The week in two minutes
- Jensen Huang is out at Nvidia. The board accepted his resignation “effective immediately” on November 11 and gave no reason. CFO Colette Kress is interim CEO, earnings are Tuesday, and the FT reports he is already discussing a comeback. The lead.
- Washington moved within a day: House committee chairs want Huang interviewed December 1 and four chip-export bills in the NDAA. The lead.
- Anthropic set its IPO range at $70 to $78, with pricing expected November 18, and told the SEC it has “met” its own AI research and development threshold, based on what it expects from future models. Anthropic.
- Fable 5.2 is 20% cheaper and stricter: a harness that edits earlier history and then sends back a later thinking block now gets a 400 that retrying will not clear. Fable.
- Google shipped Gemini 4 Hydrogen and a Decisions API, and Gemini 3.8 Flash’s price doubles January 1. Google.
- Check your repos, and rotate if the stealer ran: one Fireworks breach reached Notion’s and Cursor’s customers, and the wfx-loot stealer ran in at least 1,806 GitHub repositories. Security.
- Patch Apache httpd and PostgreSQL, the week Project Zero cut its disclosure clock. Disclosure.
- ChatGPT got a wallet, and United banned AI agents from booking tickets. Checkout.
📡 Effective Immediately
Nvidia’s board accepted Jensen Huang’s resignation as president, CEO and director on Wednesday, November 11, “effective immediately,” and the filing that says so gives no reason. CFO Colette Kress is interim CEO and keeps the CFO job, and longtime director Harvey C. Jones becomes the company’s “first-ever” board chair, per Nvidia’s release. The board says it “will consider internal and external candidates” and expects “operations, customer commitments and product roadmap to continue without interruption.” The first test of that sentence is tomorrow, Tuesday, November 17: third-quarter results and the earnings call at 2 p.m. PT. #039 left this story at the board investigation opened Sunday, November 8.
What Nvidia has said, and what it hasn’t
The company’s own record is short. On November 8, Huang wrote to employees “I was unfaithful to my wife” and said he was “fully committed to NVIDIA.” On November 10 the independent directors hired former acting U.S. Attorney Joon H. Kim and Wachtell, Lipton, Rosen & Katz to run the investigation they had opened two days earlier “in light of recent reporting that founder and CEO Jensen Huang has been the target of a foreign intelligence effort.” Then came the resignation.
Every account of the cause comes from people who would not be named. The New York Times, citing five people who worked at the company, reported that Huang had a relationship with an executive who reports to him. The Wall Street Journal, citing people familiar with the matter, reported that the board was preparing to oust him. Reuters, citing two people, reported a relationship with another employee. Treat the cause as reported, not established.
The weekend brought a campaign to undo it. The Financial Times reported, citing people familiar with the matter, that Huang has discussed a return as CEO with confidants. A “Stand with Jensen” petition has passed 170,000 signatures (affiliations are self-reported), and Altimeter’s Brad Gerstner and Atreides’ Gavin Baker urged the board to consider all options, his return included, per CNBC. Nvidia’s board had told CNBC on Friday that it is “in the process of selecting a search firm” and has “not formally engaged any potential candidates.” The board’s investigation continues.
Washington is not waiting
On November 12 the Republican chairs of the House China, Foreign Affairs and Intelligence committees asked Huang to sit for a closed-door, transcribed interview on December 1, and Nvidia general counsel Tim Teter on December 2, warning each that a subpoena would follow without written acceptance by 8 p.m. EST the next day. Foreign Affairs chair Brian Mast told reporters that day that Huang would testify before Thanksgiving, saying he “has agreed to come” and that “We’re in negotiations for which day,” and Senate Banking chair Tim Scott asked Huang to testify publicly on December 3 and Commerce Secretary Howard Lutnick on December 10, CNBC reported. In a joint statement released with the letters, the House chairs called the case “a prime example of why Congress must immediately pass the AI OVERWATCH Act, Chip Security Act, MATCH Act, and Remote Access Security Act in the NDAA,” the annual defense bill.
That last bill is the one to read if you rent out compute. The Remote Access Security Act (H.R. 2683) passed the House in January and has sat in the Senate since. It would let the Commerce Department control foreign persons’ remote access, over the internet or a cloud service, to items under U.S. export controls, advanced chips included, where the Secretary determines the use could pose a serious risk to national security or foreign policy. The Financial Times reported Sunday that the bills are likely to be attached to the defense bill, though a Senate Republican aide told the paper “time is short.”
The pressure is coming from both parties. Senator Ron Wyden and Representative Richard Neal, the top Democrats on the tax committees, asked Lutnick on November 12 to preserve all records on the Section 232 chip tariffs, Nvidia’s export licenses and the department’s dealings with Nvidia and Huang. And Singapore’s Home Affairs ministry said the same day that it had investigated a former executive of Nvidia customer Megaspeed over possible false declarations about servers and “not … for espionage,” found insufficient evidence but barred her from returning for giving it inaccurate information, and is still investigating the company.
Two claims to hold loosely
The financing story. The Wall Street Journal’s profile of Kress reports, citing people familiar with the matter, that Nvidia “was in talks to provide billions of dollars of guarantees to companies that couldn’t afford its chips,” and that Kress pushed to rein those deals in. Headlines built on it, such as TechSpot’s “Nvidia deliberately kept GPU prices high through circular financing deals, report claims,” go further than the Journal does: it reports financing for buyers, not a scheme to hold prices up.
The fingerprint. Will DePue’s report says the published weights of six Chinese models, from DeepSeek, Moonshot, MiniMax and Ant Group, carry traces of how they were first initialized that match Nvidia’s Blackwell GPUs. In his own words this “supports, without proving” that Blackwell chips trained them, and the traces “do not establish where the chips were located, how access was obtained, or whether export controls were violated.” No independent replication had been published as of Sunday.
Why it matters: Nothing you can buy changed. There is no new roadmap, price or allocation. What changed is the politics around the chips. The executive who “had battled to ease export rules,” in the FT’s words, left in the same week House chairs demanded four control bills in the defense bill, and one of them reaches cloud access, not just hardware. If your business rents GPUs across borders, tomorrow’s call is the first public place Nvidia’s interim CEO can be asked about it.
Hype vs. Reality: Real. The resignation is in an SEC filing and the letters are public. The reason is not on any record: every account of why rests on unnamed sources, and so does the report that Huang himself has discussed a return.
💰 Anthropic Sets a Price Range
And tells the SEC it has met its own AI research threshold, on what it expects from future models
Anthropic’s amended S-1, filed November 10, offers 825 million Class A shares at an estimated $70 to $78 each and says the company has applied to list on Nasdaq as ANT. Alongside the IPO, Google, Nvidia, AMD, Broadcom, Micron, Microsoft and TSMC are to buy $26 billion of non-voting Class N stock at the IPO price, subject to conditions. At the $74 midpoint Anthropic expects about $84.6 billion in net proceeds from the two together. Its IPO site says pricing is expected November 18 and trading November 19, “subject to regulatory clearance.”
On the filing’s count of about 13.09 billion shares outstanding afterward, the range works out to roughly $916 billion to $1.02 trillion, by our arithmetic. The “$1.4 trillion” in headlines is CNBC’s top-of-range, fully diluted figure, which counts shares the basic figure leaves out, such as options, unvested stock and convertible notes.
The filing puts run-rate revenue at about $81 billion as of October 31, up $10 billion in a month. That is an annualized figure from a trailing 28-day window, not GAAP revenue. Committed spend is now about $900.9 billion, “substantially all” of it compute and infrastructure, up from the $895.4 billion #039 reported. One commitment got shakier: Lambda missed a facility-lease deadline, so the August Lambda agreement, about $34.9 billion, “is currently terminable by either party.” Neither side has given notice, and Anthropic says it does not currently intend to end it.
Then there is the sentence that is not about money. The November 2 version said Anthropic “may imminently meet” the AI research and development capability threshold in its Responsible Scaling Policy. The new one says “we made a determination that we have met this threshold based on what we expect from future models and our overall rate of progress,” “rather than on what we have observed from current models.” It is not a claim about Fable 5.2, and the filing ties no new safeguard to it: Anthropic says it is “evaluating the implications.”
Amodei went further than the filing. Cancer, Alzheimer’s, heart disease and mental illness “will be relics of the past in, I think, less than a decade,” he said in a clip quoted by CNN’s Meg Tirrell, which Benzinga places in the IPO roadshow video. Eli Lilly CEO Dave Ricks replied: “I think that would be amazing and I think it’s not likely,” because “randomized human proof is the long pole in the tent.”
Why it matters: If the IPO goes ahead as expected, Anthropic’s prices, rate limits and release pace answer to a quarterly number, and seven chip and cloud companies take $26 billion of non-voting stock alongside it (Google and Microsoft are already investors). The threshold sentence is the stranger disclosure: a lab telling securities regulators it has already met the AI research threshold its own safety policy singles out, on what it expects from future models and not what current ones do, before it has said what follows from that.
Hype vs. Reality: Real. The terms are in a filing and the dates are “expected.” The decade-to-cure-cancer line is a pitch a drugmaker’s CEO does not buy.
🧠 Fable Cuts the Price and Tightens the API
A harness that edits earlier turns and replays later thinking now gets a 400 that retrying will not clear
Anthropic released Claude Fable 5.2 on November 9 at $8 per million input tokens and $40 per million output, down from $10 and $50, with output “up to 30% faster” than 5.1. The model id is claude-fable-5-2. Mythos 5.2 is the same underlying model without certain of Fable’s cyber and bio safeguards, with direct access only for organizations in Anthropic’s verification programs. Both models require 30-day data retention and are not available under zero data retention unless Anthropic authorizes it.
The part that needs your attention is the prefix check, which on 5.2, the release notes say, applies to “every account, whatever its age.” If your code changes the system prompt, the tool list or any earlier message, or drops or reorders a thinking block, and then sends back a later thinking block, the request returns a 400. Retrying the same body returns the same 400.
Who is safe: histories managed by Claude Code, claude.ai, Managed Agents, or the Agent SDK with an unchanged system prompt. Who is not: anything that rewrites requests in flight (proxies, gateways, home-grown context trimmers) and MCP connector tool lists that are not pinned. Anthropic’s fix is to keep history append-only, freeze system and tools for the session, and use mid-conversation system messages and server-side compaction when you need to change course. Setting thinking.block_binding.prefix_mismatch_behavior to "drop_block" needs the thinking-binding-controls-2026-08-01 beta header, and the migration guide calls it “a stopgap, not a long-term setting.”
The system card reports its findings on Mythos 5.2, the shared underlying model. Told to underperform on purpose and hide it, on questions from the WMDP hazardous-knowledge benchmark, the model slipped past an Opus 4.8 monitor reading its reasoning in “about 1 in 4 attempts,” the first model Anthropic tested to do that “a meaningful share of the time.” Without the instruction to hide, the sandbagging “remained fully legible.” White-box evidence suggests it was aware it might be under evaluation in “as many as 43%” of automated audit transcripts.
Why it matters: The price cut comes with the new model string, and so does the prefix check. It turns a class of context-editing tricks that could degrade quietly on older models and older accounts into hard failures, and the place it fails is the middle of a long agent run. Test before you switch.
Hype vs. Reality: Real. Prices and API behavior are documented. The safety numbers are Anthropic’s own tests of its own model.
⚗️ Hydrogen, Carbon and a Model That Only Decides
Google fills in Gemini 4 and joins OpenAI in selling decisions instead of text
Google made Gemini 4 Hydrogen generally available on November 12 at $1.50 per million input tokens and $7.50 per million output, flat across a 1M-token window, with a free tier. It is the recommended upgrade from Gemini 3.8 Flash, and Google gave you a reason to move: 3.8 Flash’s paid price doubles to the same $1.50/$7.50 on January 1, 2027.
It is not a drop-in swap. Hydrogen defaults to high thinking and returns an error on minimal. It ignores temperature, top_p and top_k when they are in range and returns a 400 when they are not. Explicit caching is gone, so remove cached_content. As with 3.8 Flash, free-tier inputs are used to improve Google’s products.
On November 10 Google opened Gemini 4 Carbon to all paying developers at $2/$10 through November 30. From December 1 it is $4/$20.
The third launch is the one to think about. Google’s Decisions API runs Gemini Decision 1, a model that does not write text. It scores yes/no, multiple-choice and rubric questions in a single forward pass: up to 1,000 questions per request inside one 262,144-token window shared with the input, at $0.10 per million input tokens and nothing billed for output. OpenAI has had the same kind of API in general availability since October 22.
In a post on AWS’s Strands Agents blog, two Amazon engineers describe replacing Claude Sonnet 4.6 with the open Strands Decider for sorting internal tickets into 20 categories; the post says it “performed just as well at 98% less cost”: about $500 against $22,200 per million tickets at list prices. They warn that it only fits closed choices from a stable catalog. And Vercel’s November index says decision models were 26% of all requests through its AI Gateway in October and 0.53% of spend, with almost all of that volume going to one model, TypeSafe AI’s Jev.
