AI FinOps

Claude Enterprise smart reports tie AI spend to workstreams by reading a sample of employee transcripts. Settle name reveals, the personal-chat count and list-price figures before an Owner turns them on.

The Microsoft Copilot harness starts billing sandbox compute on December 1 while it is still a no-SLA preview, with Microsoft setting the approval policy for its five built-in tools. Test it now with Shell off, and compare it with hosted agents and AWS AgentCore before the meter starts.

Snowflake's Cortex Training claims up to 61% lower fine-tuning cost, based on a LoRA SFT benchmark. At $9.60 per H200 hour, your own throughput decides whether you see any of it.

From December, chat on advanced models and Codex will draw on one token allowance on Plus, Pro and Business plans. OpenAI has published no sizes, date or conversion rules, so teams should baseline usage and set a credit budget now.

For one support agent at 300,000 sessions a month, Google's Agent Runtime bills $510 of compute, AgentCore $846 and Foundry prompt agents nothing. Memory, web search and idle timeouts set the rest of the bill.

Agentforce Flex Credits bill $0.10 per action, so a service agent's cost per resolved case runs from $0.50 to $3.33 depending on how many steps it takes. Help Agent's $2 per resolution wins above about 12 actions at 60% resolution, and the $2 per-conversation SKU never wins.

Baseten's A10G costs $1.21 an hour against SageMaker's $1.52, even with no scale-to-zero. Scale-to-zero saves more, but only on gaps longer than the idle window plus startup, and every such gap hands one user a cold start.

NVIDIA AI Enterprise lists at $4,500 per GPU per year. For bare-metal inference, vLLM or SGLang on your own stack, or Red Hat AI for a support contract, replaces it; vGPU and production NIM are where you still need it.

An H100 running gpt-oss-120b at Lambda's $4.29 an hour must sustain 5,958 tokens per second every hour of the month to match a $0.20 blended API rate. Business-hours traffic leaves it at about twice the API's cost.

Microsoft, Google and Salesforce raised base prices to pay for bundled AI. A bigger discount off the new list can still raise your bill, so negotiate from last term's per-user dollars and your own usage report.

Azure and Google now meter reserved model capacity at pay-as-you-go token weights, so a monthly term needs about 99% utilization to save money. Annual terms break even at 84% on Azure and 75% on Google; Bedrock publishes too little to compute.

Self-hosting LibreChat or Open WebUI on API keys costs about $3.72 a user a month on Sonnet for a typical mix, against $20 seats. A small heavy-user tail on a frontier model erases the saving, so budgets go in before launch.

Exa is the safest default web search API for AI agents, and Tavily is the one to skip. At frontier-model prices, the page text an API returns can cost 6 to 20 times the search fee.

Cohere bills rerank per search and Voyage per token, so candidate count and chunk length decide the bill. Voyage rerank-3 is cheaper at most shapes; self-host BGE for residency or high volume.

A batched small LLM translates a million characters for under $1 and Google NMT charges $20, but full post-editing of the same text runs $13,600 or more. On reviewed content, buy the engine your editors correct least.

E2B and Daytona charge the same $0.0504 per vCPU-hour and Modal Sandboxes 41% more. On a month with long idle tails, the billing shape moves the invoice by 10x or more, so price your own alive-to-busy ratio first.

Claude Haiku 5.5 matches GPT-6 Luna at $0.10/$0.50 per million tokens and beats it on Anthropic's evals, but prompts over 100K tokens cost 5x and its tokenizer counts about 30% more tokens than Haiku 4.5. Short-prompt classification and routing get the big saving; compaction and long-context work get much less.

Decision models such as Jev cost about 25x less per call than Claude Haiku 4.5 at list price, but trail LLMs on accuracy until you add labels and a calibration fit. Fine-tune if you have labels; cascade Jev into an LLM if you don't.

OpenAI's Ultrafast tier runs GPT-6 Astra at $60 input and $300 output per million tokens, 6x Standard, on US or global processing only. It ships off for Enterprise; turn it on per named user with a spend limit.

A one-year H100 reservation breaks even at 65-80% measured utilisation, depending on the provider. Commit only your 90-day floor for six months, run checkpointed training on preemptible capacity, and pay on-demand for the burst.

Microsoft's rebuilt Copilot keeps chat and Office features on the $30 seat and meters Autopilot, Code, Cowork and the Fable and Astra models in Copilot Credits. Set per-user limits and turn off Auto-apply before it reaches your tenant.

On a 30-turn agent loop, prompt caching saves 80-89% depending on the provider, not 90%. A timestamp in the system prompt drops Claude's saving to 6%, and a volatile tool definition makes caching cost more than not caching.

