inference

Know what each customer costs you, and keep your keys from costing you more

Cost and margin per customer, feature and model across every provider, your own GPUs included, with fraud protection for the keys your product runs on. Spend nobody on your team made is caught, usually within the hour, the key is shut off, and what a provider overcharged you is found and claimed back.

Run a free audit Read the docs
pip install veqen npm i @veqen/sdk

Free up to $1,000 of AI spend a month with the monthly audit, and hourly protection, shut-off, claims and the month-end close from $39 a month.

Protection first

A stolen key costs you the weekend, unless something is watching it

What one compromised key cost

In March 2026, The Register reported that a three-developer company in Mexico, which normally spends about $180 a month on Google's Gemini API, was billed $82,314 in 48 hours after one of its API keys was compromised. Most of it went on Gemini 3 Pro Image and Gemini 3 Pro Text. A spending alert on the month's total can't help when the whole bill runs up in two days.

The Register, "Dev stunned by $82K Gemini bill after unknown API key thief goes to town", 3 March 2026. Read the article

The same kind of key, with Veqen

  1. The key starts running a video model it has never used.
  2. Needs action Veqen flags it: a model the key has never run, at forty times its usual hour, with no tracked call to explain it.
  3. Security taps Shut off key in the Slack alert and confirms it in Veqen.
  4. Evidence ready The hour-by-hour evidence goes to the provider with the refund request.

The timeline with Veqen is illustrative. Veqen reads per-key usage every fifteen minutes where a provider reports it, and how soon a provider reports an hour varies from minutes to about an hour.

Install

Two lines in the client you already use

Wrap your OpenAI, Anthropic or Google client, or any OpenAI-compatible one: Together, Fireworks, Groq, Azure AI Foundry, or your own vLLM, Ollama or TGI. Chat, embedding, image and audio requests are priced in your process and recorded in the background. The SDKs also work with LiteLLM, LangChain, the Vercel AI SDK and OpenTelemetry.

Prompts stay with you
Veqen stores token counts, the model, the price and your tags, never prompts or responses.
No proxy
Requests go straight to the provider, so nothing is added to their latency.
No account needed to start
Without an API key, the SDK still prices every request locally and enforces budgets.
import veqen
from openai import OpenAI

veqen.init()  # reads VEQEN_API_KEY; without one, prices locally
client = veqen.wrap(OpenAI(), feature="support-bot")

with veqen.tags(customer=account.id):
    reply = client.responses.create(model="gpt-5-mini", input=question)

Self-hosted and GPU clouds

Models on your own GPUs, priced like any other

A model on your hardware or a GPU cloud has no price list, so you give it one: what a million tokens, an image or a second of video costs you to serve. Calls through the SDK are priced with it and tagged like any other. Import the month's GPU bill, from any GPU cloud's usage export, a cloud billing export or a FOCUS file, and Veqen sets the price from what the hardware actually cost.

Self-hosted models · Your prices · Bill check

self-hostedYour priceOctober
qwen3-coder on 8×H100$0.42 per 1M output$6,120.44
flux-dev on 2×L40S$0.0006 per image$611.40
whisper-large on 1×A10G$0.0011 per minute$94.10
Matched to the GPU bill96.8%

Product

Cost per customer, and a bill nobody can run up behind your back

Exact pricing

Every rule providers bill by: cache reads and writes, batch, priority and flex tiers, long-context surcharges and regional multipliers; images by size and quality, video by the second, speech by the character, transcription by the minute and documents by the page. The price book is checked daily, versioned and open source.

ModelBilled forCost
claude-sonnet-5-52,400 input, 12,000 cached, 380 output tokens$0.0098
gpt-image-11 image, high quality, 1024×1024$0.17
veo-3.1-generate-0018 seconds of video$3.20
nova-330 minutes of audio$0.13
mistral-ocr-latestA 40-page contract$0.16
5 requests$3.67

Cost by customer, feature and project

Tag requests with a customer, feature, user, project and environment once, around the code that makes them. See cost per customer, per active user, per team and per project, the burn rate, and a forecast to month end.

