Operations | Monitoring | ITSM | DevOps | Cloud

Picture two teams running the same million input tokens through the same model. One team’s tokens are cache hits, queued through the batch API. The other team’s are fresh, sent live. On a current-generation OpenAI model, cached input runs about a tenth the price of a fresh token, and batch processing cuts whatever’s left in half. Stack the two: at a list rate of $2 per million tokens, one team’s bill comes to 10 cents, the other’s to two dollars.

An LLM gateway is a proxy that sits between your applications and model providers, handling routing, failover, caching, and cost controls through one API. The strongest picks in 2026: LiteLLM for self-hosted control, OpenRouter for instant multi-model access, Portkey for managed governance, and Bifrost for production-scale throughput. Enterprises spent $37 billion on generative AI in 2025, a 3.2x jump in one year, per Menlo Ventures.

model-right-sizer-schema is a Claude Code skill that designs the typed contract between one agent and the controller that dispatches it. Point it at an agent plus its controller and it returns a JSON prescription with typed in/out fields, an exclusion list that keeps raw logs out of the reply, a before/after size delta, then writes the contract into the agent's own file. It picks from nine portable output-shape families, or your repo's own.

Most people rebuild the same handful of Explorer queries: the monthly close view, spend by team for the staff meeting, the filter set that isolates a service you’ve been watching for two months. When we shipped query history and favorites earlier this year, it gave you a way to save up to 12 Explorer configurations.

n8n pricing runs €24 per month for 2,500 workflow executions (Starter), €60 for 10,000 (Pro), and €800 for 40,000 (Business), with 17 percent off on annual billing and custom Enterprise pricing above that. Every plan includes unlimited users and unlimited workflows. The self-hosted Community Edition is free with unlimited executions; you pay only for your server.

CoreWeave, a GPU cloud provider, prices start at $6.16 per GPU hour for an Nvidia H100 and reaches $8.60 for a B200, sold as fixed multi-GPU nodes: an 8x H100 node lists at $49.24 per hour on demand. Spot rates run up to 60 percent below on demand, reserved contracts discount up to 60 percent, and egress is free.

The ServiceNow integration is in Labs now, which means any CloudZero admin can switch it on and start routing cost work into their incident queue the same day. What that gives you is a full loop between the money and the work. A monthly Optimize pass surfaces recommendations with a dollar figure on each one. Pick the ones worth acting on, open incidents for all of them at once, and each ticket shows up carrying the resource, the finding, and the context an engineer needs.

Across a same-store panel of 430 CloudZero customer organizations, AI reached 2.66% of the median company's cloud bill in August 2026, up from 2.61% in July and roughly four times its level a year ago. The 75th percentile crossed 11%. The share of organizations with at least 10% of cloud spend attributed to AI jumped to 28.2% from 23.9%, the largest one-month move that tier has posted. Two-thirds of the panel now spends at least $1,000 a month on AI. The typical bill barely shifted.

GPT-6 Astra is OpenAI's flagship reasoning model, released September 3, 2026. It costs $10 per million input tokens and $50 per million output tokens on the standard API tier, with cached input at $1 and cache writes at $12.50. That is 2.5 times GPT-5.6 Sol's promotional rate and matches Anthropic's Fable 5.1 on both headline numbers. Batch and Flex halve those rates, Fast mode doubles them, and any prompt past 272K input tokens reprices the entire request.

An AI cost calculator for the whole wallet adds four lanes: seats and subscriptions, API and token usage, cloud AI services, and GPU infrastructure. Average 2026 totals run $25 per employee per month at light adoption, $100 to $150 at active adoption, and $300 or more at AI-heavy companies. Getting to your number takes four lane subtotals and three corrections.

Enterprise agent platforms need a client ID and client secret in hand before they will connect to anything. An admin with the Modify MCP Settings permission can now issue that pair directly in Settings, connect the platform, and manage the credential lifecycle on whatever schedule your security policy requires. No support request, no wait.

Pop quiz: If you spend $40,000 per month on Anthropic, and you’ve got two customers, what’s your cost per customer? If you bypassed the easy answer of $20,000 and said, “Scott, you old trickster, that’s not enough information to answer that question,” you’ve won today’s prize: a lesson in the perils of average costs. Let’s flesh out the situation: You put an AI feature in your product, a document assistant powered by Claude.

Changing a Kubernetes resource request takes two numbers: what the workload requests, and what it uses. The CloudZero MCP server now returns both, by cluster, namespace, or workload. This gives you a number you can defend. Usage comes back as P95 over the date range you query, 30 days by default. When an engineering lead asks whether a service runs on a smaller request, that is the figure that settles it. Over-provisioning and under-provisioning show up on the same query.

LLM token cost is the price a provider charges per token a model reads or writes, quoted in dollars per million tokens. Input and output bill at separate rates, with output priced at roughly 5x input. As of September 2026, published rates range from under $0.10 to more than $180 per million tokens on top-end reasoning tiers. In late 2025, Hardik Sonetta of Thomson Reuters Labs published a warning about the most common prompt caching mistake in production.

S3 lifecycle rules move data to Standard-IA or Glacier to cut storage cost. CloudZero now flags the buckets where that move backfires: an early delete fee is charged when an object leaves its tier before the tier’s minimum storage duration. The cause isn’t always a misconfigured lifecycle rule. A manual delete, an overwrite, or an object written straight into the tier by a replication or backup job produce the identical charge.

AI usage tracking means measuring who and what consumes AI across your company, by team, feature, and model, then converting the usage into spend and cost per unit of work. Provider consoles stop at totals per API key. Tracking puts names on those totals: which team, which product, which model, and whether any of it was worth the money. In May 2026, CNBC reported that “almost every Fortune 500 is tracking overall AI usage,” quoting ModelOp CTO Jim Olsen. The same reporting carried his warning.

Repo rightsizing is a single-pass audit of every real model call in a codebase you already shipped: SDK invocations, sub-agent dispatch sites, and agent frontmatter pins. Each call site is scored on the job it actually does, and the result commits as one blueprint file you can diff next quarter. It replaces one-skill-at-a-time reviews, which miss files where a single model key covers two different jobs.