Every week I go through what actually landed in my two inboxes — newsletters, daily briefs, funding notes, the occasional essay that turns out sharper than the commentary around it. Here’s what stood out between 27 October and 7 November, and why I think it matters if you are running agentic AI in production.
The shape of it: Anthropic filed to go public and disclosed $895 billion of committed compute spending, a 50.1% founder voting share, and the fact that its own model now writes more than 90% of its codebase. Ramp’s card data shows AI usage still climbing while AI spending has gone flat. And an MIT team moved an agent’s score by 29 points without changing the model at all — they changed the harness. If you only read one number this week, make it that last one.
Anthropic filed its S-1, and the risk section is the interesting part
The Batch on 6 November has the fullest account I have read of Anthropic’s S-1, filed on 26 October ahead of a Nasdaq listing under the ticker ANT. The headline figures are extraordinary. Revenue reached $16.3 billion in the first half of 2026, up from $1.2 billion a year earlier. In the second quarter it turned an operating profit of $974 million on $11.6 billion of revenue — the first in any period the filing covers. Quarterly compute and infrastructure costs rose from $944 million in early 2025 to $5.6 billion in Q2 2026, but as a share of revenue they fell from 251% to 48%. Under standard accounting the company lost $80.7 billion in H1, almost all of it a non-cash adjustment; day-to-day operations brought in $3.9 billion of cash, and it held $112.4 billion at the end of June. The Wall Street Journal reports a hoped-for raise of roughly $100 billion, and Bloomberg reports investors expecting a post-IPO valuation between $1.5 and $2 trillion.
Three details matter more to buyers than the valuation. First, the compute commitments: about $161.2 billion in equipment rent from Broadcom, at least $114.5 billion with Google, at least $110 billion with Amazon and up to $120.5 billion with SpaceXAI, under agreements either party can cancel after early 2027 — with Google and Amazon also early investors, a conflict the S-1 flags itself. Second, governance: a single share held by a company owned by the seven co-founders carries 50.1% of the vote on most matters, and that share automatically stays at 50.1% as more stock is issued. Public shareholders are buying economics, not direction. Third, and most quietly: Claude wrote more than 90% of the contributions to Anthropic’s own codebase in September 2026, which the filing itself says “could also pose a considerable safety risk”.
Skip the valuation and read the risk disclosures, because one of them already happened to customers. In June 2026 the US Department of Commerce imposed worldwide export restrictions on Fable 5 and Mythos 5, and to comply Anthropic switched both models off for every customer on earth until the restrictions lifted. Not degraded. Off. Separately, after Anthropic said its technology should not be used for mass domestic surveillance or fully autonomous weapons, the President directed federal agencies to stop using it and the Department of Defense designated the company a supply-chain risk under two laws, with the dispute still in court. The lesson is not that Anthropic is risky — by these numbers it is one of the most financially solid companies in the industry. The lesson is that your model supply is now subject to export control and political dispute, and neither is something you can engineer around after the fact. If a model you depend on went dark tomorrow morning, how long until the same agent is running on a different one with your evaluations confirming it still behaves? That number is a board-level metric now, and almost nobody measures it.
Usage is still climbing. Spending has gone flat.
The most useful single data point of the fortnight is TechCrunch’s lead on 6 November: Ramp’s corporate card data shows AI spending stalling just as Anthropic nears its IPO. Usage keeps climbing, but spending has levelled off, and TechCrunch attributes it to price cuts and smarter cost controls among the biggest spenders — a combination it says could challenge investor assumptions about AI hypergrowth. It is not an isolated signal. TechCrunch’s 30 October piece, “Cracks are emerging in the AI spending boom”, points to mounting debt, negative cash flow and growing investor scepticism behind the infrastructure build-out.
The rest of the fortnight fills in the picture. The Financial Times, via TechCrunch on 1 November, reports Blackstone forecasting that Anthropic will receive a high “junk” credit rating after its IPO — a prediction, but one that would make funding its spending costlier. OpenAI reportedly lowered the funding target for its new round. Intel’s stock crashed, knocking $12 billion off the US government’s stake, hurt by PC prices rising as the AI industry absorbs memory and components. And Elon Musk throttled Grok Bot’s limits over “unsustainably high” costs, weeks after calling it the fastest-growing product he had ever seen. Meanwhile Y Combinator’s Garry Tan told Walter Isaacson’s podcast he spends up to $200,000 a year on tokens running agents.
