Our cloud
One line in your client. Nothing to deploy.
model call orchestration gateway · cost governance
AWS itself tells you to build an LLM gateway.
Veltrix already is that gateway.
22 models, 9 providers. Veltrix decides who answers each call, measures the cost of the request, and refuses it when the budget runs out.
You turn it on by swapping one line. And you start by measuring: a 4-week audit, no cost and no card.
base_url = "api.veltrixplatform.com/v1" ← the only change
the bill that dissolves · 30 days measured
01 · the product
Four decisions per call. Watch each one happen.
rerouted to a cheaper model
64.0%
OpenAI
gpt-4o
US$ 2,50 / 1M
default
Anthropic
sonnet 4.5
US$ 3,00 / 1M
Google
2.5 flash
US$ 0,30 / 1M
DeepSeek
chat
US$ 0,27 / 1M
vLLM
your GPU
cost/hour
chosen
capability
context
quality
latency
cost
criticality
2,246 of 3,512 calls outside the cache, with nothing declared by you. Deterministic, and every choice stays in the log.
02 · the cut
Your number depends on your traffic. Put yours in.
R$ 40,000
1k500k
order of magnitude: ~222,222 calls / month
The order of magnitude assumes an average of R$ 0.18 per call — our own assumption to give scale, not a measurement of your traffic.
floor across the whole bill (14.0%)
R$ 5,600 / mo
includes everything the router left alone
where the router acted (50.2%)
R$ 20,080 / mo
only on calls where it swapped the model
A projection from our measured percentages, not a promise about your traffic. If you already run everything on cheap models, there is little to cut — and we say so before you sign.
03 · the catalog
22 models, a 100× price gap — and your own vLLM in the same queue.
That gap is what makes orchestration worth it. Each dot is an active model, placed by its real price. Your own endpoint joins as one more provider.
qwen-turbo · US$ 0.00005/1kclaude-opus-4-8 · US$ 0.005/1k
100×
the price gap
input price per 1k tokens · logarithmic scale · only models active in the router, catalog as of 2026-08-20
- OpenAI×2
- Anthropic×4
- Google×2
- Groq×2
- DeepSeek×2
- Qwen×4
- Zhipu×3
- Moonshot×2
- MiniMax×1
- vLLMyour GPU
- Bedrockyour region
- Azure OpenAIyour deployment
Pricing taken from each provider's official documentation. 7 models ship disabled and outside the router — four for unconfirmed or conflicting pricing, one for a context window two of the provider's own pages contradict, two for being dominated by a cheaper model from the same provider. The reason for each is written down in the catalog. vLLM, TGI, Ollama, Bedrock and Azure OpenAI join through your own configuration, at the cost you declare.
04 · where it runs
Our cloud, yours, or pointing at your own GPU.
Your model, Veltrix in the cloud
Your GPU answers; the cap and the attribution still come from us.
Veltrix inside your infra
The embedding comes from a local model, not from OpenAI. Never run end to end on a clean machine — and we say so.
vLLM · TGI · Ollama — the case nobody solves
The fallback chain stops being a try/catch and becomes policy.
In the order you wrote it, never reordered by price — and every trigger is counted.
cost/hour
an own model has no per-token price: you declare the GPU cost and we attribute it per second occupied
counterfactual
the machine would cost the same hour if nobody called — idle GPU is shown separately
Where Veltrix has evidence — and where it cannot make a claim yet.
Coverage is not a generic savings promise. Every surface states its available signal and what that signal supports.
SurfaceReadingEvidence
Browser
Estimated usage in compatible integrations
EstimatedBrowser events; varies by provider
Desktop
Compatible tool in the foreground
PresenceOptional Local Collector on macOS
Terminal
Recognized CLI running
PresenceLocal allowlist; no arguments or paths
API through Gateway
Tokens, cost, attribution, and per-call decision
MeasuredTraffic through the Veltrix endpoint
Local presence never becomes tokens, cost, or quota. Complete financial measurement only exists in traffic routed through Gateway.
Start where you are. Move up when you need control.
Conversion does not depend on promising savings before there is measurable traffic.
Lens · discover
Map browser, desktop, and terminal with explicit sources.
Install LensGateway · measure and govern
Route API calls to measure real cost and apply rules.
See Gateway pricingTeams · respond together
Give spending, policies, and team coverage an owner.
See Teams05 · against the hyperscaler
They optimize inside their own cloud. Nobody shows the spend between clouds.
Fighting Bedrock or Foundry on routing would be losing on purpose. Click each row — the source is their own documentation.
AWS answers "No" to per-prompt cost in Cost Explorer. Minimum granularity is usage type, per day. With Veltrix it is a database row — auditable.
AWS · Bedrock cost mgmt FAQ ↗Microsoft writes that Azure OpenAI has no hard cap. AWS Budgets blocks via IAM/SCP, hours late, and denies everything — not per tenant.
Microsoft Learn · Azure OpenAI quotas ↗AgentCore does call OpenAI and Gemini — but the docs say the provider bills you directly. That spend never reaches Cost Explorer.
AWS · Bedrock AgentCore docs ↗AWS's own FAQ answers: "Not from the Amazon Bedrock side. […] Enforce tagging in a shared client or LLM gateway." In other words: they tell you to build exactly what Veltrix already is.
AWS · Bedrock cost mgmt FAQ ↗06 · the audit
Every number has a query behind it. Including when the answer is "we don't measure that".
request_logs · 30d
SELECT count(*) AS chamadas,
round(sum(estimated_cost::numeric), 4) AS pago_usd,
round(sum(cost_saved::numeric), 4) AS economizado_usd
FROM request_logs
WHERE created_at > now() - interval '30 days'
AND cache_hit = false
AND status_code < 400;the four states of a number
- measurednon-null column over a declared base
- counterfactualcomparison against a scenario that did not happen
- not measuredthe sensor did not write — absence, never a zero-height bar
- not eligiblewe measured, and the lever correctly did not act
"There was nothing to compress" is success; "the sensor went down" is an incident. In most tools both become the same green 0%.
07 · pricing
Flat subscription. No percentage of what you spend.
Your budget line doesn't move when traffic grows, and all the savings stay with you.
Lens
R$ 0
chrome extension
A capacity wallet per provider, with source and confidence per reading.
Gateway
mainR$ 179
/ month
Cache, routing across 9 providers, a cap that blocks, cost per request and per feature, alerts on Slack, Discord and webhook.
Teams
R$ 219
/ seat · min. 3
Everything in Gateway, per seat, with roles (admin, dev, finance), budgets and quota approval.
next step
Start with the audit. It's yours.
Four weeks measuring your real traffic, without changing a line of your application. At the end, an auditable report for whoever decides — including if the conclusion is that you don't need us.