Savings Layer is part of the BiVelio platform

Savings Layer Measured on every call.

Optimize your LLM calls. Verify every saving.

Run the open-source engine on your own infrastructure. Connect Savings Cloud to centralize measurement and management for your team, on flat-fee plans.

The Savings dashboard with the public bench's data: the measured saving and its split by model.

Receipt of one measurement

GSM8K · Sep 28, 2026 · SDK 0.4.31

Model
gpt-4o-mini
Sample
100 families · 200 requests per arm
Repeated requests
30%
95% interval
25.7% – 36.7%
Lower cost
31.5%
SDK A/B

Live A/B against the provider, at versioned list prices. Not an invoice reconciliation. How it was measured

Dashboard with the public bench's data
Measured live with each provider's own counterOpenAIAnthropicGoogle
Savings Layer, in action

A thin layer between your code and your provider. That's the Savings Layer.

It runs inside your infrastructure, sends the provider only what has to be paid for, and returns a receipt with every response. Your prompts go straight to the provider you configure; Savings Cloud is optional and receives counts and cost, not their content.

  1. The call enters the layer It leaves your app or your agent without leaving your infrastructure.
  2. Repeated · answered without calling the provider The layer recognizes it in its memory and returns the stored response.
  3. Data compacted · fewer tokens A new request travels lighter to the provider you configure, and the response comes back through the layer.
  4. Counts and costs only · never prompts Savings Cloud is optional: if you connect it, it receives figures, not content.

Your calls

Your appSDK
Your agentlocal gateway

Savings LayerOn your infrastructure

Memory of recent requests
Data compaction

The provider you configure

Savings Cloud OptionalCounts and costs only · never prompts

Compatible does not mean measured. What is measured with each provider

Illustration of how it works.

See how it works →What BiVelio receives, field by field →

Open engine, optional cloud

Two ways to use Savings.

The engine is the same. What changes is where you manage it and what leaves your infrastructure.

Open source · Apache-2.0

Community

The engine, the SDK and the local gateway. For one person or for a whole company.

  • No account or card
  • No telemetry by default
  • SDK and local gateway
Where it runs
On your infrastructure: the SDK inside your process, or the gateway on your machine.
Account
None needed. No sign-up, no card and no BiVelio key.
Telemetry
None by default.
Management
Locally: every call returns its SavingsReport.
Optional · starts free

Savings Cloud

The dashboard that centralizes measurement and management for your team.

  • Starts free with Measure
  • Team dashboard
  • Flat-fee plans
Where it runs
The engine stays on your infrastructure. Cloud adds the dashboard.
Account
A Savings account and a service key.
Telemetry
Counts and cost, depending on your plan. Your prompts don't pass through BiVelio's servers.
Management
Team dashboard, organization and licences.
Where it plugs in

Where do your tokens go: your app, or your coding agent?

They are two different pieces with different scopes. The deciding question is who wrote the code that calls the API.

SDK · in-process

Inside your application

For Node.js apps, in JavaScript or TypeScript, that call an LLM API with your own key and pay for every token.

  • Requirements: Node.js and your provider's API key.
  • The package's savings levers run here.
  • Every call returns a SavingsReport, with measured and estimated kept apart.

Try one lever on your data

app.ts
# 1 · the package
pnpm add @bivelio/savings-layer

# 2 · wrap the call — this is the entire code change
const layer = createSavingsLayer({ provider });
const { text, report } = await layer.generate({ model, messages });

console.log(report.cost.netSavingsRatio)  // ← the saving, measured
  • dependencies: {}Zero runtime dependencies
  • report.cost.basisOne SavingsReport per call
  • app → provider apiYour traffic still goes direct
  • layer.generateStream(request)Streaming, same report

Gateway · local proxy

In front of your application or agent

For developers who run Claude Code, Codex CLI or another coding agent on an API key.

  • Requirements: Node.js with npm, and a coding agent that uses its own API key.
  • It meters traffic for models with a list price.
  • The npm package's levers do not run here; the gateway's own apply only to Anthropic traffic, with a PRO license.
Terminal
npm i -g @bivelio/savings-layer
# optional: cloud dashboard
export BIVELIO_SERVICE_KEY=bvk_…
bvsala claude
No promises: measurements

How do we know it saves? We measure it — and we publish every trial, one by one.

It depends on how much your traffic repeats. Every battery, with its full report: public bench.

See reproducible results
OpenAISDK

Receipt of one measurement

GSM8K · Sep 28, 2026 · SDK 0.4.31

Model
gpt-4o-mini
Sample
100 families · 200 requests per arm
Repeated requests
30%
95% interval
25.7% – 36.7%
Lower cost
31.5%
SDK A/B

Live A/B against the provider, at versioned list prices. Not an invoice reconciliation. How it was measured

Why this number, and its limits

lower cost at versioned list prices (95% interval: 25.7% to 36.7%) on a stream where 30% of requests are exact repeats — 100 families / 200 requests per arm, GSM8K, SDK 0.4.31

Not a forecast: what you save depends on how much your traffic repeats. Only what your provider's counter confirms counts in your savings report — including repeated requests served without calling your provider, each valued with the counter of the identical request that already ran.

