What is AI spending and why is it growing so fast?
AI spending covers all costs a company incurs from AI: API costs from providers like OpenAI, Anthropic, and Google, AI SaaS subscriptions, and the infrastructure costs of running AI workloads. It grows faster than expected for three reasons: usage compounds non-linearly as products gain adoption; most teams lack workflow-level visibility so inefficiencies accumulate undetected; and the token subsidy that kept frontier model prices artificially low during 2022-2025 is ending. UBS research from June 2026 found that 60% of companies actively managing AI budgets have already moved to cheaper models in response to cost pressure. Full analysis here.
What is AI token cost?
AI token cost, also called LLM token cost, is the per-token price charged by AI API providers for model usage. Providers bill separately for input tokens (the text you send, including your system prompt and conversation history) and output tokens (the text the model generates). The token invoice is one of nine cost buckets in a typical AI deployment. The other eight do not appear on any vendor bill, which is why teams that manage only what they see on the invoice are usually working with an incomplete picture. Full breakdown here.
What are the nine AI cost buckets?
Identified at FinOps X in June 2026: (1) Token Invoice, the only one on any vendor bill; (2) Infrastructure overhead; (3) Orchestration cost; (4) Prompt processing waste, including system prompt bloat; (5) Caching inefficiency; (6) Agent loop redundancy; (7) Model-task mismatch; (8) Sync-on-batch overhead; (9) Uncached context. A usage export shows five of these directly: prompt processing waste, caching inefficiency, agent loop redundancy, model-task mismatch and uncached context. TokenomicsIQ diagnoses each of those and ranks recommendations by estimated monthly saving. Infrastructure, orchestration and batch overhead sit outside a usage export and are noted only where the data allows.
What file formats do you accept?
OpenRouter CSV exports and TypingMind JSON exports, including Anthropic and OpenAI usage run through TypingMind. OpenRouter data gives the richest analysis: per-request token counts, named application context and caching data. OpenRouter Activity dashboard CSVs contain metadata only (model, tokens, cost, timing); OpenRouter API-generated exports and TypingMind exports include prompt and response text. Anthropic and OpenAI dashboards do not expose request-level usage data. Other gateways, such as Portkey, can be added on request.
What does the report actually contain?
Six sections: a personalised executive summary, spend breakdown by model and named workflow, inefficiency analysis ranked by monthly saving, specific recommendations with engineering implementation notes, risk flags covering vendor concentration and model lifecycle exposure, and a methodology section. Both PDF and HTML formats included.
How long does it take?
Your free assessment arrives within 24 hours of upload. The full diagnostic is delivered within 2 working days of payment.
Is my data stored?
No. Your uploaded data file is deleted within 24 hours of report delivery, and any copy received by email is moved to a secure processing location and removed from our inbox on receipt. Where prompt content is present in your export, up to the first 500 characters of each prompt are processed solely to classify the task type. Prompt and response content is never kept beyond that deletion window, and we do not store, index or analyse client data beyond the report generation process.
What if I do not use OpenRouter?
If your team uses TypingMind, we can assess that export, including Anthropic and OpenAI usage run through it. We support OpenRouter and TypingMind exports today. Other gateways, such as Portkey, can be added on request: tell us in the form and we will confirm timing before you send anything. Direct provider dashboards do not expose enough detail for a reliable assessment.
How is the estimate calculated?
We classify every request in your export by task type, then check for paying for a more expensive model than the task needs, unnecessary input cost in every request, missed prompt caching and repeated or unnecessary model calls. “Recoverable” means spend that would not have been incurred with the cheapest adequate model, a compressed or cached prompt, or no duplicate call, at the provider prices in force on the report date. The annual figure is your average monthly recoverable spend over the most recent months in your export (up to three) × 12, shown with a ±15% range in the free assessment. With one or two months of data the estimate is less certain, so three months or more is recommended. Infrastructure, engineering time and non-LLM costs are excluded. All figures are identified opportunity, not implemented savings.
What is the free AI Spend Assessment?
A one-page assessment, formally called the Token Waste Score, built from your own usage export and returned within 24 hours at no cost: your recoverable spend %, an annualised saving estimate (±15%) and your two highest-value findings by category. The ranked fixes, named workflows and implementation notes are in the full diagnostic. See an example.
How do I know the savings in the full diagnostic are real?
Every saving in the full diagnostic shows its working: the requests affected, your actual tokens and cost, the recommended alternative and its published price. Your team can check every figure against your own export. The full diagnostic is a fixed fee of $3,500, payable when you commission it. Figures are identified opportunity, not implemented savings; the optional $1,250 delta report measures what your changes actually saved.
Can I run a follow-up report after implementing the recommendations?
Yes. The delta report ($1,250) runs the same analysis on your new usage data alongside your original export and shows exactly what has changed: saving confirmed, remaining opportunities, and any new patterns since the baseline. You re-upload both CSVs when you are ready. We never hold your data between sessions.
