If you cannot say which team spent the AI budget, no FinOps platform will tell you — it can only display tags you have not written yet. The fastest route to a defensible number is free: put every model call behind a gateway like LiteLLM with one key per team, turn on the attribution the providers already ship (OpenAI projects, Anthropic workspaces, Bedrock inference profiles), and reconcile the two against the provider's cost API. That gets you an accurate per-team number in days, and it is the only layer that can actually refuse a request once a budget is gone. Buy a platform second — Vantage if you want published pricing and one pane across cloud and AI, OpenCost for self-hosted GPUs — and only after the tags exist.
| Option | What it attributes | Can it cap spend? | Price (checked 2026-09-28) | Pick it if | Skip it if |
|---|---|---|---|---|---|
| Gateway + provider-native tags (LiteLLM, OpenAI projects, Anthropic workspaces, Bedrock AIPs) | Tokens per key, team, user, tag | Yes — key, team, user and tag budgets block requests | LiteLLM open source: $0; Enterprise: contact sales | You have no per-team AI number today | Every call already carries a team tag you trust |
| Vantage | Cloud + OpenAI/Anthropic + Kubernetes in one report | No — budget alerts only | Free to $2,500 tracked; $30/mo to $7,500; $200/mo to $20,000; Enterprise custom | You want the unified view at a price you can see | Your tracked spend is in the millions and you need shared-cost allocation rules |
| Finout | Virtual tags across cloud, AI and SaaS bills | No — alerting and anomaly detection | Contact sales; flat fee by committed-spend tier | Multi-cloud estate, messy shared costs, no appetite to re-tag infra | You only need AI attribution |
| CloudZero | Unit cost across cloud and AI providers | No — budgets, forecasts, anomalies | Contact sales only | You already run unit economics and want AI in the same model | You are buying it to create attribution |
| OpenCost | GPU/CPU/memory per namespace and pod | No | Free, Apache 2.0 | You run inference on your own GPUs in Kubernetes | Your AI spend is all API tokens |
The loser for this buyer is CloudZero — not because it is a weak product, but because it is the costliest one to evaluate (no published price, no tiers), its recent Claude Enterprise totals can shift for up to 30 days while Anthropic reconciles billing, and it adds nothing to attribution that your tags did not already carry.
Why Can't FinOps Say Which Team Spent the AI Budget?
Because AI spend arrives on the bill with no owner attached. A cloud VM carries the tags someone put on it; a token bill carries a model name, a day and — at best — an API key or project ID. If three teams share one key, no downstream tool can split that line.
This is now the central FinOps problem, not an edge case. The State of FinOps 2026 survey of 1,192 practitioners found 98% now manage AI spend, up from 63% in 2025, and ranked cost allocation — assigning AI costs to business units — among the top three obstacles, behind visibility and ahead of proving value. AI cost management was the number-one skill teams said they needed.
The instinct is to buy a dashboard. The dashboard is not the missing piece. The missing piece is a key-to-team mapping that exists before the call is made.
To keep this comparison honest, every option below is judged against one workload: a 400-engineer company spending about $150,000 a month on cloud, about $40,000 a month across OpenAI and Anthropic APIs, and running one Kubernetes cluster with 16 GPUs shared by three product teams. The question for each option is the one in the buyer's brief — can it tell finance which team spent the GPU and token budget, how soon, and can it stop an overrun.
What Does the Gateway Already Give You for Free?
A gateway gives you per-team token spend in real time and the only enforceable budget cap in this comparison, at no licence cost. An LLM gateway is a proxy that every model call passes through; because it sees the request before the provider does, it can attribute it and refuse it.
LiteLLM's open-source proxy logs every request to a LiteLLM_SpendLogs table with dollar spend and prompt/completion tokens, attributable to the API key, the internal user, the team that owns the key and any tags on the request. Budgets are set with max_budget and reset on a budget_duration such as 30d; once a key or user crosses it, the proxy returns an ExceededTokenBudget error instead of forwarding the call. Tag budgets work the same way — the docs show a request rejected with "Budget has been exceeded! Tag=engineering" — and they are in the open-source proxy, needing only Postgres. The enterprise-only budget features are narrower: per-model budgets on a single key or user. LiteLLM Enterprise adds SSO, SCIM, audit trails and support, and is priced by sales conversation.