AnthropicSDK

Receipt of one measurement

GSM8K · Sep 28, 2026 · SDK 0.4.31

Model
claude-sonnet-5
Sample
31 families · 62 requests per arm
Of which, cache-pricing contrast (approx.)
41%
95% interval
69.3% – 79.6%
Lower cost
75.1%
SDK A/B

Live A/B against the provider, at versioned list prices. Not an invoice reconciliation. How it was measured

Why this number, and its limits

reduced flow, not the OpenAI stream: 31 families / 62 requests per arm (95% interval: 69.3% to 79.6%). About 41% of A's list-price cost is a descriptive net cache-pricing contrast (arm B's cache reads valued against ordinary input pricing), not an independently identified causal effect nor a pre-existing Anthropic discount

Claude CodeGateway

Receipt of one measurement

Gateway · Oct 3, 2026

Sessions
51
Requests
19,779
Days
13
95% interval
1.50% – 3.57%
Net
2.49%
Gateway replay

Counterfactual replay with a partially reconstructed baseline. It is not a live A/B, billed savings or an invoice reconciliation. How it was measured

Why this number, and its limits

estimated net reduction in list-price cost with PRO's cache TTL change, after induced cache misses (95% interval: 1.50% to 3.57%). Counterfactual replay with a partially reconstructed baseline on 51 Claude Code sessions from one organization; not a live A/B run or invoice reconciliation.

Not a forecast for your team: it is one organization, one provider and list prices, and it is a small saving that shows where Anthropic spend is high. With PRO, on Anthropic models, it moves cache writes from the one-hour TTL to the five-minute one, only on turns where a large part of the prompt is new — such as a session's first turn or a compaction — and leaves the chained turns of a running session as they are. That changes the price, never what the model reads, and only what your provider's counter confirms counts in your savings report.

The TTL bench report: what was simulated, the per-model breakdown and its limits →

Plans

Open-source engine. Cloud control for your team.

The package is Apache-2.0 open core: no key needed, every local lever on, zero telemetry. The cloud is optional — Measure is the free plan, Pro and Team are billed per seat, in EUR, whether or not you save, and carry no commission on your savings.

Community is not a plan. The open-source engine is free and runs without an account: install it whenever you like. The four plans below are Savings Cloud.

Billing period
  • Measure

    Freeno card required

    Free cloud plan, in shadow mode.

    Measure without changing anything: your requests reach your provider exactly as you build them, and the dashboard shows what would have been saved.
    Seats
    Up to 5
    Requests per day
    Up to 20,000
    Commission on savings
    None
    Requires @bivelio/savings-layer 0.4.33 or later.See terms
  • Pro

    €15per seat per monthMonthly charge€150per seat per yearYearly charge

    A seat is a machine or a service identity, not a person.

    Every optimization, the per-request receipt and the dashboard. A flat fee, due whether or not you save.
    Seats
    As many as you subscribe
    Commission on savings
    None
  • Team

    €129per monthMonthly charge€1,290per yearYearly charge

    Includes 10 seats: machines or service identities.

    The same as Pro, for a fleet. The fee covers the included seats even if you use fewer.
    Seats included
    10
    Additional seat
    €12 per month€120 per year
    Commission on savings
    None
  • Enterprise

    Customby contract

    For what online checkout cannot cover.

    More seats, a contract of your own or agreed invoicing terms.
    Seats
    More than 500

The fee does not include your LLM provider's usage: you keep paying that yourself, on their terms. Amounts exclude VAT.

A seat is one machine or service identity, not a person. You pay for the seats your subscription covers; your dashboard shows how many were in use this month (active on at least 3 days; CI and ephemeral runners don't count).

VAT not included. If your company has an EU VAT number, the reverse charge applies.

Already on the earlier plan with a commission on measured savings? You keep it unchanged; its terms are on the pricing terms page.

See the published pricing

Buying questions

What to know before you pay.

The contractual detail is in the published pricing; what the Software sends to BiVelio is on the privacy page.

Book 15-min callWe'll tell you what you'd save on your traffic, without installing anything.
What exactly do I pay for?

The open-source engine is free. Savings Cloud has a free plan, Measure, and two flat-fee plans billed per seat, Pro and Team. The fee is due whether or not you save, carries no commission on your savings, and does not include your LLM provider's usage: you keep paying that yourself, on their terms.

What is a seat?

A machine or a service identity that runs Savings. It is not a person: the people on your team are managed separately, in the dashboard. CI runners and ephemeral runners do not count as seats.

How are Measure and Community different?

Community is the open-source engine on your infrastructure: no account, no BiVelio key and no telemetry by default. Measure is the free Savings Cloud plan: it needs an account and runs in shadow mode, so it does not change your requests or serve responses from cache; your dashboard shows what would have been saved.

Do I have to change provider?

No. You keep your provider and your own key, and your calls still go straight to it. Compatible does not mean measured: the public bench says which models have a published figure.

What data does BiVelio receive?

With Community, none by default. Once you connect Savings Cloud, your dashboard receives counts and cost according to your plan; your prompts don't pass through BiVelio's servers. The privacy page lists it field by field.

Do you guarantee savings?

No. What you save depends on your traffic; for example, on how much of it repeats. We publish every measurement with its model, its sample and its interval, and Measure lets you see what would have been saved on your own traffic before you pay anything.

Measure first. Decide later.

Measure is free and asks for no card: it runs in shadow mode and shows what would have been saved on your own traffic. The engine is open source, and its code is in plain sight.

Install open source

savings.bivelio.comPro · €15 / seat / month
Company & ecosystem

BiVelio is a member of the NVIDIA Inception Program.

NVIDIA's program for startups building with AI.

About the program
NVIDIA Inception Program