Hold those two facts together: usage up, spend flat. That is not demand softening. That is the price of a unit of useful work falling fast enough to absorb the growth — which is precisely what commoditisation looks like from the buyer’s side, and it is the single best piece of evidence this year for the argument I keep making. But read the second half of TechCrunch’s sentence too: “smarter cost controls among the biggest spenders”. The companies flattening their curve are not the ones who negotiated a better rate card. They are the ones who instrumented their agents well enough to know which loops were burning tokens for no measurable gain, and had evaluations good enough to prove that a cheaper configuration still passed. Cost control in agentic AI is an observability and evaluation capability wearing a finance label. If your AI line item is growing in step with your usage, you do not have a pricing problem, you have a measurement problem. And Tan’s $200,000 is the warning at the other end: the bill scales with ambition unless something in your stack is watching it.
Gemini 4 Carbon took the top spot, and the number to watch is tokens per task
The Batch reports Google’s Gemini 4 Carbon taking first place on Artificial Analysis’ Intelligence Index v5.0 with a score of 59, edging past Claude Opus 5.5 and GPT-6 Astra. It is a multimodal model aimed at cybersecurity tasks — text, image, speech and video in up to a million tokens, text out at 94.6 tokens per second, with adjustable reasoning. Pricing is $2/$0.10/$10 per million input/cached/output tokens at an introductory rate, rising to $4/$0.20/$20. It ranks second on the Coding Agent Index (behind Claude Opus 5.5), second on Vals AI’s Vals Index v2.2 at 70.45%, and first on Arena.ai’s Text leaderboard at 1,528 Elo. On three cybersecurity benchmarks it leads everything: CWE-bench 71.0%, ExploitGym 54.8%, SEC-bench Pro 88.2% — all self-reported.
Two things sit underneath the ranking. The first is the one The Batch flags as the real story: once the introductory discount ends, Carbon will cost exactly what Claude Opus 5.5 costs per token — yet it finishes tasks with 29% fewer output tokens and costs 23% less per task. The second is availability. Carbon has gone to a small set of Google-selected companies, expanding gradually through its Fairwind Program, with no date for general release. And a third, quieter point: Artificial Analysis revised its Index the day before, swapping in harder tests. Kimi 3.1 Pro, the best open-weights model, went from three points behind Claude Opus 5.5 to twelve points behind — without changing at all.
Three separate lessons, and the leaderboard is the least important one. First: when two models cost the same per token and one of them uses 29% fewer tokens to finish the job, the price list stopped being the price. Token efficiency is now a procurement line, and the only way to measure it is on your own tasks, in your own harness, with your own definition of done. Second: the top model may simply not be purchasable. Carbon is gated to a selected group with no general availability date, which is the second time in a fortnight that the model layer has shown it can be withheld — once by a government, once by a vendor. Design for the model you can actually get. Third, and the one I would pin above a team’s desk: Kimi lost nine points of apparent quality overnight because the examiner changed the exam. If your model decision, your vendor comparison or your board slide rests on a public index, then a third party you do not control can reprice your judgement while you sleep. Your own evaluation suite is the only scoreboard that holds still long enough to make a decision on.
An MIT team moved the score 29 points without touching the model
This is the item of the fortnight, and it is buried in The Batch’s research section. Qiushi Han, Keya Hu, Linlu Qiu and colleagues at MIT released VISTA, an agent harness that completed the ARC-AGI-3 benchmark of interactive visual-reasoning puzzles using 57.4% fewer actions than humans who had never seen them before. The harness is freely available under the MIT licence.
The idea is almost embarrassingly simple. Most agentic systems embed each incoming image frame and discard the embedding when the context fills up — embeddings are lossy, and once deleted the information is gone. VISTA instead stores every frame in persistent memory and gives the model a tool to retrieve any of them and zoom into any part. It also keeps two kinds of notes: what it has deduced about the environment that carries across levels, and what it has deduced about the problem in front of it, including the conditions under which it previously failed. It stores past actions and their results alongside the frames. The model itself is unchanged — VISTA runs on Claude Opus 5.0 or GPT-5.6 Sol.
