Start from who holds the demand data
At a renewal, the provider's account team arrives with a usage chart and a growth assumption. The FinOps Framework's Rate Optimization capability describes the counterweight: Procurement leads the negotiation, and the FinOps team supplies data on historical and projected consumption and highlights the cost and business drivers behind it.1 Procurement and Finance are both core personas in the Framework, so this is a shared job, not an engineering favor.2
The same page marks the maturity step that matters here. At "Walk", pricing, commitment and alternative purchasing options are compared against validated demand before purchase; at "Crawl", decisions default to what the team already knows, with no comparison step.1 Validated demand is what this guide builds: a short pack, drawn from gateway records, that you can put on the table whether you are talking to OpenAI, Anthropic, Google or a cloud provider that hosts the same models.
Two cautions before the data. First, the Framework notes that negotiated rates are often private and covered by non-disclosure terms, so check your agreements before sharing anything.1 Second, a gateway only sees traffic that goes through it. Anything a team calls directly, or through a cloud console, is missing unless you account for it separately (see the cut on traffic outside the gateway below).
Know which number you are looking at
A gateway records each request's model, latency, unit counts and an estimated provider cost.3 "Estimated" matters in a negotiation. The gateway prices each request with your own rate for that model and provider if you entered one, otherwise the vendor's list rate, and then reads the cost your provider reports for the same UTC day to store both figures side by side.4 The provider's reported figure is what you will be held to.
For a commit this gives three tiers of confidence:
- Reported cost from the provider's own cost API, which is what the provider counts against a commitment. INFRO's reconciliation reads it for OpenAI and Anthropic.4
- Estimate at your negotiated rate, which is good once the rate is entered correctly.
- Estimate at list rate, which overstates spend if you already have a lower contract rate and is a planning figure only.5
Do the reconciliation before you build the pack. A steady one-direction gap is usually a contract rate the gateway was not told about; reconciliation names the likely cause and suggests a rate.4 Also note what is left out: requests answered from the response cache never reached the provider, so they are excluded from the comparison.4 If the vendor's chart and yours disagree by more than the unpriced share, resolve that first, because a disputed baseline weakens every other number.
The seven 90-day cuts
Ninety days is long enough to include a month-end, a quarter-end and one launch or incident, and short enough that the model mix still resembles today's. The usage API takes granularity of hour, day or month, a group_by such as model, provider, project, key, feature or environment, and from and to timestamps; one range may not exceed 400 days and remains subject to your plan's history limit.6 Each bucket carries requests, units, cost, unpriced requests, error rate and latency percentiles.6
| Cut | How to pull it | What it supports |
|---|---|---|
| 1. Daily spend by provider | granularity=day&group_by=provider | The floor, median and peak day: what a commit can safely cover |
| 2. Spend by model | group_by=model, 90-day total | Which models carry the commit, and which rate tiers matter |
| 3. Hourly units by model | granularity=hour&group_by=model | Peak versus typical load, the input for provisioned capacity |
| 4. Spend by project or feature | group_by=project or feature | Which workloads are durable and which are experiments |
| 5. Weekly trend | Cut 1 summed by week | Whether demand is growing, flat or shrinking |
| 6. Unpriced and outside-gateway share | unpriced_requests field; provider-reported cost vs estimate | How much of the baseline is uncertain |
| 7. Cache and fallback shares | Cached-share and fallback signals | How much of the volume the provider would bill at all, and where it lands |
Cut 3 needs care. Hourly buckets hide bursts inside the hour, so they understate the minute-level peak a provisioned or rate-limited deployment must cover. For that, pull the request log (GET /v1/requests/export returns one request per row as CSV, up to 50,000 rows per file) and bucket by minute yourself.7 Percentiles do not average either: a week's p95 is not the mean of seven daily p95 values, so ask for the period's own bucket when you quote a latency figure.6
import datetime as dt, os, requests
BASE = "https://api.infro.io/v1/usage"
HDR = {"Authorization": f"Bearer {os.environ['INFRO_API_KEY']}"}
def window(start, end, **params):
q = {"from": start.isoformat(), "to": end.isoformat(), **params}
r = requests.get(BASE, headers=HDR, params=q, timeout=60)
r.raise_for_status()
return r.json()
end = dt.datetime.now(dt.timezone.utc).replace(minute=0, second=0, microsecond=0)
start = end - dt.timedelta(days=90)
# Usage buckets come back together with no cursor, so split into 30-day windows.
rows = []
cursor = start
while cursor < end:
nxt = min(cursor + dt.timedelta(days=30), end)
rows += window(cursor, nxt, granularity="day", group_by="model")["buckets"]
cursor = nxtThe response shape here is for orientation; check the field names against the API reference before you depend on them. The reason for the loop is the documented behaviour: a query returns all its buckets together and there is no cursor pagination for usage, so large analyses are split by time window.6
Cut 6 deserves its own check: what the baseline leaves out
Two gaps can quietly shrink or distort the baseline you commit against.
Unpriced traffic. Usage analytics counts unpriced requests explicitly so that an incomplete estimate is not mistaken for a complete bill.6 If a new model was added without a rate, its spend is missing from your cost column but present in the request count. Quote the unpriced share beside every total.
Traffic outside the gateway. INFRO's commitment tracking notes that an estimate can miss both unpriced traffic and traffic that does not go through it.5 Where the provider's reported cost is available, use that for the commit baseline, since it includes everything on the account. If a vendor's invoice is larger than your gateway total, the difference is either unpriced, direct-call or cached-request accounting; find out which before you accept a vendor number or dispute one.