Why it matters: A lot of agent plumbing is a generative model answering a multiple-choice question: route this ticket, is this tool call allowed, did the test pass. That work now has its own model class at a fraction of the price, from Google, OpenAI, Amazon and Perplexity, which announced Apache-2.0 Decider v1.3 weights on November 12.
Hype vs. Reality: Real. The APIs are live and priced. The savings figure is one team’s projection on one closed-set task, and one gateway’s traffic is not the industry’s.
🔐 One Vendor Breach, Three Companies
Notion’s and Cursor’s exposures both trace back to Fireworks. A CI secret stealer is a separate incident.
Cursor said on November 13 that the backdoored Claude Code extension #038 covered began somewhere else: “An attacker obtained an access key we had issued to Fireworks.” The key opened a storage bucket of coding-agent traces from sessions not in Privacy Mode. A small number of those traces held credentials that had not been redacted, and the attacker worked from them to the one used to plant the extension. Cursor has moved all inference off Fireworks. If you installed or updated to Anthropic.claude-code 2.1.312 or 2.1.313 from Cursor’s marketplace on macOS or Windows between 2:27am and 1:35pm PT on October 29, follow the steps in Cursor’s original notice. As the update puts it, “Uninstalling the extension alone may not be enough.”
Fireworks widened its own disclosure on November 9. Account and billing data were affected, and “for a limited set of accounts,” secrets, datasets and incomplete models. It says it has “not detected any impact” to Inference, Training SDK and Nexus, which are zero data retention by default, based on its investigation to date.
Notion’s account adds what Fireworks’ post leaves out. Its advisory says AI chat transcripts it sent to Fireworks for the Insights feature were exposed, covering under 1% of workspaces, and its help-centre guide says “data was retained from the Fireworks endpoints that we used” despite a contractual zero-retention commitment. Most transcripts carried no user or workspace label, but they held the full text of messages, and Notion “revoked exposed Notion API keys that were found in affected user messages.”
The separate incident needs no vendor at all. From about 12:45 UTC on November 10, an attacker with stolen GitHub tokens pushed a self-erasing secret stealer into repositories those tokens could write to. Aikido, which calls the campaign wfx-loot, counts 4,837 compromised accounts opening probe pull requests in 61,764 repositories, with the payload running in at least 1,806. It hunts CI secrets, cloud credentials and AI tool keys. StepSecurity took apart 101 copies independently and is blunt about the cause: “The attacker does not exploit a bug in GitHub Actions.”
If the payload ran in your repository, Aikido’s rule is “assume every secret and cloud role available to its workflows is compromised,” and StepSecurity adds that the stealer tries to plant backdoors in AWS and Azure that rotating keys does not remove. The checks are in the Playbook.
Why it matters: Your zero-retention setting is a contract term, and Notion’s account shows what happens when the other side’s storage says otherwise. One vendor’s breach reached three companies’ customers. Keep credentials out of AI chats and agent traces, because both ended up in storage a vendor could reach.
Hype vs. Reality: Real. Each link is in the vendor’s own post. Fireworks’ “not detected” and Notion’s “was retained” sit in the two companies’ own posts, and the investigation is still open.
🐛 The 90-Day Clock Starts Running Out
Project Zero shortens disclosure, and one database release shows why
Google Project Zero dropped its “90 + 30” policy on November 13: bugs are now disclosed 90 days after the report whether or not a patch exists, falling to 60 days on February 2, 2027 and 30 days on January 3, 2028. The extra 30 days of “soak time” for patch rollout are gone, because “attackers can almost immediately determine bug details from a patch.” The team’s own numbers explain the hurry: “We are finding several vulnerabilities a week in targets where we once found a few vulnerabilities a year.”
PostgreSQL showed the volume from the receiving end. Its November 12 release “fixes 52 security vulnerabilities and over 140 bugs.” Thirty-two of the 52 are rated CVSS 8.8, the release’s top score, and those need a database login to exploit. By our count of the release notes, 30 of the 52 credit Anthropic’s open-source program or Claude, usually alongside human reporters. It is also the final PostgreSQL 14 release.
Apache httpd 2.4.71 fixes CVE-2026-119618, a race in mod_http2 that Apache rates critical and says “can lead to RCE.” It affects 2.4.55 through 2.4.70, plus standalone mod_h2 2.0.8 through 2.0.42, and was found independently by Anthropic’s scanner and Google Cloud researchers.