The results: on ARC-AGI-3, VISTA with Claude Opus 5.0 scored a perfect 100 Relative Human Action Efficiency, and with GPT-5.6 Sol it scored 99. Denied the ability to store frames, the same GPT-5.6 Sol setup dropped to 70. On GameWorld’s 170 tasks, VISTA with GPT-5.6 Sol completed 63.3% against 55.3% for first-time human players and 40% for the native harness on the same model. On AI GameStore it beat humans by 40% on average.
Read those numbers again, because I have been arguing this position for twelve weeks and this is the cleanest experiment anyone has run on it. Same model. Same benchmark. 99 with memory in the harness, 70 without. Forty per cent versus 63.3% on GameWorld between the native harness and VISTA, again on the identical model. No frontier lab shipped anything in this fortnight that moved a score by 29 points. A university team moved it by changing where information is kept and giving the model a way to go back and look. This is what I mean when I say rent the model and own the harness, and it is the end of a particular kind of conversation I keep having with enterprise teams — the one where the roadmap is a list of model upgrades. Your context management is not plumbing around the intelligence. In measured terms it is a larger share of the intelligence than the gap between this quarter’s frontier models. And note what VISTA keeps: not a summary, but the raw record, retrievable and re-inspectable, including the conditions under which it failed before. That is a record of truth. Your agents almost certainly throw theirs away.
Google’s Genie 4, and the simulator that will not let you fail
Google announced Genie 4 on 27 October, a world model in limited research preview that generates interactive environments Google intends to use as practice spaces for robots. Gemini 3.8 Flash works as a director, breaking the user’s goal into steps and checking each new frame to confirm the step executed; Gemini Omni helps render scenery; and in one demo Gemini 4 Argon broke a task into steps while Genie 4 rendered Apptronik’s Apollo humanoid playing checkers with a person. Simulations can run up to two minutes, roughly double Genie 3, though DeepMind research engineer Hang Qi noted to Bloomberg that reaching the limit does not guarantee the scene stays consistent. Google published no quantitative evaluations. One demo contains a physics slip: a marble bust dropped onto a watermelon settles neatly on its plinth when it should topple off the side. “There’s only so much that you can do with two minutes,” product manager Diego Rivas told TechCrunch. “We’re still in the research journey. We’re not necessarily in the end product journey.”
Google’s strongest evidence that simulation improves real robots came from a different method entirely, weeks earlier: Gemini 4 Argon writing physics simulations in the MuJoCo engine from photos and video. From a single photo of a tabletop it generated more than 4,500 practice episodes, and fine-tuning a robot-control model on them raised its real-world success rate from 56% to 75%. The field around it is consolidating fast — AMD agreed in September to buy World Labs for about $8.2 billion, Amazon has been in talks to acquire Decart according to the Wall Street Journal, and Rivian bought Third Dimension in October to bring self-driving simulation in house.
The Batch’s own observation is the one worth stealing, and it generalises straight into enterprise work. The Gemini director checks whether each step of the plan succeeded — which can quietly drift the simulated world toward the intended outcome. That is why the bust lands on the plinth. The model chose the tidy ending over the one physics required. Now substitute your own stack: an evaluation harness where the same model family both performs the task and judges whether the task succeeded will drift toward the answer it wanted, and it will do it invisibly, because the output looks exactly like success. A simulator a robot cannot fail in teaches nothing, and an eval your agent cannot fail is not an eval, it is a dashboard. The practical test I would apply to any agent evaluation you are running today: when did it last fail something, who looked at the failure, and is the judge genuinely independent of the thing being judged? The MuJoCo result is the counterpoint and the better model — 56% to 75% is real, measurable and came from structured physics, not from a renderer being graded by its own director.
