AI spend analysis is the process of mapping your token consumption to named workflows, identifying which cost patterns are structural waste, and producing a ranked list of changes sorted by estimated monthly saving. It is the step between knowing your AI bill and knowing what to do about it. This article covers how the analysis works, how to compare LLM costs across providers, and what the data typically reveals when a team looks at their spend at workflow level for the first time.
The scale of AI spending in 2026
The numbers emerging from enterprise AI adoption are consistently higher than initial forecasts. UBS published research in June 2026 finding that 60% of companies actively managing AI budgets have already moved to cheaper models or open-source alternatives in response to cost pressure. The teams that have not made that shift are encountering $35,000 monthly bills and quota overruns at 200% of plan.
These are not outliers. The pattern is consistent across engineering-led companies running AI products at scale: spend grows faster than the tools available to manage it. The $81,267 single-week case is memorable, but the more common problem is slower and harder to detect: a budget that drifts 20% over plan every quarter with no clear explanation from the invoice data.
The full picture of AI spending varies by company type. For companies building on AI APIs, the provider cost is typically the dominant and fastest-growing component. For companies primarily buying AI software (Copilot seats, Notion AI, writing tools), the spend is distributed across SaaS subscriptions and is better captured by financial monitoring tools. Most mature AI deployments sit in the middle: a mix of direct API spend and AI-augmented SaaS.
Why AI spending is hard to manage
The core problem is a visibility gap, not a budget gap. Most teams know their total monthly AI invoice. Almost none know which workflows are driving it.
AI API providers bill in aggregate. The invoice total from OpenAI, Anthropic, or Google shows you what you spent. It does not show you which product feature, which agent loop, or which engineering pattern produced that spend. Without workflow-level attribution, the only management lever available is a blanket cap, and blanket caps cut revenue-generating workflows alongside wasteful ones.
At FinOps X in June 2026, analysts mapped AI spending into nine distinct cost buckets in a typical deployment. Only one of them, the token invoice, appears on any vendor bill. The other eight, including infrastructure overhead, orchestration cost, prompt processing waste, caching inefficiency, agent loop redundancy, model-task mismatch, sync-on-batch overhead, and uncached context, are real costs that drive up the effective price of every useful output the model produces. None of them are visible in the standard billing dashboard.
This is why companies that try to manage AI spending by watching the invoice total are working with incomplete information. They are optimising the one bucket they can see and leaving the other eight unexamined.
Why AI spending grows faster than expected
Three dynamics compound to produce faster-than-expected growth in AI spending.
The first is usage compounding. AI products that work attract more users. More users means more requests. More requests at frontier model pricing means costs grow non-linearly with adoption. A product that runs 10,000 API calls a month at launch can be running 500,000 within eighteen months without anyone making a deliberate decision to increase AI spending.
The second is the end of the token subsidy. For three years, frontier AI providers priced their models well below cost to drive adoption. BCA Research, writing in the Financial Times in July 2026, described this era as over. Bank of America characterised the corporate response as a shift from tokenmaxxing to token rationing. Usage-based pricing that reflects actual provider costs is now the direction of travel, and companies that have not mapped their spend at workflow level are walking into that pricing environment without the data they need to respond.
The third is that AI spending is uniquely distributed. A single engineer with API access can generate significant spend without any purchase order or approval workflow. The fintech's $81,267 single-week case is an extreme, but unsanctioned spend accumulating below finance's visibility threshold is common. Tokenminimizing, the corporate response to runaway AI spend, has become a real phenomenon: Uber's cap on engineer AI spend is a documented example of what happens when the spend finally surfaces.
What informed AI spending management looks like
The companies navigating AI spending well are not applying blanket cuts. They are running a diagnostic first: mapping spend at workflow level, identifying which of the nine cost buckets the waste sits in, and making targeted changes with an estimated saving attached to each one.
The diagnostic approach produces a ranked list. Not "reduce AI spending by 20%" but "restructuring the system prompt on the customer support workflow saves an estimated $3,200 per month; routing the classification pipeline to a mid-tier model saves an estimated $1,800 per month; implementing prompt caching on the document processing workflow saves an estimated $5,400 per month." Engineering teams can then sequence those changes by return, starting with the highest-value items and working down.
This is the difference between informed rationing and arbitrary rationing. Arbitrary rationing means cutting a percentage off the AI budget and telling engineering to make it work, then discovering three months later that you cut the wrong things. Informed rationing means knowing exactly which workflows account for what share of AI spending, and which changes will reduce it without degrading the output that generates revenue.
The tools available for managing AI spending
Several categories of tool have emerged to address the AI spending management problem, and they operate at different levels.
Hyperscaler FinOps tools, including the AWS FinOps Agent and Google's FinOps Explainability Agent, provide visibility into AI spending within their respective cloud boundaries. Both are free. Both stop completely at their own perimeter: neither sees direct OpenAI spend, Anthropic API usage, or any aggregation layer. For most companies, where AI spending is distributed across multiple providers, these tools show a structurally incomplete picture.
Financial monitoring tools like Ramp AI Spend Intelligence work at the subscription and transaction level. They can tell you what your company is spending on AI tools as a category. They cannot tell you which API calls within those tools are efficient and which are not, because they operate at the payment layer, not the token layer.
Workflow-level diagnostics like TokenomicsIQ operate at the API layer. They analyse the token usage data from your provider export, attribute spend to named workflows and models, identify structural inefficiencies across all nine cost buckets, and rank recommendations by estimated monthly saving. This is the layer where the engineering decisions that drive AI spending actually live.
LLM cost comparison: what each model tier actually costs
A structured AI spend analysis requires an understanding of the pricing landscape across providers. The table below shows approximate pricing for the main frontier and mid-tier models as of mid-2026. All figures are per million tokens. The ratio between input and output pricing matters: most workloads generate 3 to 5 times more output tokens than input tokens, so output price has a disproportionate impact on total spend.
Note: model pricing changes frequently. Figures are approximate and current as of August 2026. Verify current pricing at each provider's billing page before making routing decisions.
| Model | Provider | Tier | Input ($/M tokens) | Output ($/M tokens) |
|---|---|---|---|---|
| GPT-4o | OpenAI | Frontier | $2.50 | $10.00 |
| Claude Sonnet | Anthropic | Frontier | $3.00 | $15.00 |
| Gemini 1.5 Pro | Frontier | $1.25 | $5.00 | |
| GPT-4o mini | OpenAI | Mid-tier | $0.15 | $0.60 |
| Claude Haiku | Anthropic | Mid-tier | $0.25 | $1.25 |
| Gemini 1.5 Flash | Fast/cheap | $0.075 | $0.30 |
The practical implication for AI spend analysis is that model-task mismatch is often the highest-value finding. A classification workflow running on Claude Sonnet at $15.00 per million output tokens, where Claude Haiku at $1.25 would perform equally well, is a 12x cost difference per token. Across a workflow generating one million output tokens per month, that is a $13,750 monthly saving from a single routing change. A structured AI tokenomics analysis identifies exactly these patterns, quantified at workflow level, so engineering teams can prioritise changes by return.
The table above covers direct API pricing. Companies using a routing layer like OpenRouter will see slightly different effective prices due to markup, but gain the ability to route automatically between providers based on cost and performance. OpenAI token pricing is covered in more detail in a separate guide.