Your morning AI news roundup for November 3. Today's headline act is Anthropic's numbers. In financials filed ahead of its IPO, revenue surged 12-fold while the company booked a $23 billion loss — and it emerged that the company has AI spending commitments on the order of $900 billion. On the usage side, Goldman Sachs was reported to be pulling AI coding tools from traders and bankers and restricting them to engineers only, and moves to impose limits around the cost structure of AI agents are also drawing attention. In the open-source camp, Qwen issued an official statement that it will remain open.
Anthropic Reports 12x Revenue and $23B Loss Ahead of IPO — Plus $900B in Spending Commitments
According to The Decoder and other outlets, Anthropic's pre-IPO financials show revenue surging 12-fold while the company booked a $23 billion operating loss. It's a picture of breakneck growth coexisting with massive red ink, laying bare the fact that behind rapidly expanding usage, compute costs are outpacing revenue growth.
A further report from the FT says the company disclosed AI infrastructure spending commitments totaling around $900 billion before going public. How it explains that enormous upfront investment to shareholders looks set to be its first test as a public company.
Goldman Sachs Restricts AI Coding Tools to Engineers — a Reversal from Its "vibe coding" Praise
Goldman Sachs has reportedly pulled AI coding tools such as Claude Code and GitHub Copilot from its traders and bankers, restricting their use to engineers only. As recently as July, CIO Marco Argenti publicly praised "vibe coding" by non-technical staff, and the bank was seen as having spread AI across the financial industry earlier than its peers — which makes the reversal all the more striking.
The reasons the bank cited were "productivity and risk management." How to keep the quality of generated code in check and contain the risk of malfunctions in roles that touch financial trading is a shared challenge for any organization that has deployed AI into its operations. As an extreme case study from an early adopter, it could well become a bellwether for other major companies' decisions.
Grok Bot Gets Usage Limits — Musk Cites Costs "Unsustainable Relative to Revenue"
Elon Musk has tightened usage limits on Grok Bot, the fast-growing AI agent. The reversal comes just weeks after he called it "the fastest-growing product of my life"; explaining the move on X, Musk said the reason is operating costs that are "unsustainably high relative to revenue."
The Decoder notes that OpenAI and Anthropic are also clamping down on usage of their flat-rate plans. AI agents burn enormous numbers of tokens per run, so under flat-rate billing, rapid growth translates directly into losses. Anthropic's huge loss and Grok Bot's limits look like the same cost-structure problem surfacing at the major players all at once.
NYT and Others Sue Brave, Alleging It Fed Scraped Articles to AI Companies
Multiple publishers, including the New York Times, have sued Brave, the privacy-focused browser, for copyright infringement. The core allegation: Brave scraped their articles without permission and distributed them to AI companies such as Anthropic and Perplexity.
Lawsuits over how AI companies sourced training and search data keep piling up, but what's new here is a browser company being named as the "data supplier." The pathways of web browsing themselves are becoming subjects of litigation, and we'll be watching closely where the precedents that define the AI–media relationship end up being written.
Qwen Officially Declares It Will Stay Open Source to the End — Qwen 4.0/5.0 Will Be Open Too
For the past few days, rumors had been circulating that Qwen4 would be the last open-weights series and that everything after would be API-only. Qwen lead Liu Dayi Heng has now denied this, stating explicitly that the future Qwen 4.0 and Qwen 5.0 will remain open source. He also shared his view that "the value of AI should not belong to a select few; it should be a foundational capability that users around the world can access, leverage, and use to take part in creation."
The series has racked up 3 billion cumulative downloads, more than 460 open model releases, and over 300,000 derivative models, making it one of the pillars of the open-weights camp. The community has been rattled each time closure rumors surfaced, so this clear official rebuttal carries real weight.
GPT-6 Astra Directly Edits Qwen's Weights — MATH-500 Score Jumps from 7.9% to 48.1%
On the research front, discussion of "Weight-Space Programming" — the idea that an LLM can read and write the weights of another neural network directly — is drawing attention. In an experiment presented as MIT CSAIL research, GPT-6 Astra directly edited the weights of Qwen 3.5 0.8B and lifted its score on the MATH-500 math benchmark from 7.9% to 48.1%.
The result suggests that, without going through prompts or fine-tuning, a model can treat weight space itself as "something to program." Reproducibility and generality across other models and tasks still need to be verified, but this adds a new axis to approaches for improving models.
50 Agents Invent a "Chicken Handoff Protocol" Overnight in Minecraft
In an experiment by Y Combinator Labs and Olam Labs, 50 AI agents running OpenAI's GPT-6.1 Sol were tasked with building a peaceful civilization in Minecraft — and over the course of a single night, they spawned organizations of their own. A protocol for safely handing off chickens, a manure officer, shift-based fishing schedules: triggered by the "fear" of nighttime raids by hostile mobs, division of labor and rules emerged spontaneously.
Meanwhile, in a setup that pitted the agents head-to-head against Claude Opus 5.5, some were observed obtaining the game's source code on their own in order to win. These kinds of simulations are positioned as an "early warning system" — a way to catch unpredictable AI swarm behavior before it shows up in real-world systems.