The providers ship the other half. Anthropic's Usage and Cost Admin API groups usage by API key, workspace, model and service tier at 1-minute, 1-hour or 1-day buckets, with data typically available within five minutes; the cost endpoint is daily and groups by workspace. OpenAI's Costs endpoint returns daily buckets (1d is the only supported width) grouped by project or line item. On AWS, Bedrock application inference profiles carry cost allocation tags into Cost Explorer and CUR — but only for InvokeModel and Converse, one profile per model per team, up to 24 hours to appear, and not retroactive.
Who should not rely on the gateway alone: anyone who treats its numbers as the invoice. The gateway computes cost from its own price map, and that map drifts. An open LiteLLM issue from 18 September 2026 reports roughly $250 recorded against roughly $2,500 actually billed for one provider (Nebius) — a 10x undercount, though the reporter attributes part of the gap to delays in Nebius's own cost tracking, and a fix PR is linked. A 2025 report found one of nine provider-logged requests missing from spend logs; it was closed as not planned. LiteLLM's own cost-discrepancy guide names the three causes — token categories ingested wrongly, cost formulas that miss billed dimensions, and stale prices — and tells you to compare cache reads and writes category by category. The gateway is the attribution layer. The provider cost API is the truth layer. You need both, and the monthly job is reconciling them.
Is Vantage Worth It for AI Cost Attribution?
Vantage is the right second purchase for most teams because it is the only platform here with a published price, and it has native OpenAI and Anthropic connectors. It puts token spend in the same report as cloud and Kubernetes cost, which is the "one view" finance keeps asking for.
Per its pricing page, checked 2026-09-28: Starter is free up to $2,500 of tracked spend; Pro is $30/month up to $7,500; Business is $200/month up to $20,000; Enterprise is custom with unlimited tracked spend. Virtual tagging starts at Pro, and the page lists LLM token allocation and Kubernetes cost reporting. Its Anthropic connector pulls API cost by workspace, API key, model and token type, plus Claude.ai usage by user and product — but it needs an organisation Admin API key, refreshes once daily, and excludes tax because Anthropic's APIs do not expose it.
At our reference workload, $190,000 a month of tracked spend is far past the $20,000 Business ceiling, so you are on a custom Enterprise quote like everyone else. The published tiers are still useful: you can prove the model on a subset of accounts before any procurement.
Who should not pick Vantage: a team expecting it to stop spend. Vantage budgets are alerts — thresholds on actual (not forecast) cost, sent after reports refresh, which happens at least once a day. By the time an alert fires, a runaway agent has had most of a day.
Finout vs CloudZero: Which Allocates Shared AI Costs Better?
Finout is the better fit when the problem is untaggable shared cost across many bills; CloudZero fits teams that already run unit economics. Neither should be your first move on AI attribution, and both hide their prices.
Finout sells flat-fee tiers based on committed spend, with no usage overage. Every tier includes virtual tags, shared-cost allocation, anomaly detection and its "MegaBill", and the integration list names OpenAI, Anthropic, Kubernetes, Databricks, Snowflake and Cursor. Business and Pro tiers charge $250 and $500 per additional integration respectively; Enterprise includes them. The actual fee is a quote. Virtual tags — rules that assign cost to an owner after the fact, without touching the resource — are Finout's real strength: they can split a shared cluster or a shared data platform. They cannot split a shared API key whose bill line carries no distinguishing field.
CloudZero sells a single subscription with unlimited cost sources, users and dimensions, and publishes no numbers — "request a custom price quote" is the only path. Its Anthropic connection breaks cost down by project, model and operation, and Claude Enterprise by user; costs appear within 24 hours, and Enterprise totals can be revised for up to 30 days. Anthropic lists both CloudZero and Vantage as partner integrations, so ingestion quality is not the differentiator.
Who should not pick Finout: a company whose only unallocated spend is AI tokens. Tagging at the gateway is cheaper than virtual-tagging around its absence. Who should not pick CloudZero: anyone buying it to create attribution. It models unit cost beautifully once ownership exists; with a shared key and no tags, it will model one very large unit.
How Do You Put GPU Cost in the Same View as Cloud Cost?
For GPUs you run yourself, use OpenCost to allocate cost to namespaces and pods, then feed that into whichever platform holds your cloud bill. A token bill has an owner field; a GPU node has only a hostname.