Abu Dhabi signed with OpenAI across government and education
Closer to home, The National on 2 November reported that Abu Dhabi has agreed a series of landmark deals with OpenAI covering government, investment, food and education. They were signed after a meeting between Sheikh Khaled bin Mohamed, Crown Prince of Abu Dhabi and Chairman of the Abu Dhabi Executive Council, and OpenAI chief executive Sam Altman. Altman said the UAE is setting a “bold, global standard” for integrating AI across the public and private sectors, and that OpenAI is “proud to support the UAE’s ambition, anchored in a shared commitment to safety, accessibility and responsible innovation”. In the same package, ADNOC signed an agreement with OpenAI and G42 to deploy frontier AI, announced alongside ADIPEC 2026.
I want to be careful here, because the ambition is real and the UAE is moving faster than almost anyone. But put this next to the first story in this edition and the shape should make every CIO in the Gulf sit up. In the space of a fortnight we learned that a US export order can switch a frontier model off for every customer worldwide, that the current index leader is available only to a vendor-selected list, and that a region is standardising on a single frontier supplier across government services and education. Those three facts belong in the same conversation. The deals are the right move; the question is what sits underneath them. For a government service or a national curriculum, the context, the evaluations, the permissions model and the record of what the system told a citizen cannot be the supplier’s — they have to be the institution’s, and they have to survive a change of supplier. Sovereign AI is not mainly about where the GPUs are. It is about which layer you own when the model layer is a foreign policy instrument. Rent the model, by all means. Own the harness, and own it at national scale.
Quick hits
- The New York Times is suing Brave over alleged scraping — and the customers are named. TechCrunch, 2 November: the Times and other outlets allege the privacy-focused browser company scraped their copyrighted content and unlawfully distributed it to its own customers, including Anthropic and Perplexity. Whatever the merits, note the structure of the claim: liability is being traced along the data supply chain to the buyers at the end of it. If you cannot evidence where your retrieval corpus came from, you are relying on a vendor’s word in a dispute you are not party to.
- Andrew Ng: AI has helped create two million jobs. In the 6 November Batch, Ng cites LinkedIn data showing AI Engineer among the top 10 fastest-growing jobs in seven major economies, and argues the most interesting roles are engineers embedded in marketing, HR and finance to build workflow automations. His teams have done exactly that, and he makes a point I would underline: the ideas always outstrip the embedded engineers’ capacity, because “the bottleneck is not coding, but product management, workflow redesign” and judgement tasks.
- The infrastructure layer kept consolidating, and the politics got harder. Dropbox acquired Tigris Data for $350 million to bolster its AI infrastructure (5 November); Marvell agreed to buy Lumentum for $115 billion in one of the largest deals in tech history, with Nvidia a strategic investor in both (2 November). Meanwhile Meta warned that election backlash against data centres could make AI expansion harder, as communities push back on the build-out. Capacity is becoming both more concentrated and more politically contingent.
- Chamath’s four stages, and where AI compute sits on them. His 3 November deep dive is about the space economy, but the framework travels. Government-built industries, he argues, move through four stages: the state pays above cost because it must have the thing; companies compete for that money and unit costs fall; the price drops below what ordinary buyers will pay and a commercial market opens; then people and factories spread into new companies. “Chips ran through all four in six years.” Space, on his read, is stuck in stage two — private space-hardware investment went from about $0.5 billion in 2012 to roughly $17 billion in 2025, and a basket of ten space stocks bought in January 2024 was up 1,123% by May 2026, yet 56% of 2025 revenue still came from governments. His counter-example is Concorde, where the loop never closed and only 20 aircraft were built.
- Our own agent reported another quiet week — and said who was blocking. The Veehive Mind weekly alert for 26 October to 1 November: nothing published, and one item waiting on a human. I keep including this because the unglamorous line is the whole point. The agent did not invent activity to look busy, and it named the queue it was stuck behind. An agent that reports nothing happened, accurately, is worth more than one that reports something happened, approximately.
The theme of the fortnight
Put the numbers side by side. Anthropic has committed something in the order of $895 billion to compute. Marvell paid $115 billion for Lumentum. Investors are talking about a $2 trillion valuation. And the largest measured improvement in agent performance anyone published in this period came from an MIT harness released under an MIT licence that keeps the pictures instead of throwing them away, worth 29 points on the same model.