If part of your usage goes through a cloud provider's hosted models, the commit question changes: whether that spend burns a cloud commitment is a separate contract question, covered in Do Bedrock and Azure AI calls burn your cloud commit?. Keep a per-provider column so each negotiation sees only its own spend.
Price-change exposure: a table, not a worry
A renewal fixes some prices and leaves others open. Exposure is the part of your 90-day spend that a rate change, a promotional price ending, a model retirement or a snapshot swap could move. Build it from cut 2, one row per model:
| Model | 90-day spend | Share of provider spend | Rate basis | Retirement or step-up date |
|---|---|---|---|---|
| model id | from usage | computed | list or negotiated | from vendor notice |
The two right-hand columns are the ones procurement can act on. If a model's rate basis is list, the contract may not cover it at all; if its retirement date falls inside the term, you are negotiating for a model that will be replaced by a successor with unknown pricing. INFRO records vendor deprecation and shutdown dates and warns organizations still sending a retiring model, which gives you the dated column without a manual calendar; see Introductory AI model prices: how to budget for the step-up for the budgeting side.
Cost also moves with no price change at all. Traffic-change findings flag a model answered by a new snapshot (at least 20% of the day's successful requests and at least 50 of them), output tokens per request up 15% or more, the cached share of prompt tokens down 10 points or more, and fallback share up 5 points or more, each with an estimated weekly cost change.8 These findings read request metadata, never prompts or answers, so they cannot say whether quality moved.8 Why it matters for a commit: if your 90-day window contains a lost prompt cache or a longer-answer snapshot, the baseline carries a cost step that may not persist. Mark those dates and decide whether to normalise them before sizing.
Fallback options: leverage only if it is real
Procurement leverage comes from credible alternatives. In gateway terms that means: routes with more than one target, failover that actually triggered, and a record of what the fallback cost. Requests served by a fallback are tracked as a share of traffic, and a rise of 5 points or more is raised as a finding.8
Bring three things:
- Fallback share over 90 days, per route. A non-zero share shows the path works. A fallback share that rose by 5 points or more is itself a finding.8
- The price gap on that path. Fallbacks that land on a dearer model change the bill; What does failover do to your bill? covers the arithmetic.
- An equivalence test. A fallback you have not tested against your own quality bar is a bargaining claim you cannot defend; see Is your fallback model good enough?.
State the limits honestly in the room. Moving a large share of volume to another provider is a migration with testing, rate-limit and data-handling work, not a switch. What the evidence supports is a narrower claim: which workloads could move, in how many weeks, and what share of spend that is.
Turn the cuts into a commit you can defend
INFRO's commitment forecast is the amount consumed so far plus the trailing 28-day daily mean times the days remaining, with a band from the variation in daily spend over those 28 days. It is a straight-line projection that does not model seasonality, planned launches or price changes.5 Treat any vendor projection built the same way with the same suspicion. The FinOps Forecasting capability makes the point from the other side: forecasts should include rate optimisation and be tracked against actuals within an agreed variance, with a revisit point so the estimate is compared with real usage.9
A workable sizing rule:
- Take the lowest weekly spend in the 90 days after removing any one-off event. That is the floor; commit up to it with little regret.
- Take the median weekly spend. Between floor and median is the contested band: commit only if the owner of the project behind it signs the demand (cut 4).
- Everything above the median is burst. Leave it uncommitted, or cover it with capacity terms that can flex.
- Subtract workloads with a retirement date or a planned migration inside the term (the exposure table).
- Add known launches explicitly, with an owner, since the trend does not contain them.
For provisioned capacity the same logic applies to load instead of dollars: size to the typical hour and the minute-level peak from cut 3, then run the break-even in Provisioned throughput vs pay-as-you-go. A commit also does not rescue a workload you could still reduce: the Framework warns that commitments and usage optimisation interact, and that waiting for usage to be fully optimised before buying will almost always mean paying more than a balanced approach.1 Do the obvious cuts (response cache for repeated identical requests, a stable prompt prefix) first, re-pull the data, then size.
Package the pack and keep it current
Finance teams want the data in the tools they already use. INFRO exports the request log and usage series as CSV and, through GET /v1/usage/focus, a FOCUS v1.4 Cost and Usage dataset in JSON or CSV with per-project, per-key and per-tag attribution.7 In that dataset the gateway's per-request figure is written to ListCost and ContractedCost, never to BilledCost; BilledCost is filled only from the provider's own reported cost, and only with cost_basis=reconciled.10 Keep that distinction when you hand it over: label the estimate as an estimate.
Share aggregates (cuts 1, 2, 5 and the exposure table), not request rows. Request exports can carry custom tags or user identifiers that may include personal information, so review before sharing.7
After signing, record the commitment so burn-down is tracked against provider-reported cost, with a status of on track, under, over, not started, ended or insufficient data. The contract terms are entered by hand; INFRO does not read them, and it is not a party to the agreement.5 Enter the negotiated rate itself under custom rates, because the commitment's stated discount is only for reference and changes no estimate.5
Pre-meeting checklist
- Reconciliation shows estimated and provider-reported cost within an agreed variance for each provider.
- Unpriced requests are zero, or their share is stated beside each total.
- Direct or cloud-console usage outside the gateway is estimated and listed separately.
- Cuts 1 to 7 are pulled for the same 90-day window and the same time zone (UTC).
- One-off events and cost steps (lost cache, new snapshot) are marked.
- Exposure table has a retirement or step-up date for every model above a materiality line you choose.
- Fallback share and an equivalence test exist for each route you will cite.
- Each project owner has signed the demand in the contested band.
- You have checked what your current agreements allow you to disclose.
