A base fee, and a share of what we actually save you.

What counts as a realized saving

Only two things, both measured at the gateway on the request itself — not modelled, not projected, not estimated from a sample:

• A cache hit. An identical deterministic request served from cache costs nothing. The saving recorded is exactly what the original call cost.
• An automatic downgrade. When routing moves a call to a cheaper model, the saving is what the originally requested model would have cost for those same token counts, minus what the cheaper model actually cost.

A recommendation nobody acted on is not a saving. A projection is not a saving. Both are on your dashboard, clearly separated from the realized figure the fee is calculated on.

Changes you make from our suggestions

Some savings need a change in your code: switching a call to a cheaper model, shortening a prompt, lowering an output limit, or replacing a call with code ModelGate wrote and verified. When you make one of those changes the way we suggested, it’s billed at the same 30% — for 3 months after the change, then it stops.

• Measured for model, prompt and output-limit changes: the calls still pass the gateway, so the saving is what those calls would have cost at the cost per call before the change, minus what they actually cost.
• Estimated for a call replaced with code: it no longer reaches the gateway, so the saving uses the call volume and cost per call frozen before the change, minus anything that still runs. It needs a commit, our merged pull request, or a suggestion you accepted as evidence.
• Calls the gateway already saved on itself — cache hits, automatic downgrades — are never counted twice.
• Tracking is on by default and can be turned off per project; changes already detected stop billing from then. You can mark any change as not ours and it’s excluded.

Connecting GitHub is optional: it points each suggestion at the line of code and opens the change as a pull request in one click. ModelGate reads only the repositories you choose, never stores your source, and never pushes to your main branch. See security.

How it is billed

• The $99 base is a PayPal subscription you approve once.
• The 30% success fee is a separate PayPal invoice raised once a month for 30% of that month’s realized savings.
• Changes made from our suggestions appear as their own line on the same invoice, for 3 months after each change.
• Pro includes 1 project — one repository. Each extra project adds $29/month to the same invoice. Agencies pay per client instead, with projects unlimited.
• Months where the fee would come to less than $1 are skipped entirely — nothing is charged.

Payment processing is configuration-driven; if it is not yet live in your region or account, the billing endpoints say so rather than failing quietly, and the status page reports billing as not enabled.

For agencies running AI for clients

The agency plan is $199/month with 5 clients included, then $29 per extra client a month — same 14-day trial, same 30% success fee.

• Clients. Group projects under each client, with a monthly spend cap per client enforced at the gateway.
• One invoice, itemized. The success fee is listed per client, so you can see what each client’s savings cost — and a per-client savings report (print or CSV) to send them, without ModelGate’s fee on it.
• Or the client pays. Send a client a billing link and they approve their own PayPal subscription ($29/month + 30% of their savings). You never see their card, and that client doesn’t use one of your seats.
• Team and client logins. Invite staff as admins or members, and give a client read-only access to their own projects, under your agency name and logo.

The Reliability add-on

Reliability is a separate, flat $49/month add-on — a different buyer (whoever owns whether the bot behaves) and a predictable price, not tied to savings. It unlocks the judge-based detectors (self-contradiction, instruction drift, grounding), the confusion map, response capture and reliability alerts.

• On every plan. The deterministic detectors — format breaks, degenerate output, refusals — and the consistency score run on every plan. That’s the passive score.
• Metered active tests. Each account gets 20 active Consistency Test runs a month included; beyond that they’re $0.50 per run, because each one spends your provider tokens through your own stored key. Overage appears on the same monthly invoice as the success fee.
• During the free trial the add-on is included, so you can see the reliability report before you decide.

The flat $49 is a PayPal subscription you approve once; the metered test overage is billed on the monthly invoice. See how reliability works.

What happens when the trial ends

The trial starts when you add a card to turn savings on. Because your card is then on file, the trial converts to the paid plan automatically on day 14 — nothing in your setup changes, and the first base fee is charged then. If you cancel before day 14 you are never charged. Cancel later and you keep your ledger — logging, pricing, the waste score, the recommendations and detect-mode guardrails carry on; what stops is the part that changes your requests: automatic caching, automatic model routing, enforce mode on the guardrails, and generating new code replacements (scanning for them carries on).

When ModelGate is worth it — and when to wait

Two things decide it, and we would rather say so now.

• Spend. Below roughly $500 a month on model APIs, even a large percentage cut is a small absolute number — rarely worth changing how you call your models.
• Traffic history. Waste is a pattern. Week one prices your spend to the token; what repeats, what is oversized and what is overpaying only becomes visible once there is volume behind it.

Questions

The Terms of Service cover fees, cancellation and refunds in full. Anything else: support@modelgatehq.com.