Want to appear here? Talk with us

A new survey called State of Tokenomics asked almost 500 large companies how they track and manage AI spending.

Together these companies represent $4.6 trillion in revenue, so the results carry real weight for FinOps teams.

The biggest problem is proving value. Three out of four companies cannot confidently show their CFO that AI spending leads to real business results.

Even companies that feel good about tracking their AI costs often cannot connect that spending to outcomes their finance team would accept.

Ownership is starting to take shape. Most companies have picked an owner for AI costs, usually someone in technology, but many are still sharing that job across teams.

Tools that route AI requests to the right model are becoming popular, and teams using them are four times more likely to prove value to their CFO.

When asked what they want from AI providers, companies did not ask for lower prices.

They asked for clearer bills and better data, something FinOps professionals know well from cloud cost management.

More than half of companies are already changing their pricing because of AI costs.

Companies do not need cheaper AI, they need to understand what they are paying for.

Read More

Opus 5.5 sets clear token prices for Claude Code: The model charges $4 per million input tokens, $20 per million output tokens, and $0.20 per million cached tokens, making costs easier to estimate.

Auto mode skips billed classifier overhead: In supported setups, Claude Code's new server-side classifier doesn't add extra charges like the local fallback did.

Slack cost totals now match real usage: A bug that showed inflated cost and token numbers after worker restarts was fixed, giving teams more accurate spend tracking.

Gemini Code Assist no longer bundled with some plans: New or renewing Gemini Enterprise Standard and Plus subscriptions won't include Code Assist, so check your renewal costs.

Gemini 3.8 Flash is now the default model: This change may shift usage patterns and costs for teams using Gemini Enterprise across global, US, and EU regions.

Prompt caching for GPT-6 gets cheaper and easier to track: New breakpoints and diagnostics improve cache hit rates, which can lower token costs and speed up responses.

GPT-6 Sol and Luna offer lower token prices: These new models cost less than GPT-5.6 predecessors, giving teams cheaper options for coding and high-volume tasks.

ChatGPT for Word bills by model usage: Business, Enterprise, and Edu plans now pay standard API rates based on the selected model, making costs predictable.

Real-world savings: Ringg cuts costs 90% with GPT-5.6: The company's multilingual support agents saw major savings while resolving up to 65% of calls, showing the real cost impact of model upgrades.

Read All AI Provider Updates

Discover why organizations need to move beyond cost control and focus on AI business value, cost per task, AI observability, governance, and strategic decision-making.

We speak with Andrée Lundberg, FinOps Lead at H&M Group, about AI Economics, AI cost optimization, FinOps, business value, AI governance, and the future of AI spending.

Forecasting AI spend is turning out to be much harder than tracking cloud costs ever was.

Knowing gas costs four dollars a gallon does not tell you the total cost of the trip.

You also need to know the vehicle, the distance, and whether you take a wrong turn.

AI spend works the same way. Token prices are just one piece of a much bigger picture.

Business demand does not directly create cost. It creates AI activity, and that activity can vary wildly for the same request.

One customer question might trigger three steps behind the scenes, or it might trigger thirty.

Architecture choices, like routing simple tasks to cheaper models, can cut token use by ninety percent without changing the amount of work done.

Cheaper AI can actually raise total spend, not lower it, because teams start using it more often.

Instead of one exact forecast number, ranges like a low, expected, and high estimate may be more honest.

The real question is not just what AI cost. It is whether the outcome was worth paying for. Tokens tell part of the story, but value tells the rest.

Read More

Forecasting AI spend feels a lot like planning a road trip. Knowing gas costs $4 a gallon tells you nothing about the total trip cost.

You still need to know the vehicle, the distance, and whether you take a wrong turn.

That's the core insight from a recent FinOps deep dive on AI cost forecasting.

Tokens are measurable and tied to the bill, but they don't tell you if the work was worth it.

A $100 AI task might replace hours of engineering, or it might just repeat low-value work. Same spend, very different economics.

The real forecasting model should follow a chain: business demand drives adoption, adoption drives execution complexity, complexity depends on architecture, and only then do we get token and infrastructure costs.

Time horizon matters too. This month, a simple run-rate estimate may work fine. Next quarter, forecast from actual usage drivers. Next year, use scenarios instead of one falsely precise number.

One striking example: Spotify engineers cut Claude token use by 90% just by routing routine work to a cheaper model.

Same business output, totally different cost baseline. The bigger question is whether the outcome was worth paying for.

AI just adds more layers between demand and dollars to work through.

Read More

Counting AI tokens tells you cost, not value. FinOps teams need to look past token usage and cost, and instead measure real business outcomes.

He shares a simple framework called Observe, Net, Convert, and Book.

First, observe the baseline, like how long a task took before and after AI. Then net out any extra work AI creates, such as time spent reviewing its output.

Next, convert that net time savings into something useful, like hours, headcount, or dollars. Finally, book the value where it actually shows up, whether that is new revenue, saved spend, or freed up capital.

The article also lists ten ways AI can create value, including revenue growth, cost avoidance, faster delivery, better quality, lower risk, and even employee engagement.

A single AI project can touch several of these at once.

For FinOps leaders, this framework offers a clear way to prove AI spend is paying off, instead of just tracking usage numbers.

Real value comes from outcomes, not tokens.

Read More

AI spending has a new name in FinOps circles: Tokenomics.

The FinOps Foundation recently made it an official domain under the Linux Foundation, putting AI token costs on the same level as cloud infrastructure spend.

AWS breaks the challenge into four familiar pillars: See, Save, Run, and Plan. Visibility comes first.

Turning on IAM principal tags in AWS Cost and Usage Report 2.0 lets teams trace Bedrock spend back to specific users, projects, or coding tools like Claude Code.

The optimization numbers are striking.

The same coding task cost 1 cent on Kiro versus 7 cents on Claude Code, a 7x difference before any prompt tuning even starts.

Writing prompts like a computer instead of a person cut token use by 55% in one test. Prompt caching can cut costs by up to 90%.

For governance, AWS Budgets, Cost Anomaly Detection, and Service Control Policies help catch runaway spend before it becomes a surprise bill.

Proving ROI remains tricky. Time saved does not always turn into money saved, since teams often use that free time for other work instead of cutting costs.

Amazon's own Sales Finance team offers a real example. AI cut deal analysis time from 15 minutes to 1 minute, saving about 14,000 hours.

AI costs need the same discipline traditional cloud costs already get, and it starts with seeing where the tokens go.

Read More

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: