The week closes with Nasdaq wiping out all of the 2026 gains and with Anthropic, just listed, sliding to $88.30 against $83 at the offering. In the same few days, however, concrete tools arrive: Claude joining meetings, Gemini 4 Neon with clear pricing, 9,000 plugins in the MCP marketplace. The two things must be read together. The market stops paying for AI at any price, while daily work absorbs agents into meetings, sheets and whiteboards. Those who remain standing are those able to put cost per result at the center, not the model.
Image generated entirely with AI.Who pays for AI when the price collapses?
Anthropic has filed the prospectus for the largest listing in tech history: a trillion-dollar valuation after five years of existence. Inside the document there are three cracks written in black and white.
The first is commoditization. Open models run about four months behind closed ones according to Epoch AI and cost half as much: Moonshot's Kimi K3.1 Pro is at 50% of the price of Claude Opus 5.5. The price for the same level of performance has collapsed by 97% in less than two years, from $17 to 50 cents. On Vercel, more than half of the text processed in August went through open models, against 13% in April. The second crack is slowing growth, from 145% to 45%. The third is $417 billion in commitments to rent chips and servers, almost 13 years of sales at the first-half pace.
In the same week Nvidia closed the quarter at $115.84 billion in revenue, double the year before, with the board closing the door on a return by Jensen Huang after the allegations that engulfed him. About 90% of the AI chip market passes through there. It is the kind of risk concentration that in production would be avoided on principle: a single node carrying the entire load is an outage waiting to happen.
On Friday Nasdaq closed 5% below January 1, with about $4 trillion in market value gone in a week. Anthropic fell to $88.30 against $83 in the offering, Nvidia lost 14%, SanDisk 14%, Micron 12%. Analysts read it as the IPO draining liquidity: those who want to buy the new shares sell something else. On the same day Google launched Gemini 4 Neon at $3 per million input tokens and $15 output, with no free tier, and a one-million-token context. The first independent review says it handles complex tasks but gets elementary counts wrong, like the beeps in an audio track.
The common denominator is one: cost per result. Those who build with these models already have a dashboard with cost per feature, a daily spending cap and an alert when it is exceeded. The market is doing the same check on billions, and Friday's verdict was bad. The question that remains open is how much a completed task really costs, and on this the Anthropic prospectus answers with an eloquent silence.
Where does work happen when the agent joins meetings?
Anthropic has brought Claude outside the code enclosure with three tools. Claude Meetings brings the model into calls to take notes, Whiteboards gives it a shared whiteboard on which to arrange ideas, Sheets puts it to work on the spreadsheet alongside colleagues. Whiteboards and Sheets are in beta for all paid plans, Meetings remains reserved for Team and Enterprise. There is also a transformation of a presentation into a one-minute explainer video without leaving Claude.
The real number, however, is 9,000 plugins in the MCP marketplace. Interactive connectors open partner apps such as Adobe and Atlassian inside Claude: a PDF can be edited with the native editor without changing tabs. An Anthropic product manager says the next gap is the one for people who do not write software. The point is right: the barrier for non-developers is context, not model capability. Tools that live where people already work attack the right problem, and call latency, note quality on overlapping voices and permissions on transcripts become the real selection criteria.
The same model of work emerges from Satya Nadella at the DealBook Summit, speaking about the buyout offered to 60,000 Microsoft employees. The team building Cowork, the AI agent inside Microsoft 365, has 11 developers. They do not write the product; they build a software factory that produces it. Every employee becomes a manager of agents, work shifts from direct production to directing automated work. On The Pragmatic Engineer podcast Nadella admits that the industry has created enormous anxiety and that the transformation will take longer than expected. In Europe 28% of companies have already cut staff because of AI, and BMW governs hundreds of agents across thousands of developers: the direction is the same, the speed is not.
The crux, in both cases, is control. When an agent writes into a sheet, must it do so in place or return a CSV to reload? The difference decides whether it is a button that writes into the cell or a side chat. Those building a product that lives inside a model's ecosystem know that the model changes under their feet: what remains when the model changes is the question that decides whether a product has a defense. One data point from the week says it clearly: more than 100 AI agents fail from silent drift, not from errors. It is the kind of failure that no benchmark catches and that only human checkpoints see.
Image generated entirely with AI.
Why have permissions become a product?
Meta Platforms Ireland will have to pay €805 million in Italy. The Garante alleges that the Facebook app read photos, videos and metadata stored on the Android smartphones of about 20 million Italian users without specific consent and without adequate information. It is one of the highest fines ever issued by the Italian authority, but the number is the least interesting part.
The point is jurisdiction. Meta raised the GDPR one-stop-shop objection: jurisdiction belonged to the Irish authority. The Garante rejected the objection by building the case on Article 122 of the Privacy Code, which transposes the ePrivacy Directive and regulates access to information stored on terminal devices. On that matter jurisdiction remains national, so the one-stop-shop does not apply. A French precedent in a similar case is cited as the basis for the reasoning.
For a company it means multiplied compliance costs and legal risk that cannot be covered with a single office in Dublin. If other authorities copy the model - Germany, France, Spain - the one-stop-shop is effectively finished for an enormous slice of cases. Those designing mobile products must map permissions device by device. The logic applies identically to agents: when an agent leaves the perimeter, more is understood from edge cases than from benchmarks.
In the practice of MCP connectors - Gmail, Drive, GitHub - explicit permission is the first thing to define, and the user must see what is being read. If the rule becomes that whoever touches the device answers to the local authority, permissions are architecture and must be decided before release, not after the first report. An assistant that listens to company meetings is a governance question before a product one: who accesses the transcripts, with what retention, with what possibility of deletion. The authorities have just shown that the local perimeter is a skeleton key, and it is worth treating it as such.
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 22, 2026
The Riemann hypothesis falls to Claude, and mathematics shifts gear
Which tools are worth trying now?
If there is a common thread in the week's tools, it is measurement. Three items are worth a try.
TokenEconomizer is an open-source plugin that reuses past sessions to cut tokens and costs for coding agents. It is not a model, it is not a platform: it does one thing, and does it at the point where the bill burns. It stands as an example of how to tackle cost per token without changing vendors.
AWS Emulator is a local HTTPS endpoint for testing more than 200 AWS services without an account. Useful for running coding agents in an isolated environment, without touching real resources and without bill surprises. Same for Inkling 2 Mini, a 26-billion-parameter open model that runs offline on a 32 GB Mac: it does not replace a frontier model, but it covers local tasks at fixed cost.
On the production agents front, Gemini API Tool Retrieval brings only relevant tools into context when an agent has hundreds of them, with a sharp cut in tokens. Databricks Genie ZeroOps is an agent in public beta that detects, investigates and proposes the fix for broken data pipelines, with human review for high-impact changes. Amazon Textract Insights extracts fields from documents in natural language, with a confidence score and per-page grounding: the interesting part is precisely the score, because it makes uncertainty measurable.
On the governance side, Copilot SDK 2.0 limits tools and credentials to each individual user and survives the app being closed, while AWS Agent Registry keeps typed and validated metadata to govern agents, MCP servers and skills inside the organization. These are the pieces that make it possible to run multiple agents without each one becoming a case of its own. Before choosing them, it is worth returning to a simple question: which tools remain useful when the model underneath changes, and which instead live only on the week's novelty.
The week teaches one thing only: clear pricing is an advantage, not a problem.
The rest will be told by the next four quarters: the ratio of people to output per team, published inference costs per product, diversification of hardware customers. If the AI bill does not appear anywhere, the story remains a restructuring with a good press office. If it does appear, we will know who really pays.





