TOGETHER WITH STACKLET
The Virtual Token Governance Summit You’ve Been Waiting For (Oct 27th)
AI token usage is growing fast across developer tools, internal agents, and production apps. Most teams can see some of that spend. Few can steer it toward the work that pays off.
Join FinOps leaders, platform teams, and AI builders for 90 minutes on governing AI token usage. Learn where token spend really goes, how to route work to the right model, and how to set budgets that don’t stop developers mid-task.
We’ll close with a first look at Stacklet Token Custodian, the control plane that attributes every AI token and governs it with policy.
Tuesday, Oct 27 | 12:00 to 1:30pm EDT | Free livestream
AI COST OPTIMIZATION
Your AI agent may cost more because it keeps changing what it sends, not because the model costs too much.
Asana reports cutting a browser agent’s cost per task by 76 times and making it five times faster by changing how it reused past input and switching models.
Using the original model, the best setup still cut costs by 29 times. The problem was repeated edits to the agent’s history.
Each edit broke cache reuse, which lets a model read saved input at a lower price. Asana kept that history unchanged between larger cleanup steps and gave it more room.
The next steps:
Check cache use with the provider’s token counts, not just total spend.
Test history limits for each model.
Set caps on steps, tokens, and cost per task.
Keep full test records so teams can check results.
The study used only three or four runs per setup, so small differences need more testing. The best runs found all 192 facts in the task.
Before choosing a cheaper model, check whether your agent keeps paying to read the same input.
AI PROVIDERS
Claude Cache Reads Drop to $0.10 per Million Tokens; Asana Cuts Agent Costs 76x
Anthropic
Claude Code cut Sonnet 5.5 cache read prices to $0.10 per million tokens, down from $0.20. Cost tools and budget limits now use the new rate.
Idle compaction in Claude Code shrinks long chats before the one-hour cache runs out, so you waste fewer tokens.
Claude Code now shows cache creation token metrics, and it gives better /usage warnings. It's easier to see what drives spend.
Claude Opus 5.5 and Sonnet 5.5 are now in the Gemini Enterprise app. You pay as you go, so you can compare model costs across regions.
OpenAI
Asana made its browser agent 76x cheaper and 5x faster using Codex. It shows that tuning can cut costs a lot.
Composer predictions in Codex use no credits during beta, so teams can try them for free.
VIDEOS & PODCASTS
How to Control AI Costs, Token Spend and ROI
We discuss the similarities between AI adoption and the cloud era, tokenomics, AI gateways, model routing, caching, hybrid AI architectures, local models, frontier models, data sovereignty and the importance of measuring business outcomes rather than simply counting tokens.
AI GOVERNANCE
Per-User AI Spend Limits in Snowflake That Don’t Lock Out the People Who Run Them
One heavy AI user can spend in an afternoon what another spends in a month.
Snowflake’s per-user AI quotas help FinOps teams set daily, weekly, and monthly credit limits without building their own blocking tools.
Jeremy Demlow’s tested setup offers three useful lessons:
Set limits from real use. Give new and untagged users a default limit, with separate tiers for heavy users and service accounts.
Daily limits curb short-term spend; monthly limits protect the budget.Roll out controls in steps. Apply tags first, start with blocking off, and exclude the admin by name. Tags can take about two hours to update.
Wait at least 10 minutes after quota creation or scope changes, then check who is covered before turning blocking on.Use small top-ups for short-term needs. Moving from a 250-credit monthly limit to 350 gives someone 100 more credits, not a fresh 350.
Remove them from the old quota too, and set a process to take back the top-up.
These quotas are not exact spending caps: blocks can take minutes, allowing extra spend. The scripts are a community example, so test outside production first.
Good AI cost control pairs clear budgets with safe rollout steps and a clear path for extra spend.
AI COST OBSERVABILITY
Five Ways Claude Prompt Caching Quietly Breaks On Amazon Bedrock
Your Claude requests can succeed while your Amazon Bedrock bill climbs. One reported LiteLLM upgrade cut cache hits from about 90% to 25–45%, and daily costs rose 2–3 times for six days.
Prompt caching lets Bedrock reuse shared input at about one-tenth of the normal input price. But failed reuse can mean paying full input price plus a 25% cache-write premium.
Five checks for teams running AI agents:
Move the cache point as tool results build up, rather than leaving it at the first user message.
Keep an earlier cache point after large groups of tool calls.
Put changing text, such as dates and request IDs, after shared content.
Confirm that software updates and model changes still send cache points.
Check that saved streamed replies do not become hundreds of text blocks.
Track CloudWatch’s CacheReadInputTokenCount and CacheWriteInputTokenCount by model.
Many writes with few reads can signal wasted spend; both near zero can signal missing cache points.
Treat cache health as a cost check after every software or model change, not just when the bill jumps.
AI UNIT ECONOMICS
Agentic FinOps, Part 5 of 5: I Built Outcome Economics Studio — Because Model Cost Is Not Outcome Cost
A lower AI bill does not always mean a cheaper business result. A low-cost model may need more retries or human checks before the work is done right.
Kartik Anand’s Outcome Economics Studio compares the full cost of an AI task, not just its tokens. It adds model, tool, human review, testing, and shared system costs, then measures cost per correct result that meets policy rules.
Three checks stand out:
Track both average cost and the highest cost allowed per task.
Skip extra model calls when they cannot change the final action.
Keep estimates separate from measured results.
Routing based on the final action cut model cost by about 73% versus using only a larger model. That is not a claim about total business savings.
The Studio supports planning and comparison; it does not control live spending or check invoices. Try it with your own task volumes, review times, and quality data.
The goal is the lowest full cost for a correct, policy-safe result, not the lowest token price.
🎖️ MENTION OF HONOUR
Agent FinOps in the Microsoft Ecosystem
Your Microsoft AI bill can span three admin centers, but the teams using it still share costs.
This guide explains where to track agent spend and which controls can stop it.
Track cost per result, such as one invoice cleared, rather than just counting users or chats. One request can trigger several paid steps.
Set monthly limits per user for Copilot and Cowork. For Copilot Studio, set limits per agent and turn on the hard stop.
Treat Azure budgets as warnings, not spending caps. Foundry agents need added controls to keep runaway use in check.
Review the shared credit pool with both Microsoft 365 and Power Platform owners. Credits assigned to one team leave less for others.
Check the monthly bill before charging teams. Usage reports and cost screens may show different totals.
Watch testing costs, too: some agent builds use credits before launch. Give each agent a cost owner, set clear limits, and judge its value by the work it gets done.
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 Certified Practitioner
FinOps AI Value
FinOps Technology Value
FinOps Certified Engineer
FinOps Certified FOCUS Analyst
AI Economics Community
Join the Slack community where the real AI Money conversations happen
