A customer's AI agent gets stuck in a retry loop late one night, and the usage-based billing system behind your product, whether that's Metronome or a rival, records every call until morning.
Nobody set a cap, so the meter keeps climbing while the customer sleeps. The invoice that follows will be accurate to the cent, and the customer will dispute it anyway because no one on their side approved that spend.
Stripe completed its acquisition of Metronome in January 2026, and since then, a Metronome evaluation has increasingly turned into an evaluation of a Stripe-owned billing stack. That pushes the buyer's question past invoicing: which tool decides whether the next call should run?
If you're comparing Metronome alternatives for usage-based billing in 2026, start with Stigg: it's the best fit for AI products that sell credits or tokens, because it decides in real time what each customer, user, or agent may use, alongside your billing.
The other five each fit a different pricing model. Orb, now owned by Adyen, suits enterprise commitments and drawdowns, while Amberflo ties AI cost tracking to wallet billing.
Lago is the self-hostable open-source engine for hybrid plans. OpenMeter brings open-source metering inside Kong Konnect, and Flexprice offers open-source billing behind a flat monthly fee.
An invoice summarizes what already happened. Usage-based pricing for AI products also needs things that happen before and during each request, and these are the four I'd check in any demo.
Real-time limits: Your product needs entitlements, meaning the rules for what each customer can do and how much, checked fast enough to sit in front of every metered call. A limit has to answer before the call runs, which puts it in the request path.
Credit balances: Prepaid credits need a live balance, rules for which grant burns first, and expiry dates. You also need a clear answer for what happens at zero: block the call, or allow it and track the overage.
Idempotent usage events: Agents retry, networks drop, and clients resend. Each usage event should carry an idempotency key so a retried request is counted once and your customer pays for one call.
Grandfathered plans: When you change pricing, existing customers often stay on their old version for a while (nobody enjoys learning by email that their plan got pricier). You want versioned plans and a controlled way to migrate people later.
You stop it with a check that runs in the request path, before the expensive call fires. The check reads the customer's remaining credits or limit and returns allow or deny, and your code decides what a deny means: block, switch to a cheaper model, or prompt a top-up.
Alerts help, and most tools here send them, but an alert sent overnight lands on a phone nobody is looking at. When you test a vendor, ask where the decision runs, what happens when its API is unreachable, and whether retried events are counted once.
If your customers buy credits and spend them through agents, this is the layer I'd put in front of the model call first.
Fits pricing like: prepaid credits, per-request token or credit burn, seats plus usage, and hybrid plans where an enterprise contract sets the limits.
It's a usage runtime and credits engine that works alongside your billing system, deciding what each customer, user, or agent may do while billing keeps payments and invoices.
What it handles: Your app calls a credit or entitlement check before the expensive operation, and the runtime returns the balance plus an allow or deny decision. The published figure is p99 entitlement checks under 10ms, served from cache.
If the service is unreachable, checks fall back to defaults you configure. Credit grants carry an effective date, expiration, and priority, and expiring or promotional credits burn first.
Reserve-and-settle holds credits up front for agent runs: in the docs' example, 20 simultaneous reservations of 75 credits against a 500-credit balance grant exactly six and deny the other 14.
Every usage report accepts an idempotency key, so a retried request is counted once. Hard limits (the default) deny access, while soft limits let usage continue as tracked overdraft and return a flag your code can read.
Plans are versioned, so existing subscribers can stay grandfathered while new ones get the new terms, and Miro launched its AI pricing in under six weeks. Billing sync is native for Stripe and Zuora. Metronome and Orb run through its workflow integrations, still in beta for select customers.
Its certifications cover SOC 1 Type II, SOC 2 Type II, and ISO 27001.
Price: As of November 2026, Build is a free tier, and Pro runs $499 per month if you pay monthly or $399 per month if you pay annually. Scale is custom, BYOC (Bring Your Own Cloud) begins at $40K a year, and plans are billed on managed entities plus usage events.