Nebius bought Inferize to cut GPU cold starts in Token Factory from minutes to seconds, but published no date, price change or capacity guarantee, and its own docs warn released capacity may not come back. Keep your minimum replicas until both are in the contract.

Prompt caching cuts a 400K-token long-context answer to $0.085 on Claude Sonnet 5.5, but RAG still answers for $0.017. The crossover formula, the accuracy-at-depth evidence and the latency users notice.

No FinOps platform can split a shared API key. Tag AI spend at the gateway and the provider first, reconcile against the cost API, then add Vantage for one view and OpenCost for GPUs. Only the gateway can cap a budget.

GPT-6 Sol costs $2/$10 per million tokens, half GPT-5.6 Sol's promo rate, permanently. But OpenAI's headline evals ran at xhigh or max effort, and only one carries a dollar cost per task.

Claude Opus 5.5 cuts per-token prices 20% and cache reads 60%, but forced tool_choice and disabled thinking now return HTTP 400, the default effort drops to medium, and the 40% saving compares different effort levels.

Grok 4.7's per-token price didn't move. Its per-task cost did: up 47% at matched effort on Artificial Analysis's index, down 10% on Cursor's own benchmark. Price it on your workload.

A 40-call black-box audit from USTC flagged 7 of 15 OpenAI-compatible GPT services for behaviour consistent with output-token inflation. The flags are not proof, but padded answers pass quality evals and invoice checks alike, so measure output tokens per task against the first-party API.

At 10,000 users, Postgres with pgvector costs $163-$332 a month against Zep's $375-$750 and Letta's $1,020 — and no independent party has confirmed that a commercial memory store retrieves better. Buy one when you can name the column you could not write yourself.

A model router sells two things: mechanical failover, which works, and a learned classifier that picks your model, which the neutral benchmarks say does not. Buy the first, build the second, and plan on 35% savings rather than 60%.

Priced on one normalised agent workload, a month of inference costs $370 on GPT-5.6 Luna and $10,725 on Claude Opus 5. Output, reasoning and cache-read rates explain the gap, not the input price.

Training a custom model costs about $30. Serving the same weights costs $326 to $28,908 a month depending on the platform. The fine-tuning-versus-RAG decision is a serving-contract decision, and the break-even is 683,000 queries.

A Copilot seat lists at $19 to $30 a month, but agent consumption, a credit pool that expires and per-gibibyte indexing decide the bill. One GitHub Copilot developer running a daily agent session costs $174 a month against a $19 seat.

Sixty-four B200s cost between $1.93M and $9.03M for a year depending only on how you buy them. But at the 5% GPU utilisation Cast AI measured across 23,000 production clusters, a committed $3.44/hr rate bills out at $68.80 per useful GPU-hour — and the break-even duty cycle is 48%.

GitHub's HydraFusion router bills every model it invokes at standard rates, and two of its three patterns fire two or three models per turn. The measured saving swings from 36% to 67% across just three benchmarks — and the one that looks most like enterprise work saved least.

Databricks took managed Iceberg tables to GA, so the file format is no longer the lock — but external writes are still Private Preview, and the real markup is GPU serving at $5.50 an A100-hour against $2.79 at Lambda. Move the serving, keep the lakehouse.

A published trace of one 800-line pull request shows a coding agent billed roughly 156 million tokens to produce 289,000 — 98% of the volume was cache re-reads of context re-sent on every one of 512 turns. Coding-agent cost is set by turns per task and the width of each read, not by the model's rate card.

Google began billing Memory Bank and Sessions on September 1, charging $0.30/GiB-month on total data stored including every revision. Memories carry no default TTL, so four months of agent state accumulated before the first metered invoice.

Anthropic cut Claude Fable 5.1 cache reads 75% to $0.25/MTok, half of Opus 5's rate on a model with twice the base input price. But Opus 5 stays cheaper until cache reads pass a third of your bill, and Anthropic's own savings ceiling of 45% falls short of the 50% crossover.

OpenAI bought tens of thousands of Mac minis to train computer-use agents, and Anthropic rents Mac capacity through AWS. The reason matters to your pilot: a desktop agent's unit of cost is a machine, not a token, and Apple's licence caps you at two virtual desktops per Mac while every cloud Mac host bills a 24-hour minimum.

Anthropic's August 13 Claude Tag update replaced the per-message classifier with full-channel context. The agent's read scope is now the channel's whole conversation, the member list is the control that governs it, and Anthropic's own docs now say there is no per-action log of who asked.