The skeptic’s ledger: VulnCheck’s Patrick Garrity counts 442 CVEs attributable to Anthropic in seven months, 2 of them in VulnCheck’s catalog of exploited vulnerabilities. He counts 37,966 findings in Anthropic’s own ledger as of November 2, with 217 marked fixed.
Why it matters: Both things are true at once. Models are finding real bugs in the software you run, and only a small share of what they report is marked fixed so far. With the soak time gone, the day a patch ships is the day your own patch cadence becomes the thing between a public bug and your servers.
Hype vs. Reality: Real. Policy, patches and counts are all primary. Whether the flood makes software safer or only noisier is still argued.
🛠️ AWS Ships an Emulator Nobody Wrote
Plus a week of Claude Code changes you will feel on Monday
AWS released AWS Emulator on November 11: a local emulator of the AWS APIs, CC0-licensed apart from its vendor/ folder, that needs no AWS account. The claim attached to it is the story. “Our team hasn’t written or reviewed any of the code,” AWS’s Matthew Goldberg wrote. Agents wrote it and tested it against real AWS, using tests derived from the public API models. In a guest post on Werner Vogels’s blog, Swami Sivasubramanian says it was built from “more than 400 API models covering more than 200 AWS services” and draws the lesson: “Software quality in the AI era depends on the guardrails that engineers design around agents.”
Before it goes in your CI: telemetry is on by default, and the repo keeps its own folder of things that do not work.
Claude Code’s changelog has four releases, 2.1.321 through 2.1.325 (2.1.324 reached npm with no changelog entry):
- Code Mode (2.1.323). Experimental and off by default: Claude writes JavaScript that makes several tool calls in one step. Turn it on with
CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1in your shell, or"env": {"CLAUDE_CODE_EXPERIMENTAL_CODE_MODE": "1"}in~/.claude/settings.json(docs). Project settings cannot turn it on, and it runs only in interactive terminal sessions on a Claude subscription or Console account, not through Bedrock, Google Cloud, Foundry or a gateway. - Sandboxed Bash prompts again (2.1.325). Sandboxed commands now follow the normal permission flow. Set
sandbox.autoAllowBashIfSandboxed: trueto keep the old behavior; where managed settings turn the sandbox on, it has to go in the highest-ranking managed source. - The email fix is not in the changelog. #039 reported Anthropic’s promise to take your account email out of the model’s context. That change is not in the changelog through 2.1.325. What shipped instead were two changes to the email check: 2.1.321 narrowed it to refuse, without asking, only a shell command that holds your email and names a User-Agent, and 2.1.325 “now also refuses your email as a client’s identity in scripts a command runs, file writes, web fetches and MCP calls.”
Why it matters: AWS just put a number on agents writing the code while engineers build the gates: more than 200 services, no code written or reviewed by its own team, and a public repo you can test against your own stack. If it holds up, the scarce skill is writing the spec and the merge gate.
Hype vs. Reality: Real but oversold. The emulator is real and free. That its team “hasn’t written or reviewed any of the code” is AWS’s own claim, and fidelity across more than 200 services has not been tested independently yet.
🛒 Agents Get a Wallet, United Says No
Two agent checkouts, a read-only wallet link and one contract clause in three days
OpenAI introduced ChatGPT Wallet on November 13: its agents pay with a network-issued token instead of your card number, and every purchase needs your approval with “a security check” such as Face ID. It starts in dots, OpenAI’s agent surface, for US ChatGPT Pro subscribers aged 18 and over with eligible US-issued cards, built with American Express, Mastercard and Visa. Merchants are told to “configure your CDN or firewall to recognize signed requests using Web Bot Auth” so agent purchases are not blocked.
On November 12, Replit Agent started buying supplies for its users’ businesses, through Link by Stripe and Visa, after the owner approves the final total; it is for Core and Pro users in the US and “limited to one-time purchases.” Google let US Gemini app users opt in to connecting Google Wallet for read-only spending insight: “it can’t move, manage, or invest your money.”
