The LLM cost layer anyone can install — licensed to runCut the cost of your AI applications without switching providers or degrading quality.
Savings analyzes every request, avoids unnecessary executions, strips redundant context and picks the cheapest option that still meets your quality contract. Every published figure is measured against the provider's own meter.
Install in five lines. The license activates from your dashboard.
It installs in front of these providers
gpt-4o-miniAnthropicGeminiresolvers · singleflight · avoided_calls
pruned tools · retrieval · redundant context · eliminated_tokens
leaner encoding · only when it pays off · compressed_tokens
Cheapest execution wins
OpenAI gpt-4o-mini
≥ 20.0% less bill, verified at 95% · against an unoptimized integration, no streaming
The second control
Against an integration that was already hand-optimized
Here every scenario also runs against an integration a diligent engineer would already have tuned: the provider's native cache configured on the stable prefix, and only the tools the task genuinely needs, picked one by one. Everything else — contract, cap, temperature and seed — is identical.
23.2%
Against an unoptimized integration
Almost everyone benchmarks against a naive integration, and against that anyone looks good.
22.9%
Against an already hand-optimized integration
That second comparison is the only one that measures the layer's own merit: what is left to save once the low-hanging fruit is gone.
We always announce the smaller of the two figures, and both come out of the same public artifact: promising above the second control would be selling you, as our merit, what you already had.
measured on gpt-4o-mini · 95% floor · d1c8233 · 2026-09-12 · 193
Avoid the call entirely
resolvers · singleflight
avoided_calls
Remove what never needed sending
pruned tools · retrieval · redundant context
eliminated_tokens
Compress what remains
leaner encoding · only when it pays off
compressed_tokens
Each step falls into ONE different ledger, so these figures do not add up to each other. A step whose ledger does not clear zero does not appear here. A percentage without its model next to it means nothing: the same lever can save on one model and cost on another.
How savings are measured — and every trial, published one by one.
Both arms are metered by the provider's own counter — no modeled baseline. For each scenario the bench sends the unoptimized request and the BV-SALA request to the same model, prices both from the token usage the provider itself reports, and aggregates the scenarios with a stratified bootstrap, weight one per scenario: the headline therefore does not depend on how many repetitions each case happened to get. The pooled aggregate is still in the artifact as aggregatePooled, labelled for what it is: not the headline. We publish the floor of the 95% confidence interval, never the point estimate.
- Same model, two arms: the naive request vs. the BV-SALA request.
- Cost priced with the provider's own tokenizer — no modeled baseline.
- We publish the floor of the 95% confidence interval, never the flattering point estimate.
- Two control arms: against an unoptimized integration, and against the same integration already hand-optimized. Only the second figure is the layer's own merit.
- Ledgers that do not clear zero are published all the same, with their label.
Download the raw trials (CSV) d1c8233 · 2026-09-12 · 193 · gpt-4o-mini
Check it with your own numbers, before installing anything
The verifier runs the real optimizer in your browser: pick an example or paste your own data, say how much you use the API, and see the projected saving computed exactly the way your receipt will compute it later. Nothing you paste leaves the page.
The verifier lives inside the product, in your Savings dashboard — it is not a separate marketing calculator.
Built for developers
Fewer bill headaches. Without changing provider.
Three facts about the package. None of them is a promise.
Install in five lines
npm install @bivelio/savings-layerLicensed SDK for Node.js; the license activates from your dashboard. Works with any endpoint compatible with the OpenAI API and with the native Anthropic adapter.
npmjs.com/package/@bivelio/savings-layer ↗- 01
Zero runtime dependencies
A single ESM bundle. Nothing gets installed next to your code that we did not write; the exact tokenizer is an optional peer.
- 02
One SavingsReport per call
Ledgers that are never summed, the cost with its basis — measured or estimated — and traces per stage. It is the same thing your dashboard reads.
- 03
Your traffic still goes direct
The SDK wraps the call inside your process. Your prompts never pass through BiVelio: we only ever receive counts and cost.
Three lanes that are never summed — and only one is billed
Every receipt splits the saving by how well it can be proven, because not every proof is worth the same. Adding the three together would give a bigger, less certain number: that is why they are never added.
Measured
The provider declared the same tokens and a different rate was paid: it is a subtraction between two public prices, and you can reconcile it against your provider's console, line by line.
measured · billedFrom the provider
Their own counter says how many tokens they dropped. The count is theirs, not ours; what it is worth in money depends on the price those tokens would have been paid at, and that part is our estimate. It is shown, not charged.
estimated · not billedHypothetical
What a policy would have saved had it been applied. It comes from simulating the turn: nobody ran anything. It can come out negative, and it is shown anyway. It is never charged.
simulated · not billedOnly the measured lane is billed. Estimated and hypothetical never enter your invoice, even when they come out bigger — and if a figure cannot be reconciled against your provider's invoice, it says so on the very line where it appears.
bvsala · coding terminals
Claude Code on an API key? See its real spend, token by token.