Nvidia carries $279B in supply commitments, an increase it attributes primarily to memory, and still guided gross margin from 75% down to a 71-72% trough on RAM costs. Your server quote has no such contract behind it, and every self-hosted inference TCO model built on 2025 hardware prices is now understated.

At two million pages a month, Amazon Textract's Forms + Tables + Queries bundle bills $125,000 and a Gemini Flash-Lite schema prompt bills $3,155. The gap is a pricing structure, not a technology gap — buy OCR and field extraction separately.

Priced against one normalised workload — 5M LLM calls a month across 4 models — Azure API Management runs about $700/month, Kong about $1,400 and Apigee about $3,931. None of the three enforces a hard budget cap.

A controlled 90-trial experiment found Claude Haiku 4.5 spent 735K tokens where Sonnet 4.6 spent 640K, for a result 1.9 points worse. Downgrading a coding agent to a cheap tier saves less than the rate card implies, varies fivefold by vendor, and only holds up if you replace prose design docs with machine-checkable contracts.

OpenAI cut GPT-5.6 Sol to $4/$20 per million tokens but guarantees the rate only "at least through November 21, 2026," with no successor published. The cut was asymmetric, so no two workloads saved the same amount — and the reversion is +50% on output, not +33%.

Priced on one 10,000-conversation workload, the same agentic work costs $5,940 or $20,000 a month depending only on the billing unit. Per-resolution is the only meter where the vendor absorbs the 30x token variance instead of the buyer.

Eleven AI observability options priced against one workload — 500,000 agent runs a month, 5 million spans, 10 GB of traces. The same telemetry costs $49 at SigNoz Cloud and about $930 at W&B Weave, and retention is a bigger multiplier than traffic.

Modelled at three workload sizes on 18 August 2026: managed pgvector runs about $1,910/month at 10M vectors against Pinecone's $19,716 on-demand. But the real problem is that Pinecone's cheaper Dedicated Read Node rate and Qdrant Cloud's per-unit rates are not published anywhere.

DoiT bought Attribute, the eBPF-based AI cost-attribution startup, for an estimated $65M. DoiT also resells the cloud it now meters — and the kernel sensor you approved has a new owner nobody re-reviewed.

A production RAG pipeline at 10M tokens a day costs about $1,308 a month, 61% of it generation and one dollar of it embeddings. The line that wrecks the budget is document parsing, which swings from $2,100 to $90,000 over the same 2 million pages.

AgentCore's new Instances compute type runs agents on EC2 in your own account for up to 14 days. AWS's documentation states that agents sharing an instance get no security boundary, that the platform does not verify a session ID belongs to the caller, and that the billable instances are hidden from your EC2 console by default.

Seven agent monitoring tools priced against the same workload: 50,000 runs a month at 12 steps each. Self-hosted Langfuse wins on cost and audit depth — but only one of these tools sits in the request path where it can actually stop a runaway agent, and that is the part you have to build yourself.

Self-host LiteLLM: per-team budgets and virtual keys are in the free open-source tier, while everyone else gates enforcement behind a sales call. Priced through one 50M-request workload, the platform layer ranges from $7 to $10,300 a month.

DeepSeek warned of a significant API price rise with no number and no date. Because the weights are MIT-licensed, the ceiling is already public: independent hosts serving the identical model charge 3-4x on posted rates and 40x on cache hits.

Manifest killed its four-tier LLM router after four months and 7,000 users, and the arithmetic explains why: cache reads bill at 10% of base input, so routing an agent step to a model 2.5x cheaper makes it 3.5x more expensive. Route at the session boundary, not the request.

Avalara surveyed 1,505 CFOs: 92% feel pressure to prove AI ROI, 50% see only limited returns. McKinsey's AI FinOps framework is the answer most teams are missing.

OpenAI CFO Sarah Friar published a four-question AI ROI framework — 'Useful Intelligence per Dollar' — as enterprise AI spending hits $2.59 trillion with 72% of investments destroying value through waste. The scorecard demands enterprises stop measuring tokens consumed and start measuring work accomplished. Enterprise AI ROI calculator, cost-per-successful-task measurement framework, and AI investment maturity assessment inside.

On July 2, 2026, Crunchbase published data that should have triggered an emergency vendor review at every Fortune 500 company. Global startup investment hit a record $510 billion in the first half of 2026 — surpassing the $440 billion invested in all of 2025. But the real story isn't the record. It's the concentration: OpenAI and Anthropic alone absorbed $217 billion, or 43% of every venture dollar deployed on Earth. Your AI stack almost certainly runs on one or both of these companies. And whether they succeed spectacularly or collapse under their own economics, you have a problem either way.