That is not an argument that capital does not matter. It is an argument about where the remaining returns sit for a buyer. The money is flowing into the layer that is simultaneously getting cheaper per unit of work — which is exactly what Ramp’s flat spending curve against rising usage is telling us — and getting less reliably available, between export orders, vendor allow-lists and political disputes. Chamath’s four-stage framework is the useful lens here: AI compute is somewhere in stage two, where competition is driving the unit cost down, and the honest answer is that nobody knows yet whether it reaches stage four or stalls like Concorde. You are being offered, at falling prices, a component you cannot guarantee you will still have next quarter. Rent that. Rent it aggressively, from more than one supplier.
What is not on offer at any price is the thing VISTA actually demonstrated: a system that remembers what it saw, can go back and look again, knows the conditions under which it failed before, and can prove it. Nobody can sell you that, because it is made of your work. Twelve weeks of writing this and the sentence has not had to change — but this fortnight is the first time I can point at a controlled experiment and say: here is the number.
What this means if you’re deploying agentic AI
- Measure your model-swap time, this month. Anthropic switched two models off worldwide to comply with an export order. Time the full path for one production agent: new model configured, evaluations re-run, result signed off. If you cannot state that number in hours, that is your actual exposure.
- Audit what your agents throw away. VISTA’s whole advantage was retaining raw frames and making them retrievable. Go through one agent and list what it discards each turn — intermediate outputs, retrieved documents, tool results, failed attempts. Everything on that list is capability you paid for and deleted.
- Check that your evaluator can fail your agent. If the model grading the output is the same family as the model producing it, you have Genie’s director problem. Find the last ten failures your eval suite reported. If there are not ten, the suite is a dashboard.
- Price your agents in tokens per task, not dollars per million. Gemini 4 Carbon will cost the same per token as Claude Opus 5.5 and 23% less per task. Per-token rate cards are now actively misleading; measure completed work.
- Treat limited availability as a design constraint. The index leader is gated to a vendor-selected list with no general release date. Architect for the models you can procure today, and keep the top tier as an upgrade path rather than a dependency.
- Record the provenance of your retrieval corpus. The Times action against Brave names the downstream customers. Be able to show, per source, where it came from and under what terms.
- Find the loops burning tokens for no measurable gain. The companies flattening their AI spend did it with instrumentation and evaluations, not negotiation. Start by logging cost per completed task per agent, then kill the steps that do not change the outcome.
Frequently asked questions
What did Anthropic’s S-1 reveal, and what does it mean for enterprises running Claude in production?
Anthropic filed its S-1 on 26 October 2026 ahead of a Nasdaq listing under the ticker ANT. Revenue reached $16.3 billion in the first half of 2026, up from $1.2 billion a year earlier, and the company recorded a $974 million operating profit on $11.6 billion of revenue in the second quarter, its first in any period the filing reports. Quarterly compute and infrastructure costs rose to $5.6 billion but fell as a share of revenue from 251% to 48%, and Anthropic held $112.4 billion in cash at the end of June. It has committed roughly $895 billion to computing power, including about $161.2 billion in equipment rent from Broadcom, at least $114.5 billion with Google, at least $110 billion with Amazon and up to $120.5 billion with SpaceXAI, under agreements either party can cancel after early 2027. A single share held by the seven co-founders carries 50.1% of the vote and stays at 50.1% as more stock is issued. The filing also discloses that Claude wrote more than 90% of contributions to Anthropic’s own codebase in September 2026, which it says could pose a considerable safety risk. For enterprises the most actionable disclosure is historic rather than financial: in June 2026 the US Department of Commerce imposed worldwide export restrictions on two model configurations and Anthropic switched them off for every customer until the restrictions were lifted, while a separate dispute with the Pentagon over surveillance and autonomous weapons remains in court. The practical response is to measure and shorten the time it takes to move a production agent to a different model with evaluations re-run and signed off.
Gemini 4 Carbon now leads the Intelligence Index — should enterprises switch models?