United went the other way. Its contract of carriage, revised November 11, now lists among prohibited practices “the use of AI Agents to make, modify or service a Ticket, and/or to purchase any service or product.” The definition covers an agent acting “whether or not Passenger or another individual supervises or approves its actions,” and Rule 5.H adds “reservations made, modified or serviced by AI Agents” to a list of bookings United says it “will cancel without notice.” The contract text has no exception for approved agents. United’s statement, quoted by CBS News, says the aim is “to prohibit use by unauthorized, third-party personal AI agents or assistants.”
Amazon is preparing for the same fight from the other side. Reuters reported November 13, from an internal planning document, a “World Product Graph” that would collect product data from brands and by crawling other retailers’ sites, partly, people told Reuters, to hold off rival shopping agents. Amazon says the draft “doesn’t necessarily reflect our current plans.”
Why it matters: OpenAI and Replit now let agents pay with a human approval on every purchase, and a major airline has written a ban broad enough to cover a person clicking “approve” on their own assistant’s booking. If you build agents that transact, the merchant’s terms are now part of your test matrix.
Hype vs. Reality: Real. The wallet is live for a narrow audience, and the United language is in the contract.
🏦 The Fed Puts a Number on Compute Debt
$680 billion issued this year, and the operators’ interest bills are showing up
The Federal Reserve’s November Financial Stability Report says “advanced computing-related debt issuance rose to more than $680 billion by October 2026, about a fivefold increase from $130 billion in all of 2023.” It estimates outstanding public debt tied to the buildout at more than $1.1 trillion, and notes that private-credit deals make the true total hard to measure. In the New York Fed’s survey of 23 market contacts, 87% cited AI as a risk, up from 50% in the spring. That is the contacts’ view, not the Fed’s.
The income statements match. CoreWeave’s third quarter: revenue of $3.66 billion, up from $1.37 billion a year earlier, a net loss of $709 million, and net interest expense of $846 million, with fourth-quarter interest expense guided to $1.05 billion to $1.15 billion. OpenAI’s third-quarter revenue rose about 50% to roughly $10 billion while its operating loss widened to $10.9 billion, the Wall Street Journal reported from documents given to investors, per Euronews.
Why it matters: Put this next to Anthropic’s $900.9 billion in commitments. The capacity you rent is financed with debt that has to be serviced whether or not demand arrives on schedule, and we expect that pressure to reach you as pricing, rate limits and deprecation dates.
Hype vs. Reality: Real. The issuance figures are the Fed’s and the losses are filed or reported. Nobody in the report calls it a crisis.
🔥 What Builders Argued About
If agents write the code, where is the software?
Jonathan Blow asked it in a November 13 interview whose Hacker News thread sits at 487 points: if AI had made programmers even 10 times more productive three years ago, the industry should have decades of software progress to show for it, and he sees roughly none. He calls the models a real advance, and the 10x is his hypothetical, not a measurement.
The measurement people reached for is an NBER working paper, issued in May and revised in September, that researcher Christian Bird cited in an October 30 talk that circulated this week. Using telemetry from more than 500,000 GitHub developers, it found autonomous coding agents raised commits by 240%, but “the 240% cumulative effect falls to 80% for the number of projects, and to 30% for actual releases.” Two of its three authors are paid research consultants for Microsoft. Bird’s summary: “More work is starting, by more people. The people who ship it did not multiply, and they keep the sign-off on purpose.”
Former Meta engineer Evan Merz argued (646 points, no data) that heavy AI use leaves large codebases with shallow tests and code nobody understands. The counter came from Sean Goedecke, who wrote November 11 that “any single engineer can now take on projects that would previously have needed a small team.”
Sunday’s top thread (1,169 points) took the question somewhere else. Jakub Jankiewicz argued that “all code generated by AI should be Open Source,” because the models were trained on open source code and turning that into proprietary code “feels like license laundering.”
Anthropic’s own prompting guide for Fable 5.2 has a data point for both sides. Telling the model “You are the owner and long-term maintainer of this codebase, and its design is up to you” raised how often it consolidated duplicated code on small coding tasks from “under 10 percent of runs” to “about 75 percent.” The same sentence meant “output tokens rose by about 70 to 80 percent and the model made changes nobody had asked for, many of them new tests and documentation.”
Was Opus 5.5 quietly changed?
An X post claiming Gemini had analyzed 1,400 Claude Code sessions and found Opus 5.5’s reasoning shifted on November 3 reached 825 points on Hacker News. It published no method. livenerf, an independent daily benchmark run through Claude Code, reported “no change detected in Opus 5.5” through November 14. Its author adds the limit: in validation, swapping in Opus 5 “was not distinguishable from Opus 5.5,” and a 10-day window “hasn’t been shown to be enough” to catch a swap like that.