Catch: Payments, tax calculation, dunning, and revenue recognition stay outside its scope, which is why it runs paired with a billing system such as Stripe Billing. Recurring credit grants don't roll over, and Zuora, NetSuite, and Salesforce billing integrations sit on the Scale and BYOC plans.
Orb fits best when your largest deals run on annual commitments, minimums, and prepaid drawdowns.
Fits pricing like: committed annual spend with minimums and drawdowns, prepaid credits with allocations, and usage plans that need threshold invoicing when a customer crosses a spend or usage line.
What it handles: Orb ingests usage through an idempotent API or from S3 and GCS, and you define billable metrics in SQL over the raw events. Plan versions let you migrate customers on your own schedule, and before a pricing change goes live, you can simulate its effect on real customer data.
Every plan includes Orb Invoicing and tax integrations such as Avalara and Stripe Tax, payments can be collected through Adyen or Stripe, and finance gets recognized and deferred revenue reports. Orb holds SOC 1 and SOC 2 Type II certifications.
Adyen acquired Orb in mid-2026, and Orb says it continues as a stand-alone product whose customers can keep their preferred payment processor.
Price: As of November 2026, Orb's Business plan charges $6.00 per credit, with each credit covering 0.60% of your billing volume plus event ingestion and other meters. You start with 100 free credits, and the Enterprise tier is priced custom.
Catch: Orb reserves the NetSuite and Salesforce integrations, collections automation, data exports, and its 99.9% SLA for the custom Enterprise plan.
If you process elsewhere, ask how long Orb guarantees multi-processor support, since Adyen names "convergence" of billing and payments as its long-term intent.
Fits pricing like: bulk-buy and wallet-based plans where customers prepay and usage draws the balance down, plus tiers, bundles, and overages.
What it handles: Amberflo brings AI cost visibility and customer billing into one platform, so the usage data that shows what a workload cost also prices it for the customer. Metering runs in real time, usage passes through a rating engine, and wallets draw down as usage occurs.
Each billing period closes with an invoice that generates automatically. Cost Guards trigger an alert when a condition is met on usage or cost across a rolling window, and the default ingestion guardrail allows 1M events per hour per meter, a ceiling Amberflo says it can raise.
Price: Amberflo keeps pricing out of its docs, so you'll be talking to sales before you see a number. In a trial, have the team rate your messiest invoice end to end, then push a burst of events at a test wallet and watch how fast the balance moves.
Catch: Payment collection runs through Stripe, which receives each invoice total. No entitlement or feature-gating product appears in its docs, so plan on enforcing access limits somewhere else.
If self-hosting is a hard requirement, Lago is the first billing engine I'd test.
Fits pricing like: hybrid AI pricing that mixes subscriptions, prepaid credits, and token usage, with input, output, and cached tokens metered separately.
What it handles: Lago ingests events and meters usage in real time, so a wallet balance typically reflects a new event in under a second. Billing and entitlements share one system, and per-subscription overrides leave the underlying plan untouched.
It also covers quotes and order forms, invoicing, and tax, with Adyen, GoCardless, Stripe, or MoneyHash collecting the money. Official SDKs cover JavaScript, Python, Ruby, and Go, and Lago holds SOC 2 Type II.
Its named customers include CoreWeave, Mistral AI, and PayPal, and the company is independent.
Price: As of November 2026, Lago Cloud's Launch plan carries no minimum. Your first $100K of billings and 5M events cost nothing (a one-time allowance), then the rate is 0.7% of billings plus $20 per million events. Scale sets a $12,000 annual minimum, and the self-hosted AGPLv3 edition is free.
Catch: The open-source edition comes with AGPL obligations and a limited feature set, and self-hosting means your team operates the deployment. Getting onto Cloud is team-assisted today, and you'll need Scale or Enterprise for SSO, custom roles, and audit logs.
Fits pricing like: per-request, per-token, and per-millisecond usage on APIs and LLM traffic, where event volume is the hard part.
What it handles: OpenMeter ingests any usage you can express as CNCF CloudEvents, and for LLM traffic it counts tokens per model, per customer, and per request. Trigger.dev uses it for per-millisecond compute billing.
