FinOps for AI: A Playbook for Managing Your AI Spend

AI adoption has outpaced almost every other line item in the technology budget, and for a lot of organizations, spend on AI models is now growing faster than anyone can track it. Unlike traditional cloud costs, which most finance and engineering teams have spent a decade learning to manage, AI spend comes with its own vocabulary, its own pricing logic, and its own blind spots. That's exactly the gap FinOps is meant to close — and it's worth applying that same discipline to AI before the bill becomes a surprise.

What FinOps actually means

The FinOps Foundation defines FinOps as "an operational framework and cultural practice which maximizes the business value of technology, enables timely data-driven decision making, and creates financial accountability through collaboration between engineering, finance, and business teams." In practice, that means optimizing outcomes based on the technology available, the people using it, and what's actually being spent — with engineering, finance, product, and IT asset management all at the table. AI spend deserves the same treatment, but most organizations haven't built the muscle yet.

A quick primer on "AI speak"

Before you can manage AI costs, it helps to know what you're actually paying for. AI models generally fall into three categories:

• Closed-weight (frontier) models — Anthropic Claude, OpenAI's ChatGPT, Microsoft Copilot, Google Gemini. These are the highest-performing models, hosted by the vendor, and the most expensive.
• Open-weight models — DeepSeek, Alibaba's Qwen, Meta's Llama, Google's Gemma, Microsoft's Phi, Mistral AI. The model's parameters are released, but that's not the same as open source. Performance is typically a step behind frontier models, but cost is dramatically lower, and they can be self-hosted or run by third parties.
• Open source models — some releases from DeepSeek, Z.ai, Alibaba, and OpenAI's gpt-oss-120b go further, releasing weights, architecture, and training code under licenses like Apache 2.0 or MIT, so they can be freely used, modified, and distributed.

Fewer models are fully open source today, but that list is growing.

Then there's the unit you're actually billed on: tokens. Input tokens are the prompts coming from users, apps, or agents; output tokens are what the model generates in response. Everything about AI pricing comes back to tokens.

How AI is actually priced — and where the real cost hides

There are two fundamentally different pricing models to watch. Per-user plans cover people typing prompts, sending images, or using voice, and they typically come with request limits — this is not where the spend problem lives, because a human being can only send so many requests. The real cost driver is API calls made by agents and applications, priced per input and output token. Agents and automated workflows can generate far more requests than any person could, and that's where AI budgets quietly spiral. If you're only watching your per-seat AI licenses, you're watching the wrong number.

What you should be measuring

The metric that matters is results per dollar spent — for every model, application, agent, and deployment option (SaaS or self-hosted) you're using. Since "results" and "spend" aren't always easy to attribute cleanly in real time, a useful proxy is:

Performance = Benchmark results ÷ Token price

To get the "results" side of that equation, look to composite benchmarks like the Artificial Analysis Intelligence Index (which scores math, science, coding, and reasoning) or the Epoch Capabilities Index (a composite of roughly 50 AI benchmarks with open documentation), alongside industry-specific benchmarks for legal, healthcare, finance, customer service, or software development use cases. On the cost side, published token prices vary enormously across vendors and model tiers — often by an order of magnitude for comparable capability — which is exactly why a side-by-side comparison, not a single vendor's list price, needs to anchor your decisions. It's also worth noting that self-hosting an open-weight model only pays off at very heavy usage volumes, and that calculation still needs to account for the IT overhead of standing up and maintaining the infrastructure.

The open-vs-closed debate isn't settled — but the economics are getting hard to ignore

A widely cited Linux Foundation analysis (Frank Nagle, November 2025) makes the disconnect stark: closed models account for roughly 80% of usage and 96% of revenue, even though they cost, on average, six times more than competing open models — while open models routinely achieve 90% or more of the performance of closed models on widely used benchmarks. The same analysis estimates the global AI economy could save $20 to $48 billion a year by shifting more workloads to open models. MIT's Irving Wladawsky-Berger has framed the broader arc of proprietary-versus-open adoption as a progression — experiment, learn, specialize, scale — and by most measures, we're still in the early stages of that curve. None of this means closed, frontier models are the wrong choice; for many high-stakes use cases, the performance gap is worth paying for. But it does mean every high-volume application deserves an honest look at whether an open-weight or open-source alternative could do the job at a fraction of the cost.

Building your AI spend playbook

Putting all of this together, here's where to start:

• Focus first on your highest-cost applications and agents — this is almost always API/token usage, not per-user seats.
• For each one, ask whether it could run on an open-weight, open-source, or "mini" model instead — and evaluate US, French, and Chinese options, plus SaaS versus self-hosted, on their merits.
• Check license terms before you commit. Some open-weight models carry display requirements or usage limits at scale; open source releases under Apache 2.0 or MIT are generally the cleanest to work with, but restrictions on any tier could tighten in the future.
• Get accurate, real-time cost and usage data, attributed down to the specific agent or application generating it — averages and monthly invoices aren't granular enough to manage this.
• Take advantage of caching and batch-processing discounts wherever your provider offers them.
• Revisit the playbook continuously. Model pricing, performance, and licensing in this space are all moving quickly enough that a plan built six months ago is already out of date.

The bottom line

AI spend is following the same trajectory cloud spend did a decade ago: fast growth, unclear attribution, and a widening gap between what's being spent and what's being measured. The organizations that get ahead of it will be the ones that treat AI cost the way FinOps treats cloud cost — as a shared responsibility between engineering, finance, and the business, backed by real data rather than assumptions.

If you'd like the underlying pricing and benchmark data referenced here — there's a lot of it — reach out and we're happy to share the detailed spreadsheets.

Belarc, Inc. helps organizations get accurate, real-time visibility into technology spend and usage. Questions? Contact us via our contact page.