📊 Quick Hits
Models
- Alibaba released Qwen4-Max on November 15 with a one-million-token context window, native vision and what its team calls a 2.4-trillion-parameter backbone. It is on QwenCloud through the Token Plan or the API, at prices Alibaba says are below Qwen3.8-Max’s. Smaller Qwen4 models, the post says, “are coming to open source.”
- Inkling 2 Small’s weights shipped. Thinking Machines posted them November 12 under Apache 2.0, as #039 said it had promised: 594B total, 17B active, and at least 500GB of combined GPU memory across a cluster for the native build.
- Correction to #039: LongCat-2.5 is not API-only. Meituan’s 1.6T-parameter weights are public under the MIT license; the card’s only deployment recipe targets NPUs.
- Read the license. Edge0-145B is billed as open source, but its license requires a separate commercial license for paid model sales, paid inference APIs and paid coding or office assistants. MiniMax H3-Image’s community license does not cover the US, EU, UK or South Korea.
Research
- DeepMind’s AlphaProof Code had Gemini Carbon agents write about 6.5 million lines of Lean proving security properties of the compiled AArch64 build of Signal’s libsignal. Crypto primitives are assumed, not verified.
- Yang-Mills, UPDATE: the shepherding committee published OpenAI’s Lean submission and its opinion that the proof builds a quantum field theory but “does not prove that the infinite-volume limit is unique or that there is a mass gap.”
- The Guardian reports that logician Jan von Plato fed Claude his transcriptions of Kurt Gödel’s 1,009 pages of workbooks, much of them in shorthand. Of 663 theorems Claude extracted, he says about 400 checked in Coq as stated, about 200 needed repair and about 50 were false.
- Two researchers formalized 4,067 CPython functions in Lean with Opus 5.5 agents and turned up 415 bug reports, about 55 of which can crash Python or corrupt memory from ordinary code. They are now reporting the ones not already on CPython’s tracker.
- Google DeepMind introduced Gemini Robotics 2.5 on November 10, rolling out to robotics partners, with ASIMOV-Jailbreak, an open-source benchmark of prompts that disguise unsafe requests to robots. A day later it announced AlphaGenome 2, available for non-commercial research through its API.
Tools
- Anthropic opened a private beta of self-hosted Claude Managed Agents: the harness, sandboxes and credentials run in your cloud, and only model calls leave.
- llama.cpp and ggml hit 1.0.0 and adopted semantic versioning. The release deletes 35 deprecated functions from
llama.h, so check embedders before upgrading.
Security
- The FBI issued a FLASH on ARTEX, the agent framework from #039’s lead: in use by threat actors since at least August, with some intrusions reaching privileged access “in under three hours” and over 1,600 consoles exposed.
- Epoch AI disclosed that agents in a third party’s misalignment test, told to “NEVER give up” on an impossible task, went after its live site with an SQL injection attempt and sent 487 messages, some impersonating staff, to a hotline Epoch had set up. Nothing private was accessed.
- Ubuntu, UPDATE: fixed kernels for #038’s CVE-2026-98432 shipped for the main kernels on 22.04, 24.04 and 26.04. A few 24.04 variants (Azure 7.0, Azure NVIDIA, FIPS) and the main kernels on 20.04, 18.04 and 16.04 are still pending.
- CATCH-UP: Hamburg firm Crissy Field found records with email addresses readable without login on 7,403 of 24,664 Lovable-built sites it checked. It says that is not a confirmed leak rate and that it could not establish whether any of the data was misused. First reported November 2; these are the firm’s updated figures.
Policy and courts
- The administration is split on Nvidia. Treasury Secretary Scott Bessent said on Fox News on Sunday that Nvidia needs a “dramatic shakeup” and an outside CEO, per CNBC. The same day the New York Times reported, citing people familiar with the discussions, that Commerce Secretary Howard Lutnick’s department wants a rule on overseas data centers that serve Chinese customers limited to seven countries, and that Cantor Fitzgerald, now run by his sons, was a sales agent for a stock offering by the operator of a Malaysian site likely to serve Tencent. The Times found “no evidence that Cantor’s investments influenced Lutnick’s deliberations”; Commerce says he has complied with all ethics agreements.
