LLM spend, under control.
Explore the productYour AI can move fast.
Your spend should have limits.
Track LLM API costs by customer, feature, and model. Set spend budgets and stop over-budget requests before they reach OpenAI, Anthropic, or Gemini.
Demo workspace
Example request data
MTDMonth-to-date
$7,241BlockedBlocked spend
$844CacheCache hit
42%Example p95Example p95
4.1 msRequest path
before provider spend
customer_982
gpt-4o
support_copilot
Blocked
team_enterprise
claude-3-5
agent_review
Routed
customer_123
gpt-4o-mini
ai_writer
Allowed
customer_771
gpt-4o-mini
research_agent
Routed
metadata.customerId=customer_123 / metadata.feature=ai_writer / policy=project_budget
Integration
One base URL
Attribution
Customer + feature
Enforcement
Spend budgets
Alerting
Webhook first
02 Control before the invoice
A retry loop. A busy agent. One customer using far more than expected. Put a limit in the request path, not just a chart on a dashboard.
Follow spend by customer, feature, and model. Get a warning as a budget fills up. When it runs out, Halvr rejects the request before it reaches your provider.
Budget envelope / Example data
$7,244
of $10,000
The chart shows request load rising through the month while an eighty percent budget threshold remains visible as a dashed warning line.
Peak request load
18.2k/min
Budget window
rolling 30d
Warn threshold
80%
Block mode
pre-provider
Works with
03 How it works
Halvr is not a reporting SDK after the fact. It is an instrument in the request path, so it can decide, route, meter, and reject before spend happens.
Keep your SDK, payload shape, and provider mental model. The only required change is routing through Halvr.
Send stable customer and feature identifiers so spend maps to product ownership instead of raw provider accounts.
Project, key, customer, and feature budgets are evaluated before the upstream provider call can spend tokens.
Usage events, alerts, dashboards, and audit logs make the next cost decision more precise than the last one.
OpenAI-compatible base URL
No SDK rewrite.
const openai = new OpenAI({
apiKey: process.env.HALVR_API_KEY,
baseURL: "https://api.halvr.io/v1"
});
await openai.chat.completions.create({
model: "gpt-4o-mini",
messages,
metadata: {
customerId: "customer_123",
feature: "ai_writer"
}
});
// Attribution now follows the request.04 Savings calculator
An illustrative provider-cost scenario, not a benchmark or promise. Subscription fees are separate and are not calculated from estimated savings.
Traffic assumptions
Adjust these example traffic assumptions to explore the model. They are not observed customer results.
5,400,000
monthly calls
Illustrative provider savings
$43,546Before Halvr subscription fees. Actual results depend on workload, cache eligibility, model quality, and provider pricing.
Cache
$18,662Routing
$13,997Model selection
$10,88605 Pricing
Evaluate Halvr for 14 days or 10,000 requests, whichever comes first. No card required. Subscribe explicitly to continue; provider inference is billed separately.
For a developer bringing an AI feature into production.
50,000 gateway requests per billing period. No automatic overages.
For a team managing customer and feature costs across applications.
250,000 gateway requests per billing period. No automatic overages.
For teams operating higher-volume AI workloads.
1,000,000 gateway requests per billing period. No automatic overages.
Request allowances reset with the paid billing period. Admitted gateway requests, including cache hits and failed calls, count toward the allowance. At the limit, new requests stop; there are no automatic overage charges. Need higher limits? Contact us.