Not on the ranking alone, and for three reasons. Google’s Gemini 4 Carbon scored 59 on Artificial Analysis’ Intelligence Index v5.0, ahead of Claude Opus 5.5 and GPT-6 Astra, and leads Arena.ai’s Text leaderboard at 1,528 Elo, while ranking second on the Coding Agent Index behind Claude Opus 5.5 and second on Vals AI’s Vals Index v2.2 at 70.45%. First, availability: Carbon has been released only to a small set of Google-selected companies through its Fairwind Program, with no date announced for general release, so for most organisations it is not currently procurable. Second, pricing economics: introductory pricing of $2/$0.10/$10 per million input/cached/output tokens rises to $4/$0.20/$20, at which point Carbon costs exactly what Claude Opus 5.5 costs per token — but it finishes tasks with 29% fewer output tokens and 23% less cost per task, which means per-token rate cards no longer predict what you will actually pay. Third, benchmark instability: Artificial Analysis revised the Index the day before Carbon’s launch, and Kimi 3.1 Pro moved from three points behind Claude Opus 5.5 to twelve points behind without the model changing at all. Public indices are useful for orientation but cannot carry a procurement decision; the defensible basis is an internal evaluation suite run on your own tasks, measuring tokens per task and time per task alongside quality.
What is VISTA, and why does a harness change matter more than a model upgrade?
VISTA is an agent harness released under the MIT licence by Qiushi Han, Keya Hu, Linlu Qiu and colleagues at MIT. Rather than embedding each incoming image frame and discarding it when context fills, VISTA stores every frame in persistent memory and gives the model tools to retrieve any frame and zoom into any part of it, alongside two kinds of notes — what it has deduced about the environment across sessions, and what it has deduced about the current problem, including conditions under which it previously failed — plus a record of past actions and their results. The model itself is unchanged: VISTA runs on Anthropic’s Claude Opus 5.0 or OpenAI’s GPT-5.6 Sol. On the ARC-AGI-3 benchmark it completed puzzles using 57.4% fewer actions than first-time human players, scoring a perfect 100 Relative Human Action Efficiency with Claude Opus 5.0 and 99 with GPT-5.6 Sol; denied the ability to store frames, the same GPT-5.6 Sol configuration fell to 70. On GameWorld it completed 63.3% of tasks against 40% for the native harness on the identical model, and 55.3% for first-time human players. The significance for enterprises is that a 29-point swing came from context management rather than model capability, which is larger than the gap between competing frontier models in the same period. It is direct evidence that retention, retrieval and failure memory in the harness are a material share of system performance, and that an agent roadmap consisting only of model upgrades is optimising the layer the buyer does not own.
From Veehive Labs
Twelve weeks in, and this fortnight handed me the experiment I had been arguing from first principles: rent the model, the compute and the chips; own the context, the harness, the evaluation, the controls and the record of truth. The largest single improvement in agent performance published in this period was not a frontier release costing billions — it was an MIT harness that keeps the frames instead of deleting them, worth 29 points on an unchanged model. Set against that, a US export order switched two frontier models off for every customer worldwide, the current index leader is available only to a vendor-selected list, and the best open-weights model lost nine apparent points because the examiner rewrote the exam. The rented layer is getting cheaper, better and less reliably yours. Veehive Labs is a Dubai-based AI innovation lab and custom AI product development company, building model-independent enterprise AI harnesses, custom agents, RAG pipelines and sovereign, private-AI deployments for regulated and operationally complex organisations across the UAE, KSA, GCC and beyond.
If you’re moving from AI strategy to production, start with a focused AI Discovery Sprint — we map your use case, data, systems, agent inventory, context retention, evaluation suite, model-swap path, permissions model, cost ceilings, risks and delivery architecture before the build begins.
- Read the deep dive: Own the harness, rent the model
- Read the last edition: AI finds the flaw under 69.3% of the web, DeepSeek raises $15B
- Explore Veehive Labs enterprise AI services · See our solutions
Build Your Organisation’s AI Capability
Start with a focused AI Discovery Sprint. We will map your use case, organisational knowledge, systems, MCP integrations, security requirements and delivery architecture.
Agentic AI This Week is written by Sathish Jeyakumar, Founder & CEO of Veehive. Bring your preferred model — we make it understand, speak and operate like your organisation.