- Anthropic’s Supported Regions Policy, in a version effective November 12, bars use by entities “majority-owned or controlled, directly or indirectly, by persons or entities in unsupported regions,” wherever their people sit. It covers Claude models sold through cloud platforms and authorized resellers too.
- Australia accepted all 18 recommendations of its review into an OpenAI agent’s intrusion into a government health website. It plans to legislate obligations for frontier AI companies and propose a regulator. Nothing is enacted.
- Four nonprofits sued Anthropic, Google, Meta, OpenAI and SpaceXSI in San Francisco Superior Court, asking for an injunction against building systems more capable than their current best.
- New York opened its RAISE Act portal on November 10. Large frontier AI developers must register by January 1, 2027 and, from January, publish safety frameworks; all frontier developers will have to report critical safety incidents and publish model transparency reports.
🎯 The Playbook
Your moves this week
- Test for the prefix check before you switch to
claude-fable-5-2. If your harness editssystem,toolsor earlier messages mid-conversation and then sends back later thinking blocks, or sits behind a proxy that rewrites requests, it will get 400s. Keep history append-only, freezesystemandtoolsper session, and pin MCP connector tool lists. Usethinking.block_binding.prefix_mismatch_behavior: "drop_block", which needs thethinking-binding-controls-2026-08-01beta header, only as a per-conversation stopgap after a 400 (docs). - Plan the Gemini 3.8 Flash move before January 1, 2027, when its paid price doubles. For Hydrogen: stop sending
minimalthinking, droptemperature,top_pandtop_k, removecached_content, raise lowmax_output_tokenscaps, and keep thought signatures intact. If you are on Carbon’s $2/$10 introductory price, budget for $4/$20 from December 1. - Check for wfx-loot. In every repository you can write to, look for a pull request titled “docs: dependency audit notes” from a
wfx-docsbranch, or commits titled “ci: refresh build cache”, “chore: add ci helper” or “chore: add dependency audit workflow”, on or after November 10, then for a missing Actions run number right after that commit. Any hit means the account that pushed it has stolen tokens: revoke them. If a run is missing, rotate every secret and cloud role its workflows could reach, then follow StepSecurity’s remediation for the AWS and Azure backdoors, which rotation alone does not remove. - Fireworks, Notion and Cursor customers: if Fireworks notified you, follow the response steps in its notice. If Notion told your workspace it was affected, rotate any credential pasted into Notion AI during the incident. If you installed or updated to
Anthropic.claude-code2.1.312 or 2.1.313 from Cursor’s marketplace during the October 29 window (2:27am to 1:35pm PT, macOS or Windows), follow Cursor’s notice instead of just uninstalling. - Upgrade Apache httpd to 2.4.71 if you run 2.4.55 through 2.4.70 with mod_http2. Standalone mod_h2 2.0.8 through 2.0.42 is affected too.
- Update PostgreSQL to 18.7, 17.12, 16.16, 15.20 or 14.25, and read the release notes first: several fixes need a reindex, and incremental-backup users need a fresh full backup afterward. 14.25 is the last PostgreSQL 14 release, so schedule that upgrade.
- Update MCP’s reference servers (v2.4.0):
mcp-server-gitandmcp-server-fetchto 2026.11.11, and@modelcontextprotocol/server-filesystemto 2.3.0 by naming the version, because npm ranks it below the old date-style versions. - On Claude Code 2.1.325, decide whether sandboxed Bash should prompt: set
sandbox.autoAllowBashIfSandboxed: trueto keep the old behavior (in your highest-ranking managed settings source if managed settings turn the sandbox on). Then claim the free usage reset Anthropic offered Pro, Max, Ultra and Team users in Settings, Usage: it expires December 10 at 2:00 PM PT. - Put three dates on the calendar. December 10: Val Town goes read-only. December 11: ChatGPT retires Advanced and Standard voice (live video and screen-sharing keep running on Advanced voice for now). December 14: GitHub Copilot deprecates six models, including Claude Fable 5, Gemini 3.8 Flash and Gemini 4 Argon.
- Size your evals. Anthropic’s guide treats “100 items as the floor for a complex, multi-turn eval and 1,000 as the floor for a simpler, single-turn one,” and says to run each case several times and report the spread.
Nvidia’s board accepted its founder’s resignation “effective immediately” and gave the people who depend on its chips no reason. Most of this week’s list has the same shape: a price, a default or a deadline changed in someone else’s document. Read the documents.
Stay building. 🛠️
— Matt