The week has a single thread that holds different news items together. Google burns 51.4 billion in a quarter and ends with negative free cash flow, Revolut moves 2,000 engineers to Z.ai's GLM and cuts the bill by 90%, OpenAI puts ChatGPT and Codex on the same token meter. Three pieces of the same puzzle: how much it costs to complete a task. This week's headline is not a rhetorical question; it is the yardstick for reading everything else.
Image generated entirely with AI.Who pays for infrastructure?
Alphabet's third-quarter accounts tell the story of an expensive transition. The group invested $51.4 billion in AI infrastructure and ended with negative free cash flow of $9.1 billion. It is the second consecutive quarter in which Google spends more cash than it generates, something that had not happened since its 2004 listing.
Revenue grows 25% to $127.6 billion and operating profit grows 30% to $40.6 billion, both below expectations. Capex runs at $132 billion in the first nine months, with guidance of $195-205 billion at year-end and 2027 still rising according to CFO Anat Ashkenazi. Data centers, energy, networks, and compute capacity are needed: capex becomes a physical constraint, not a number on a sheet.
The sensitive point is Google Search. Search revenue is $63.4 billion, growing 12% versus rates near 20% at the start of the year. Queries in AI Mode monetize less than traditional ones, as the company itself admits. Google has yet to find a way to earn from new searches as much as it earns from those it is progressively transforming.
The figure that matters at company scale is negative free cash flow for two quarters in a row. Spending $51.4 billion in one quarter to build compute capacity, with $42.3 billion generated from operations, is a balance-sheet problem before it is a product problem. Inside a company the same logic is called cost tracking by feature and by model, with a daily spending cap and alerts when it is exceeded. The scale is different, the problem is identical. The question of how much agent autonomy costs is the same, only with more zeros.
What to watch: revenue per thousand AI Mode queries compared with classic search, and the curve of incremental revenue against 2027 capex. If free cash flow remains negative for three or four quarters, the market stops financing the transition lightheartedly.
What does the token meter teach?
Starting in December OpenAI unifies the Chat and Work modes of ChatGPT in a single interface. Plus, Pro, and Business plans share a token allowance between ChatGPT and Codex: more complex models consume more allowance, simple instant chats are not counted. The old flat fee gives way to a meter.
The point is the decoupling between price and perception. Today a Plus user pays a fixed amount and consumes as much as they want, without ever seeing the cost of a single request. Starting in December the same person decides whether it is worth switching to the reasoning model to refactor two hundred lines or whether a lightweight model is enough. It is the same logic that anyone putting AI into production already applies: routing by cost, decided by the budget instead of habit.
The interesting part is that Codex enters the same allowance as chat. Agentic coding and conversation end up on the same meter, and tokens spent on refactor and tests weigh exactly like a question in chat. The risk is psychological: a meter that goes down changes behavior, people stop asking the hard question and use the small model even when the large one would be needed.
In the same week Revolut moved the default model for about 2,000 engineers to Z.ai's GLM, with a bill about one tenth that of Anthropic and OpenAI. The path starts with GLM-5.1 in April 2026 and reaches GLM-5.3 on 56 B200 GPUs, with a stated cost of $0.30 per million tokens. On OpenRouter, in the week from October 26, Chinese providers together collect 49.2% of text requests against 18.8% for OpenAI and 5.7% for Anthropic.
The logic you recognize is that of marginal cost. When a token costs one tenth, what you can afford to automate changes: document reading, double-checking, the extra attempt. The savings turn into work surface covered by AI, and this is the real point. On where competitive advantage ends up the question is open: if the token price collapses and weights become open, the defense shifts to orchestration capability, not the model itself.
There is only one metric that matters, cost per correctly completed task. If GLM-5.3 closes the task on the first attempt, Revolut is right and the rest is noise. If three corrections and a human recheck are needed, the savings are paper.
Who benefits from the agentic web?
In 2001 Tim Berners-Lee published the manifesto of the Semantic Web: an internet where agents navigate on our behalf, provided that every page is annotated. The idea never took hold, because annotating costs work and the benefit is not visible to those who publish.
Large language models have removed that bottleneck. An agent trained on text understands where an address is without tags, and multimodal agents read screenshots and click. From this comes the agentic web: Meta with Muse, OpenAI with Dots, Anthropic with Claude Duet are giving agents access to mailboxes and bank accounts, along with a series of summer incidents in which agents allegedly escaped their sandboxes.
The market is moving. Time publishes versions of its articles designed for agents and sells advertising space addressed to them. Etsy and Expedia have opened dedicated interfaces. Two ways of building all this remain: train the agent to use the browser like a human, or expose tools that speak text and code. The choice decides who controls access, and therefore whether the result will be an open web or a porthole web.
On the opposite front, Apple has opened Siri AI to all apps with the iOS 27.3 betas. App skills let developers describe to Siri how to chain multiple actions inside the app, and the Spotlight semantic index exposes any content without a rigid schema. The constraints weigh: Siri AI is not available in the EU on iOS and iPadOS, requires an Apple Intelligence device in six languages, and server-side functions have daily limits with a likely future cost.
The distance between an agent that opens a screenshot and one that reads a JSON is the distance between a demo and a production system. The part that holds up is the MCP connectors on GitHub, Gmail, Drive: the agent calls a declared action and receives a clean result. When the buyer is an agent, the audience to convince stops being the person who reads and becomes the model that chooses. On why agents leave the perimeter the guardrail issue ties exactly to this choice: who publishes a machine-readable contract and who is left to be clicked by other people's agents.
Image generated entirely with AI.
SundAI comes out once a week, but AI news doesn't wait for Sunday. Every morning I pick the best one and comment on it.
TodayNovember 8, 2026
Open-weight models catch up with closed ones on cyberattacks at a tenth of the cost
Which tools stand up to comparison?
Three signals from this week's tools.
First, tracing and orchestration. Langfuse to see prompts, tokens, and latency for each step of an agent, n8n to line up agents, APIs, and webhooks. Claude Code and Cursor as daily assisted-development tools. Databricks Lakebase with branching to create a database for every pull request and test on realistic data without touching production.
Second, model adaptation. Unsloth for lightweight fine-tuning of open weights such as Qwen on a single GPU, NVIDIA NeMo to turn agent failures into tests and fixes. And the MIT result on GPT-6 Astra: by modifying 0.016% of a Qwen's parameters, math accuracy goes from 7.9% to 48.1%. If the method holds up on replication, you specialize a small model instead of paying a frontier model for every request. Stability after an update and dependence on open weights remain to be verified.
Inkling 2 Small from Thinking Machines also moves along this axis: 594 billion mixture-of-experts parameters, with the challenge shifted to efficiency per agentic task. Token consumption to complete a task rises sharply, and for those buying inference on demand a task that costs three times as much must be justified by a measurably higher success rate.
Third, memory and identity. Mem0 with OpenAI Agents SDK to keep preferences and context alive across different sessions. Okta Cross App Access with Bedrock AgentCore to govern which agent can act for which user. DuckDB Skills for Claude Code to query CSV and Parquet with SQL from the same session.
On the news front: OpenAI's autonomous pipeline finds 948 flaws in the Linux kernel in a single run. Dropbox buys Tigris Data for $350 million and builds storage for agents. Anthropic's $1.5 trillion IPO hits the investors' wall. No agent exceeds 12% on the benchmark for real work, and this gives the measure of how much the cost-per-completed-task metric is still entirely to be built.
The cost per token is a number on an invoice. The cost per correctly completed task is the number that decides which products survive.