Since Kong acquired OpenMeter in September 2025, the hosted version is sold as Konnect Metering & Billing, and the project remains open source under Apache 2.0.
The hosted side adds entitlements as plan components, hard and soft quotas, prepaid credits, itemized invoices, and sync to Stripe and NetSuite.
If Kong already runs your gateway, you can apply token-based rate limits and budgets there before costs reach your LLM. SDKs cover Node.js, Python, and Go, and the company is SOC 2 Type 2 certified.
Price: Kong's price as of November 2026 combines 0.4% of billing volume (up to $250,000) with $20 per million ingested events (up to 10M), and Kong Konnect Plus begins at $25 a month plus usage. The Apache 2.0 project is free to self-host.
Catch: Because the hosted product sits inside Kong Konnect, gateway-native enforcement expects you to run Kong. If you self-host the open-source project, your team runs and maintains it.
Flexprice suits teams that want a predictable platform fee while their usage pricing is still finding its shape.
Fits pricing like: usage-based, credit-based, and hybrid plans, including commitments and overridden prices for negotiated contracts.
What it handles: Flexprice meters usage in real time and lets you change usage limits and feature access without deploying code. It configures credit grants, expirations, rollovers, and overage rules.
Paid tiers add wallets, entitlements, coupons, add-ons, and tax management, along with a customer portal, PDF invoices, and Stripe payments that include dunning and retries.
It's open source under AGPL-3.0, can run entirely inside your own infrastructure, and supports MCP, a CLI, and multiple SDKs. Flexprice is SOC 2 Type II compliant and says it can sit on top of an existing billing tool.
Price: As of November 2026, Launch is a flat $99 a month for up to $300K in billing volume. Growth adds 0.40% of billing on top of a $500 monthly fee, and Scale is custom with zero revenue share. (Revenue share is a vendor congratulating you on your growth and then asking for a slice of it.)
Catch: Flexprice is a younger product, so check that every integration you depend on exists today. VPC, on-prem, or air-gapped deployment, SAML SSO, and SCIM sit on the custom Scale plan, and Launch excludes premium support.
Start from the pricing model you sell today, then the one you expect to sell in a year, and shortlist the tools that fit both.
Credits burned by agents or per request: start with the runtime at #1 and keep your billing engine downstream, because the hard part is the allow-or-deny decision before the call.
Annual commitments and enterprise drawdowns: Orb's contract tooling is the closest match, and you should ask about processor support beyond Adyen.
AI cost visibility next to customer billing: Amberflo puts both on one platform, so test its rating engine against your messiest invoice.
Open source or self-hosting required: Lago or Flexprice under AGPL, or OpenMeter's Apache 2.0 project if AGPL terms worry your legal team.
Usage that already flows through Kong: OpenMeter, since enforcement can run in the gateway you already operate.
The riskiest invoice in any migration is the first one a new tool builds from old events. If you import historical usage and let the new system rate it under its own rules, the same events can produce different totals, and customers notice when last month's numbers change.
Keep the old tool as the system of record for every period you've already invoiced. Export the raw events anyway, since you'll want them for audits and for testing your new price book, then run one billing cycle in parallel and compare totals line by line before you cut over.
When a number differs, find out whether the rule changed or the data did before anyone sends an invoice.
Unbilled usage stays in the old system until you either invoice it there or move the raw events to the new tool and rate them again. Pick a cutover date, close the period in the old tool, and send only post-cutover events to the new one, so no usage gets billed twice or skipped.
No open-source or self-hosted edition appears on Metronome's pricing, product, or integrations pages, so plan on its hosted service. If you need code you can inspect and run yourself, Lago and Flexprice publish theirs under AGPL-3.0, and OpenMeter's project uses Apache 2.0.
Stigg enforces usage limits in real time by checking entitlements and credit balances in the request path, before the expensive call runs. OpenMeter can apply hard and soft quotas in the Kong gateway, while Orb leans on threshold invoicing and spend alerts.