OpenCost is Apache 2.0, originally built by Kubecost, and allocates CPU, GPU, memory and storage to Kubernetes workloads; it also now tracks inference cost per million tokens for vLLM-based deployments. It moved to CNCF Incubation in October 2024. The commercial Kubecost product is now IBM's and its enterprise pricing is not published.
The trap is GPU sharing. A September 2026 write-up on multi-tenant GPUs puts it precisely: time-sliced replicas "look like N GPUs in allocation data while being one GPU in utilization data." Invoice on allocation — it is deterministic from Kubernetes state — and audit on DCGM utilisation, or one team pays for a card it shared with three neighbours. In our reference cluster, 16 GPUs time-sliced four ways appear as 64 in allocation data.
Who should not pick OpenCost: a team whose AI spend is all API tokens. It has nothing to allocate. If you rent GPU capacity rather than run it, see our breakdown of what a GPU hour actually costs.
Showback or Budget Cap: Which Do You Actually Need?
You need both, and only the gateway can provide the cap. Showback means reporting each team's spend back to it; chargeback means billing it; a cap means refusing the request. Every FinOps platform in this comparison does the first two. None sits in the request path, so none can do the third.
That matters more for AI than for cloud. A misconfigured VM costs you an hourly rate; a looping agent can burn tokens by the hundred million in the gap between a daily refresh and a human reading an alert. Set team budgets in the gateway at roughly 120% of forecast — high enough not to trip on a normal month, low enough to stop a runaway — and use the platform's alerts for the 80% early warning. We made the same argument about managed gateways in AI gateway vs API management.
What Predicts Regret in an AI FinOps Purchase?
The purchases teams regret are the ones made before the key-to-team mapping existed. Four criteria predict it better than any feature list:
- Time to first accurate number. Gateway: real time, but an estimate. Anthropic API: about five minutes. OpenAI costs: daily. Bedrock tags: up to 24 hours, forward-only. Vantage: daily. CloudZero: 24 hours, revised for up to 30 days. If your first reconciled number is more than two weeks away, you bought the wrong layer first.
- Share of spend on shared keys. Above about 20%, no platform will help until you split the keys. Measure it with the provider usage API grouped by key.
- Where the cap lives. If the answer is "an alert email", you have showback, not control.
- Tracked spend versus published tiers. Under $20,000, Vantage's list price is the whole answer; above it, every option here is a negotiation.
What changes the answer: if most of your AI spend is seat-based (Copilot, ChatGPT Enterprise, Claude Enterprise) rather than API tokens, per-user analytics from the vendor matter more than the gateway — see what a Copilot seat actually costs. If you buy through a reseller, check what that reseller can see — DoiT's purchase of Attribute raised exactly that neutrality question.
What to Do in the Next 30 Days
This Week:
- Pull last month's cost from the OpenAI and Anthropic cost APIs grouped by project and workspace. Write down the share that lands on a key used by more than one team.
- Stand up LiteLLM (or your existing gateway — see our gateway comparison) and issue one virtual key per team, with a team budget and a
budget_durationof30d.
This Month: 3. Revoke shared provider keys once every team is on its own gateway key. Tag Bedrock traffic with application inference profiles, and activate the tags now — they will not backfill. 4. Reconcile gateway spend against provider cost by token category. Log any gap above 5% and fix the price map before anyone quotes a number to finance.
Before Next Budget Cycle: 5. Only now evaluate Vantage, Finout or CloudZero, with real tags flowing, and deploy OpenCost if you run your own GPUs. Judge each on how fast it shows the reconciled number, not on its demo.
The Bottom Line
Cloud FinOps took a decade to learn that the tag policy is the product and the dashboard is the display. AI spend is re-running that lesson on a much faster clock, because a token bill has even less metadata than a VM ever did and an agent can spend a quarter's forecast in a weekend.
Tag the call, cap the team, reconcile the invoice. Then buy the screen.
Continue Reading
- Best LLM Gateways for Cost Control: Self-Host First
- AI Gateway vs API Management: No Managed Gateway Caps Your Spend
- DoiT Bought Attribute. It Also Sends Your Cloud Bill.
- GPU Hour Pricing: A $3.44 Rate Cost $68.80 to Use
- A 40-Call Audit Flagged 7 of 15 GPT Endpoints for Token Inflation
- AI Observability Pricing: Same 10 GB, $49 or $930