One command puts a local meter in front of your terminal. No code changes, nothing to configure — and your prompts never leave your machine: only counts and cost reach the dashboard.
One command, zero configuration
bvsala claude
bvsala doctorIt works with Claude Code, Codex CLI, Goose and Qwen Code — and bvsala doctor prints the exact configuration for the rest.
- The gateway comes up on an ephemeral port, launches your terminal pointing at it and shuts down when you exit.
- Your spend shows up in the dashboard token by token, counted with your own provider's counter.
- bvsala doctor only checks for existence: its report shows paths, never the contents of your configuration files.
Today the gateway measures; automatic optimization there is not in place yet.
savings-mode · free skill
The skill that stretches your subscription caps
No invoice doesn't mean no pain: the five-hour window and the weekly cap. savings-mode is a pure-instructions skill — no proxy, your traffic untouched — that cuts the two masses that actually burn cap: model output and tool results.
Claude Code
One command: it installs into ~/.claude/skills and switches itself on when the conversation turns to limits or saving tokens.
~/.claude/skillsCodex
It has no skills, so the same discipline ships as a marked block inside your AGENTS.md: always on, and switched off by deleting the block. Your AGENTS.md outside the marks is never touched.
AGENTS.mdClaude.ai
The skill travels as a zip: you enable “Code execution and file creation” in settings, upload it under Customize → Skills, and it appears with its own toggle, private to your account. The download asks for your Savings account, which is free; the skill itself carries no telemetry at all.
savings-mode.zip
The skill's own effect will be published once it is A/B-measured — never before.
It's free; what you pay Savings for is knowing how much it saves you.
The commercial model
How what you pay us is calculated
A fee per seat, and a commission that comes out of the measured saving — never out of your pocket.
90%
You always keep 90% or more of every euro saved.
The commission only exists when your measured saving exists, and no bracket goes above the top rate.
Only on measured savings
After that, commission only on measured savings: the kind you can reconcile against your provider's invoice, line by line. Estimated savings are shown, but never billed.
Marginal brackets
Marginal brackets, each bracket at its own rate and from the first measured euro: the more you save, the smaller the percentage we take.
No lock-in
A month with no savings? That month there is no commission. No lock-in — and your first variable invoice only contains numbers you already saw on your statement.
What a seat is
A seat is one active machine or service identity, not a person, and it is billed only if it was active on three or more days in the month. CI and ephemeral runners do not count.
Your first month is calibration
There is no free tier and no trial: what you can check before paying are the product's public measurements. Your first month is calibration, and it is paid: you pay the seat only, we measure all of your traffic and we charge no commission — at close you get the exact statement of what it would have cost.
- 1
Connect
Savings sits in front of your model calls and starts counting with your own provider's meter.
- 2
Calibrate
For the first month you pay the seat only: we measure all of your real traffic and charge no commission.
- 3
Decide
At close you get the exact statement of what it would have cost. From then on, the commission comes out of the measured saving, never out of your pocket.
The exact rates live in the product's price list and are not copied here, so nobody reads a figure on this page that has already moved.
bivelio-savings-layer · apps/web/lib/billing/tarifa.ts · caa7310 · 2026-09-12
Frequently asked questions about Savings
How BV-SALA measures and cuts your LLM bill.
Does it really reduce my LLM bill?
Yes, and we don't ask you to take it on faith: every figure on the page is measured against the provider's own meter, with its confidence interval right next to it. We report the 95% floor —the conservative saving—, never the flattering estimate.
How is the saving measured?
With an A/B bench: for each scenario the unoptimized request and the BV-SALA request go to the same model, both are priced with the provider's own meter, and we report the lower bound of the 95% confidence interval. There are two control arms — an unoptimized integration and the same integration already hand-optimized — and scenarios are aggregated with a stratified bootstrap, weight one per scenario. No modeled baseline, no assumptions.
Does it degrade the quality of the answers?
No. On each request, BV-SALA chooses the cheapest execution that still meets your quality contract. If cutting costs would compromise the quality you've set, it doesn't: the saving never comes at the expense of the result.
Which providers does it work with?
Two things, kept apart. Compatibility: it sits in front of your provider and works with any endpoint compatible with the OpenAI API, plus the native Anthropic adapter. Measurement: today there is exactly one published figure, OpenAI gpt-4o-mini, measured against the provider's own meter. A brand being compatible does not mean it has been measured; when we measure another, we will publish its figure with its model next to it.
How is it installed?
In five lines. It's a licensed SDK for Node.js (@bivelio/savings-layer); the license is activated from your dashboard and works in front of your current integration without rewriting it.
How much does it cost?
A fee per seat plus a tiered commission on the measured saving, from the first euro. A seat is one active machine or service identity, not a person. There is no free tier and no trial: your first month is calibration — you pay the seat only and we charge no commission. And in a month where you save nothing, you pay no commission.
Book a 15-min call
We'll tell you what you'd save on your traffic, without installing anything.