The LLM cost layer anyone can install — licensed to run

Prove the savings on your own data.

BV‑SALA sits in front of any LLM provider and, for every request, picks the cheapest execution that still meets your quality bar. Don't take our word for it — paste your data below and watch the meter move.

Run the verifier See pricing

Runs entirely in your browser. No signup, nothing leaves this page.

Cheapest execution winsper request

01Avoid the call entirelyresolvers · exact & semantic cache0 calls

02Avoid retrieval & toolsroute: none / single / multi / toolfewer calls

03Eliminate needless contextdedupe · rerank · adaptive‑k−tokens

04Compress what remainsleaner encoding · net‑positive only−tokens

05Constrain the outputschema · minimal contract−output

Step 1 · Preliminary tests — in your browser

Try the optimizer yourself. Nothing leaves this page.

These two instruments run the real optimizer locally so you can sanity-check the mechanism before touching a provider. The left panel runs the real BV‑SALA optimizer on data you paste — real token accounting, the same code the SDK ships. The right panel estimates a monthly bill by running the optimizer over a sample workload against a mock model. For provably-measured dollars, see Step 2 below.

Savings verifierESTIMATE — heuristic, not provider-metered

{
  "users": [
    {
      "id": 1000,
      "name": "user_0",
      "role": "admin",
      "active": false
    },
    {
      "id": 1001,
      "name": "user_1",
      "role": "member",
      "active": true
    },
    {
      "id": 1002,
      "name": "user_2",
      "role": "member",
      "active": true
    },
    {
      "id": 1003,
      "name": "user_3",
      "role": "admin",
      "active": true
    },
    {
      "id": 1004,
      "name": "user_4",
      "role": "member",
      "active": false
    },
    {
      "id": 1005,
      "name": "user_5",
      "role": "member",
      "active": true
    },
    {
      "id": 1006,
      "name": "user_6",
      "role": "admin",
      "active": true
    },
    {
      "id": 1007,
      "name": "user_7",
      "role": "member",
      "active": true
    },
    {
      "id": 1008,
      "name": "user_8",
      "role": "member",
      "active": false
    },
    {
      "id": 1009,
      "name": "user_9",
      "role": "admin",
      "active": true
    },
    {
      "id": 1010,
      "name": "user_10",
      "role": "member",
      "active": true
    },
    {
      "id": 1011,
      "name": "user_11",
      "role": "member",
      "active": true
    },
    {
      "id": 1012,
      "name": "user_12",
      "role": "admin",
      "active": false
    },
    {
      "id": 1013,
      "name": "user_13",
      "role": "member",
      "active": true
    },
    {
      "id": 1014,
      "name": "user_14",
      "role": "member",
      "active": true
    },
    {
      "id": 1015,
      "name": "user_15",
      "role": "admin",
      "active": true
    },
    {
      "id": 1016,
      "name": "user_16",
      "role": "member",
      "active": false
    },
    {
      "id": 1017,
      "name": "user_17",
      "role": "member",
      "active": true
    },
    {
      "id": 1018,
      "name": "user_18",
      "role": "admin",
      "active": true
    },
    {
      "id": 1019,
      "name": "user_19",
      "role": "member",
      "active": true
    },
    {
      "id": 1020,
      "name": "user_20",
      "role": "member",
      "active": false
    },
    {
      "id": 1021,
      "name": "user_21",
      "role": "admin",
      "active": true
    },
    {
      "id": 1022,
      "name": "user_22",
      "role": "member",
      "active": true
    },
    {
      "id": 1023,
      "name": "user_23",
      "role": "member",
      "active": true
    }
  ]
}

Your data0

BV-SALA0

Already optimal here — BV‑SALA never makes it worse.

Bill projectorESTIMATE

Requests / month

Avg input tokens

Avg output tokens

Calls avoided: 25%

Input cut: 40%

Output cut: 30%

projected saving / month · at gpt-4o-mini prices

$170

$330 → $160 · 52% lower

Step 2 · How savings are measured — reproducible

Both arms metered by the provider's own tokenizer — no modeled baseline.

This is the output of the A/B benchmark (pnpm ab): for every scenario it sends a naive request and the BV‑SALA request to the same model, prices both from the provider's own token usage, and reports the 95% confidence lower bound — never the flattering point estimate. Nothing is modeled.

MOCK · referenceESTIMATED · mock provider — this proves the mechanism works, not your bill. A committed LIVE run replaces every number here.

≥ 44.5%

net savings, 95% CI lower bound. Point estimate 54.1%, upper 58.4% — we headline the floor (reference workload, mock provider).

