Want to appear here? Talk with us
AI ECONOMICS
Who Ordered All These Tokens? Giving AI Spend a Name in FOCUS
Every FinOps team knows this problem. Your CFO asks how much you spend on Copilot versus your own AI agents. You open the cost report and find the answer buried under vague labels like "Other" or "SaaS."
Two data sources should help: FOCUS, the standard cost report format, and the Azure price sheet, which lists every billing meter. The problem is neither one clearly labels what is an input token, output token, or which agent used it.
Partner models like Claude get mixed into generic "SaaS" categories. Microsoft's own Copilot and agent products scatter across different service names. None of it lines up for clean reporting.
The fix is simple but takes work.
Build your own reference list, called a meter map.
Classify each billing meter once, by hand or with rules you control.
Then join that map to your cost data.
This turns confusing rows into clear categories like Provider, Token Type, or Agent Name. Keep an "Unclassified" bucket so new charges do not slip through unnoticed.
Next time your CFO asks about AI spend, you will have a real answer instead of a shrug.
AI PROVIDERS
Google adds Gemini seat caps and usage dashboards; Claude Code trims token costs
Gemini Enterprise now has subscription seat caps, with 25 seats for self-serve accounts and 1,000 for invoiced accounts, helping teams plan costs.
New usage dashboards track Gemini developer adoption and token use, giving cost teams more visibility into spending.
Gemini 3.7 Flash is now available in more regions, giving teams another model option to balance cost and performance.
Gemini 3.6 Flash is now generally available in US and EU multi-regions, adding a lower-cost option for high-volume workloads.
Anthropic
Claude Code now attributes gateway spend to individual users, making it easier to track who is using API budget.
Prompt caching is restored for gateway-routed Claude Code sessions, cutting repeated token costs for teams using custom gateways.
A built-in Claude Code skill now uses about 25,000 tokens instead of 200,000, lowering costs for repeated sessions.
Claude Code now reuses cached context across parallel agents, reducing duplicate token processing in multi-agent workflows.
OpenAI
Codex chats now show estimated lifetime credit usage, helping Enterprise teams track spend at the individual chat level.
WEBINAR
FINOPS FOR AI
Join this FinOps for AI webinar to learn how to control AI costs, improve visibility, implement governance, and justify AI investments with confidence.
📅 August 27, 2026
🕚 6:00 PM Spain / 12:00 PM ET
VIDEOS & PODCASTS
Why Tokenomics is Wrong
I break down why terms like "Tokenomics" create confusion with the blockchain world, the right word to call it and why managing AI costs (AI Economics) needs to be a standalone practice separate from FinOps.
AI ECONOMICS
What an Agent Actually Costs in Production
A finance leader once asked her team why the AI agent bill did not match the vendor's price sheet. Turns out the sticker price for AI models only tells part of the story.
A new breakdown of true agent costs shows something surprising.
Model inference, the number most teams quote, is only about one fifth of the real bill.
The two biggest costs are human review and approval time, at 43% and the governance and control systems built to keep the agent safe, at 25%.
Neither shows up on a pricing calculator. Some more findings worth knowing.
The real cost per verified outcome came out 17 times higher than the simple per-call estimate.
Cheaper models are not always cheaper overall.
They fail more often, need more fixing, and get verified less, so total cost per successful outcome can actually go up.
Scaling volume only helps up to a point, then costs flatten out.
The biggest savings come from improving verification and approval processes, not from shopping for a cheaper model.
Track cost per verified outcome, not per call. Put human review time in the same cost table as token spend. The teams that do this will make smarter bets than the ones still staring at a vendor price page.
AWS
Cost-Optimized GenAI on AWS
GenAI costs on AWS are messy because they are built on tokens, not simple compute resources. Here’s the framework to follow:
Tag first, optimize second.
If you cut costs before you can track where they come from, you end up with a smaller bill nobody can still explain to Finance.Start with the easy win.
Prompt caching cut inference costs by 29 percent with just a config change, no big infrastructure rebuild needed.Do not trust one alert system alone.
A real incident showed a 30,000 dollar spike that anomaly detection missed completely, because it does not watch AWS Marketplace billing.
Budgets with fixed limits caught what smart detection missed.Show costs before you charge for them.
Billing teams for shared cloud resources before your tagging is accurate just creates disputes, not accountability.
Good FinOps for AI is not about picking the cheapest model.
It is about building trust in your numbers, one tagging decision at a time, before you ever send an invoice to another team.
FINANCE
Designing AI Workflows Like a P&L: Cost, Tokens, and RoAI
How to break down AI spending into simple numbers you can actually manage using workflow P&L.
It means tracking cost and value for one AI task at a time. The example used:
A support team handles 10,000 tickets a week with AI help.
The AI costs about 4.50 dollars a week in tokens.
That same AI frees up 500 hours of human time, worth 12,500 dollars a week.
That works out to a huge return, almost 2,777 times the AI spend.
Follow the 4 step method: Pick a workflow, track its AI cost, measure its value, then optimize with FinOps tools like model tiering, token budgets, and usage tagging.
It also connects this thinking to token pricing for teams working in web3 or usage-based billing.
If you cannot see the cost per request and value per outcome, you are just guessing. Start with one workflow, run the math, and let real numbers guide what to scale, fix, or cut.
🎖️MENTION OF HONOUR
Semantic Cache in AI Economics
Many companies build a cache to save money on AI costs. But most caches only work if someone types the exact same question, word for word.
In real life, people never do that. They ask the same thing a dozen different ways, like "reset my password" or "forgot my login."
An exact match cache treats all of these as brand new questions. That means you keep paying for the same model call over and over.
Semantic caching fixes this by checking meaning instead of exact words.
It turns each question into a set of numbers that capture what the question means. Then it compares new questions to past ones using a similarity score. If the score is close enough, it reuses the old answer instead of calling the model again.
This matters most for support bots, FAQ tools, and help desks. These tools often get the same small set of questions asked in many different ways.
Semantic caching can cut costs by treating those variations as one question, not many. Picking the right similarity threshold is tricky though. Too loose, and you risk giving a wrong answer that looks right. Too strict, and you barely save anything at all.
Teams also need to expire old answers so pricing or policy changes do not get served as stale truth.
If your AI bill still feels high even with a cache in place, the problem is not the idea of caching. It is the way your cache decides what counts as the same question.
Save 20% on AI Value & FinOps Certifications
The job market is hungry for certified professionals who can prove results. Don't let your company's budget leak because of a lack of specialization.
Use code: FINOPSWEEKLY_20 to get an instant 20% discount on the most prestigious certification bundles:
FinOps for AI
FinOps Certified Practitioner
FinOps Certified Engineer
FinOps Certified FOCUS Analyst