CustomerTop featureThis month
acmesupport-bot$412.20
initechcode-review$268.94
globexsummaries$97.31
Free plan, 1,204 userschat$506.12

Margin, and the accounts that abuse it

With a read-only Stripe key, revenue sits next to AI cost for each customer. Accounts whose compute runs far past what they pay are flagged and sent to your fraud tool by webhook or as Stripe metadata, before a free trial turns into someone's GPU farm.

CustomerRevenueAI costMargin
hooli Flagged$49.00$1,071.80−2087%
globex$299.00$97.3167.5%
acme$1,990.00$412.2079.3%

Budgets that stop the next call

Set budgets per organization, customer, feature, user or project. Veqen alerts, and blocks new requests once recorded spend reaches the limit; in-flight requests and reporting delays can overshoot it a little. How enforcement works

veqen.budget(5.00, customer="*")  # every free user gets $5 a month

try:
    client.responses.create(...)
except veqen.BudgetExceeded as e:
    return upgrade_prompt(e.limit_usd)

Overcharges found and claimed

Wherever a provider reports both its bill and its usage per key, today OpenAI and Anthropic, its usage is priced at its rate card, cache, batch and long-context rules included, and compared with what it billed for the same day and model. Overcharges become claims with the evidence attached, and credits you win back go into Provider credits and the month-end close.

Day and modelBilledAt the rate card
Oct 03 · claude-sonnet-4-5$1,180.40$904.12
Oct 04 · claude-sonnet-4-5$1,212.88$931.40
Billed above the rate card, ready to claim$557.76

A cost review on every pull request

When a pull request changes a model, the GitHub Action comments with what the change costs at your real production volume, before it ships.

gpt-4o → gpt-5-mini
monthly impact: −$1,912.40

What each agent run costs

Tag the calls of one agent run, conversation or batch job with the same id, or let the OpenAI Agents SDK and OpenTelemetry do it, and Veqen adds them up: what a typical run costs, the 90th and 99th percentile, and the runs that ran away, step by step. How runs work

typical run    $0.042
99th pctile    $1.87
run_8f2c       $14.20  ▲ 41 calls

Free up to $1,000 a month

$39 up to $5,000

$109 up to $25,000

$349 up to $100,000

All bands

Frequently asked questions

Which models and providers does it cover?

Text, embedding, realtime voice, image, video, speech, transcription, rerank and OCR models: over 2,000 of them from model labs, open-model hosts such as Together, Fireworks and Groq, clouds such as Bedrock and Azure AI Foundry, and media providers such as fal, Runway and ElevenLabs. Models on your own GPUs or a GPU cloud take the price you give them, and the month's GPU bill checks it.

How fast do you catch a stolen key?

On paid plans Veqen reads per-key usage every fifteen minutes from OpenAI, Anthropic, Fireworks AI, fal and Baseten. A key woken after a month unused, a model it has never run, or an hour far above its usual is flagged as soon as the provider reports that hour, which is usually within the hour. Spend that only shows up against your own tracked calls waits two hours for those calls to arrive. At other providers Veqen watches the key list and its totals. Owners, admins and security members get the alert by email, Slack or webhook as soon as it's flagged.

Does Veqen see my prompts?

The SDK sends token counts, the model, the price and the tags you set, without copying prompts or responses. Use pseudonymous tags, and when forwarding third-party telemetry, configure content redaction at the sender too.

Does it slow my app down?

Pricing happens in memory and events are sent in background batches. If Veqen is unreachable, your requests carry on and the events are retried.

Do I need an account?

Not to start. The open-source SDK and CLI price requests on your machine. Add a free key when you want the dashboard, budgets shared across servers, and checks against your provider bills.

How is this different from an observability tool?

Observability tools trace requests and estimate cost from them. Veqen treats cost as accounting data: each request is priced at a versioned rate, totals are checked against the provider's invoice key by key, and the results feed fraud protection, profit per customer and your accounting.

Can we use Veqen without the dashboard?

Yes. Nearly everything the dashboard does is in the REST API with a full-access key, and Claude, ChatGPT, Cursor or any MCP client can query your spend and manage budgets, credits, bills and month-end over MCP. Roles apply the same way everywhere.

Find out what your keys did last month

The free audit reads 30 days from each provider with read-only keys, and the report is ready within the hour.

Run a free audit Start with the SDK