−5.0%0%63.4%

avoided_calls

35.2%

95% ≥ −27.8% · Δ $0.000015

estimated (mock)

eliminated_tokens

60.6%

95% ≥ 30.8% · Δ $0.000351

estimated (mock)

compressed_tokens

57.3%

95% ≥ 57.3% · Δ $0.000771

estimated (mock)

output_restriction

−25.0%

95% ≥ −25.0% · Δ $-0.000012

estimated (mock)

Four ledgers, never summed — each is a separately measured lever, so a dollar can't be double-counted. Reconciliation residual ≈ 0 proves it. These are reference levers: your live rail measures a different four — provider-cached replaces output-restriction — so we never map one onto the other.

Quality gate 100% passedLedgers reconcileModel fingerprint stableCache break-even K=2unproven (mock)A regressed answer voids its saving

scenario	kind	A in/out	B in/out	A cost	B cost	savings ▾	qa
Local arithmetic	resolver	10/15	0/0	$0.000010	$0.00	100.0%	2/2
Tool-heavy request	ran	815/15	190/20	$0.000131	$0.000040	69.1%	4/4
Large uniform table	ran	2,182/15	877/20	$0.000336	$0.000144	57.3%	4/4
Summarize provided text	ran	31/15	31/20	$0.000014	$0.000017	−22.0%	4/4
Constrained structured answer	ran	20/15	20/20	$0.000012	$0.000015	−25.0%	4/4
FAQ (first ask)	ran	12/15	12/20	$0.000011	$0.000014	−27.8%	4/4
FAQ (repeat ask)	cache	12/15	12/20	$0.000011	$0.000014	−27.8%	2/2

cache · amortizedFAQ (repeat ask): cold miss $0.000014 → hit $0.00. Breaks even at K=2 repeats — cache savings are repeat-traffic, not a single-call number, so a cold call honestly reads ≈0%.

Honest accounting

Four separate ledgers. Never summed into one vanity number.

A cache hit removes a whole call. BV‑SALA re‑encodes structured data more compactly. Compression drops redundant words. A provider prefix cache makes tokens cheaper without removing them. Mixing them lies to you — so we don't.

avoided_calls

LLM, retrieval and tool calls that never executed at all.

—

run the optimizer to measure

eliminated_tokens

Information dropped because it simply wasn't needed.

—

run the optimizer to measure

compressed_tokens

Information kept, but encoded with fewer tokens.

—

run the optimizer to measure

provider_cached_tokens

Still in the prompt, but billed or processed more cheaply.

—

run the optimizer to measure

Five lines to start

Run it in your own app. Your prompts never touch our servers.

Drop the SDK in front of your provider. Every response carries a full SavingsReport you can chart — the same numbers the dashboard reads.

// your app — the data plane. LLM traffic goes straight to your provider.
import { createSavingsLayer } from "@bivelio/savings-layer";

const layer = createSavingsLayer({
  provider: yourProvider,
  licenseKey: process.env.BIVELIO_LICENSE_KEY,       // from your dashboard
  licenseServerUrl: "https://savings.bivelio.com",   // license check + kill-switch
});

const { data, report } = await layer.generate({
  model: "openai/gpt-4o-mini",
  messages,
  response: { contractId: "answer:v1", schema, maxOutputTokens: 220 },
});

report.cost.netSavings;   // → your net savings on this request
report.calls.avoidedLlm;  // → whole calls you never paid for

One plan, priced per seat

Pennies to run. Multiples to save.

No free tier, no trial — test BV‑SALA right here in your browser, then subscribe to run it in your app. Priced per seat, in USD.

$4 / seat / month

PRO — billed per seat. Or $39 / seat / year (save ~19%). Enterprise, with packaged features, comes later.

A seat is one active machine or service identity, not a person — billed only if active on 3+ days in a month. CI and ephemeral runners don't count.

Every BV‑SALA optimization, uncapped
Managed semantic cache & learned thresholds
Savings dashboards & team analytics
Your prompts never leave your infrastructure

Get started →

Why the price can be this low

We move no tokens. Your LLM traffic flows straight to your provider — we only see counts, never content.
You self‑host the heavy part. The optimizer runs in your process; our servers just license, learn and report.
Infra scales with revenue. It starts on one small box and only grows once paying customers cover the next tier.
The report is the meter. Billing reads the same SavingsReport you verify above — no black box.

Prove the savings on your own data.

Try the optimizer yourself. Nothing leaves this page.

Both arms metered by the provider's own tokenizer — no modeled baseline.

Four separate ledgers. Never summed into one vanity number.

avoided_calls

eliminated_tokens

compressed_tokens

provider_cached_tokens

Run it in your own app. Your prompts never touch our servers.

Pennies to run. Multiples to save.

Why the price can be this low

Savings Layer

BiVelio

Legal