Want to appear here? Talk with us

One team found out the hard way how fast AI tooling costs can pile up. Their AI code review tool cost about $1000 a week.

Then they tried running smarter AI review checks through GitHub Actions.

That single experiment cost $1000 in one day. The reviews were also noisy and slow, taking up to 30 minutes per pull request.

The fix was simple and cheap.

Instead of paying an AI model to review every single pull request, the team used AI once to write custom linters based on their own coding rules.

Linters are small programs that check code in seconds, not minutes. Their slowest check now takes 27 seconds. The comments are consistent, so engineers can skim them fast.

Best of all, these checks can run before code is even pushed, so mistakes get caught early instead of costing time and money later.

This is a clear lesson for anyone managing AI spend. Pay for AI to build a smart tool once.

Do not pay AI to repeat the same expensive judgment call thousands of times a week.

Small automation choices like this can quietly save real budget.

Read More

Gemini Enterprise gets better monitoring, adding new tool, engine, and response-code dimensions plus a latency metric to help teams track connector costs and performance.

Claude Code adds a cost-optimize command, guiding teams step-by-step to review and cut their Claude API spend.

Claude Code lets teams set prompt-cache time limits, giving more control over cache costs and cutting wasted reprocessing.

Claude Code fixes cost estimate resets, so usage numbers stay visible when switching between agent views.

New Admin plugin manages usage and limits, giving workspace admins tools to track spend and control access in ChatGPT Work and Codex.

GPT-5.6 arrives in Kiro, offering better price-performance for planning, building, and testing software.

Jalapeño chip boosts inference efficiency, cutting latency and power use to lower inference costs and carbon impact.

Read All AI Provider Updates

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

We’ll deploy Kestra locally, configure a FinOps workflow using a pre-built blueprint, monitor Azure Batch resources and automatically scale idle resources down to reduce unnecessary cloud costs.

AI spending is not like buying software licenses anymore. Tokens, the building blocks of AI costs, work more like an electric meter than a flat fee.

The same task can cost very differently each time it runs, depending on how much context a model uses.

The problem is spending without knowing what you are getting back.

Smart teams are fixing this by building visibility first. They track every dollar, figure out the real cost per interaction, and connect that spend to actual business results.

One company found that a rise in token spend matched a big drop in customer support time. That single insight turned a scary bill into a clear success story.

Teams are also splitting budgets into two parts. One budget runs the daily AI work. The other improves and fixes the system over time.

This keeps maintenance costs from quietly eating the whole budget.

The biggest idea here is that finance leaders do not hate spending.

They hate not knowing if the spending is working.

Once teams can prove the return, leaders feel confident giving more room to spend.

Good visibility does not slow AI work down. It is what makes fast, confident AI spending possible.

Read More

Rippling almost spent 40 percent of its engineering budget on AI tokens. That number should make any FinOps team pay attention.

One engineer alone burned $50,000 a month, and just 10 to 15 percent of employees drove 60 percent of the total spend.

Rippling's CFO and product team built an internal tool to fix this, then turned it into a product called AI Spend Console.

It tracks who spends what, on which AI models, and whether that spend actually leads to good work.

It even flags engineers whose code often needs rework, tying cost directly to output.

This setup routes simple tasks to cheap or local models, and only sends hard problems to expensive ones.

Paired with a tracking tool called Langfuse, teams can see exactly where their money goes..

If you do not track your AI spend by task and person, you are almost certainly paying more than you need to.

Cost control does not require a big budget, just the discipline to route work to the right tool.

Read More

AI cost gateways are the go-to fix for a real problem: cloud bills can be tagged and tracked, but AI token calls often cannot.

Teams route every AI call through one gateway so they can tag, meter, and cap spend by team. That advice is sound, but it skips a big step.

The gateway now holds every provider's login keys for every team using it. That makes it as sensitive as an identity system, not just a cost tool.

The most popular open source gateway had three serious security problems in one year.

- One let attackers reach the database with no login at all.

- Another was serious enough to land on a government list of actively exploited flaws.

- Even worse for budget-watchers, the spending cap feature failed in two opposite ways.

One bug let a five cent limit balloon to fifty four cents before stopping. Another bug blocked spending that was still well under budget.

People running these tools in real jobs report these budget failures again and again, not as rare glitches.

This is what happens when one tool gets trusted to hold every key without the same checks given to identity and security systems.

The fix is not to avoid gateways.

It is to test the budget caps yourself, read the incident history first, and name someone who owns security review, separate from whoever owns the cost dashboard.

Right now, most companies have no one in that role. Cost control tools deserve the same trust and testing as security tools, because for AI spend, they have become one and the same.

Read More

AI agents are showing up everywhere in the enterprise, and most companies still treat them like an on/off switch.

They are either fully locked down or fully trusted. This approach is failing.

Gartner found that 40 percent of enterprises will shut down autonomous AI agents by 2027, not because the tech broke, but because no one set clear limits on what the agent could do.

Regulators in Singapore and China both landed on the same fix, even though they worked separately.

They built tiered systems that sort agent decisions into levels, based on how much authority the agent should have.

A simple version looks like a traffic light. Red means only a human can decide. Yellow means the agent can prepare an action but needs approval first. Green means the agent can act on its own, but only within limits set in advance.

Agents should start at the bottom and earn more trust over time, similar to a new employee proving themselves.

Keeping error rates under 2 percent for 30 days before giving an agent more freedom.

Setting these limits now, before a costly mistake forces the issue, is the real lesson here.

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: