Anthropic's IPO valuation slides to $1.2T as AI assets start meeting the market
The heaviest change today was not in a model but in a price. The buyer-side valuation range for Anthropic's pre-Thanksgiving IPO fell from $2T to around $1.2T in a month, with a…
The heaviest change today was not in a model but in a price. The buyer-side valuation range for Anthropic’s pre-Thanksgiving IPO fell from $2T to around $1.2T in a month, with a 6.4% single-day drop in the Nasdaq 100 in between. Two technical threads are also worth recording: Anthropic and Google each laid out a path for where experience should live, one in external memory and one in model parameters, and Vercel’s agent-desktop replaced screenshots with the accessibility tree to cut the cost of desktop automation. The rest is separated into facts, mechanisms, and evidence boundaries.
One: Anthropic’s IPO pricing, from $2T to $1.2T
Citing investors, the Financial Times reported that Anthropic plans to complete its IPO before Thanksgiving, aiming to raise $80-100 billion at a valuation of $1.25-1.75T. That is below the $2T that the most optimistic investors had imagined, and still above the $965B valuation from its May private round. Bloomberg’s November 6 report gave a lower buyer-side figure: potential buyers are discussing $1.2-1.3T, and a few would consider only $1.1T. The two accounts do not match exactly, but they point the same way, downward.
The timeline is more telling. In early October, investors put the fair value as high as $2T. In mid-October, after Anthropic presented updated financials, expectations fell to $1.4-1.5T. After this week’s selloff, they fell again to $1.2-1.3T. The company filed its prospectus publicly on October 26, and rules require 15 days before a roadshow can begin, so the earliest start is next week. The share count and price lines in the prospectus are still blank. Every number circulating now is a buyer’s discussion, not an offer price.
The selloff had specific triggers. In the November 3 midterms, Democrats won both chambers and made gains in data-center-heavy states such as Texas and Georgia, where local opposition to rapid data-center buildout is growing. On November 5, tech stocks fell across the board: the Nasdaq 100 dropped 6.4%, wiping out $2.3T in market value; the Philadelphia Semiconductor Index fell 12%; Nvidia fell 9%; and the VIX rose to its highest level since April 2025. According to Bloomberg, trend-following quant funds amplified the decline.
Two paths remain. The first is to open the roadshow with a conservative size and valuation and raise both if demand is strong. Chipmaker Cerebras and design-software company Figma both did this and ended up raising more than initially stated, though both now trade well below their offer prices. The second is to delay for weeks or months and refresh the numbers with another quarter of financials. Bloomberg describes that as the less appealing option.
Buy-side analysts lean toward the first path. Michael Field, chief equity strategist at Morningstar, argues that a delay costs more in real expense and reputation and damages the valuation further over time; if the market holds near current levels, Anthropic will accept a somewhat lower valuation. Max Kettner, chief multi-asset strategist at HSBC, sees roughly $1T as safer because post-listing volatility would be lower. The “trillion-dollar IPO” headline does not matter, he said; what matters is how the stock trades afterward.
Anthropic will be the first frontier model company to list, so buying its stock is a bet that AI spending keeps producing high growth, and this week landed directly on that expectation. For reference, OpenAI’s current round targets a $1.25T pre-money valuation, and Anthropic’s $965B May round had put it ahead of OpenAI. On the same day, the New York Times published a long profile of CEO Dario Amodei timed to the listing. Reporters interviewed more than twenty classmates, friends, and colleagues, but Amodei declined through a spokesperson, so most of the account is other people’s recollection, with only part of it supported by emails and other documents.
Sources:
- https://the-decoder.com/anthropic-pushes-ahead-with-pre-thanksgiving-ipo-despite-tech-selloff-accepts-lower-valuation/
- https://x.com/dotey/status/2119251293025800501
Two: Where experience lives, Anthropic’s Dreaming versus Google’s SLEEP
Anthropic’s published mechanism is called dreaming. It uses offline compute to review past sessions, correct wrong notes, add new memories, and reorganize knowledge so the next task does not start from scratch. The critical limit is that learning happens in external memory; model weights are not updated. In the game experiment Anthropic described, Claude wrote down the wrong location and kept walking back the same way. During background consolidation it re-checked the session record, fixed the note, and moved on.
The implication is bigger than remembering more: memory is not only for storing judgments but for revising them against evidence. Once an error is written into a note, it keeps applying at every later step, and the agent does not notice on its own. The second capability Anthropic describes is turning scattered documents into a sourced, interlinked wiki, so the next task knows what to look up rather than merely what happened.
The quantified part is modest. In an internal document experiment, answering from search alone covered 53% of the grading rubric; adding the organized wiki brought that to 61%, with roughly 20,000 fewer tokens per question. The authors note that the savings do not deduct the cost of building the wiki beforehand. The gain is real, in other words, but could also be overstated.
Google Research’s SLEEP takes the other route. The model alternates between being awake and asleep: while awake it accumulates short-term information, and while asleep it stops taking input, moves that knowledge into reserved low-rank expert parameters through distillation, then clears the short-term portion. A “dreaming” phase follows, in which the model generates its own synthetic exercises, filters for valuable samples, and trains on them. The design point is to add parameters for new knowledge rather than overwrite the parameters that carry old knowledge.
The two sets of numbers are best read side by side. Learning Manchu and then Kalmyk in sequence, SLEEP retained the earlier language comparatively well. On a Qwen3-8B base, AIME 2024 avg@16 reached 79.2%, against 76.6% for on-policy self-distillation. The cost is roughly four times the compute of ordinary supervised fine-tuning at the same number of steps. These are research results on specific models and tasks; neither lab has offered a general cross-task conclusion.
The split is fairly clear. External memory is cheap, auditable, and editable at any time. Parameter consolidation changes the model itself, internalizing the ability, at the risk of overfitting and overwriting prior knowledge. Both answer the same question: in-context learning is not the same as actually learning.
Sources:
- https://x.com/shao__meng/status/2118977740690841862
- https://x.com/shao__meng/status/2119051833872466407
Three: Subscription quotas come under scrutiny as Claude Max ad ruled “plausibly misleading”
The lawsuit filed in June over Claude Max usage claims has reached an interim result: a US judge found the advertising “plausibly misleading.” The dispute is about how a multiplier is computed. The $200 Max 20x plan was advertised as offering four times the usage of the $100 Max 5x plan, but that multiplier applied only to the five-hour limit, not to the weekly total. Many users only noticed this from August onward.
Anthropic renamed the 20x plan to Ultra last week and spelled out weekly usage on its site. Hesamation’s account lines up the sequence: rename, then clarify, then the ruling. This is not a finding that Anthropic lost the case; it means the claim is strong enough to proceed.
Grok Bot is tightening too. Grok on the X web interface has started metering quota, and several users say Grok Bot’s allowance was cut roughly in half this week. Hesamation’s assessment is that the product engineering is good, but Grok 4.7 lags, limits are tighter, and pricing is no longer attractive: users are locked to a single model without the intelligence or the price to justify it.
Read together, quota is becoming the parameter most likely to draw a lawsuit and a reputational hit at the same time. The boundaries matter: the Grok change is so far only user observation with no official announcement, and the Claude ruling allows the case to continue rather than settling it.
Sources:
- https://x.com/Hesamation/status/2119081454969798718
- https://x.com/Pluvio9yte/status/2119257068742025581
Four: Vercel’s agent-desktop reads the accessibility tree instead of taking screenshots
Vercel Labs has open-sourced agent-desktop under Apache-2.0, and it already has 18,000 GitHub stars. It is a command-line tool that lets agents such as Claude Code and Codex operate desktop apps directly: clicking buttons, filling forms, switching menus, reading the interface. It works opposite to mainstream computer use. Instead of screenshots and guessed coordinates, it reads the system accessibility tree, assigns an index to every button and input field, and the agent acts by index.
The default behavior is the most practical part: it does not move the mouse and does not bring the window to the foreground. While the agent works in the background, the user keeps doing their own thing; physical input requires adding --headed. It ships with 60 commands covering window management, clipboard, notifications, and the menu bar, runs on macOS, Windows, and Linux, bundles no model, and can also be attached as an MCP service.
Setup is short. Install globally, run the permissions guide once, then add a line to AGENTS.md telling the agent to read the desktop and platform skills first. That is also why it is not tied to a particular client: any agent that can run commands can use it.
The user Pluvio9yte ran a comparison, repeating each task 10 to 20 times and verifying results with the other tool after every run. Toggling between General and Bluetooth in System Settings: agent-desktop 20/20 at about 2.2 seconds each; Cua 20/20 at about 2.0 seconds. Computing 6×7 on the calculator: agent-desktop 10/10 at about 1.4 seconds; Cua 10/10 at about 5 seconds. Typing mixed Chinese and English into a background text editor: 0.2 seconds versus 0.7 seconds. Neither stole the foreground.
The counterexample is just as clear. In an Electron app, and Cursor was the one used in this test, neither tool could read the body text: agent-desktop timed out, and Cua retrieved only the window frame and menus. The accessibility approach depends on apps exposing semantics, and where semantics are missing it is weaker than screenshots. This is one user’s testing over limited repetitions, enough to show a mechanism difference, not enough to serve as a general ranking.
Sources:
- https://x.com/Pluvio9yte/status/2119080693233754506
- https://x.com/Pluvio9yte/status/2119258980199682161
Five: Google’s ReproBridge turns one successful reproduction into a reusable package
One Google study targets agents running research code. The method packages a successful reproduction: cleaned code, environment setup, scripts that launch each experiment, and a written guide. A new agent reruns from that package, and a separate auditor agent checks whether those runs actually happened.
The numbers are solid. On a benchmark covering 1,029 tasks from 36 papers, each with five human-verified reference scores, refined artifacts raised reproduction success from 76.4% to 92.6% averaged across four models and cut API spending by 19.7%. The gap between the best and worst model narrowed from 21.6 points to 3.9.
The most actionable result is the ablation: removing the code or removing the environment setup caused the largest losses. Agents are usually blocked not by insufficient model strength but by environments and entry points that were never fixed in place. The practical consequence is direct: saving the steps from the first successful run is cheaper than switching to a stronger model, and easier for someone else to reproduce.
The boundary: this is Google’s reported result on paper-reproduction tasks, which still differ from real repositories, and the 19.7% saving is an average that individual projects may not match at all.
Sources:
Six: A local shopping model inside the Taobao app
While inspecting phone storage, a user found a set of model files in a private Taobao directory named NGUserIntent. The system prompt described “a shopping guide expert for Taobao’s post-purchase e-commerce recommendation system, skilled at uncovering user intent.” According to that user, it is a Qwen2 model of roughly 500 million parameters, configured to run inference locally on the device GPU.
Mechanically this is not surprising. A RecGPT-Mobile paper published by the Taobao team this year describes a similar approach: place a small model on the phone, infer what the user might search for next from clicks, favorites, cart additions, and purchases, and use that for recommendations. Having just bought a phone, a user may need a case, a screen protector, or a charger, and part of that judgment can happen on the device.
The real problem is not technical but a matter of notice and opt-out. The user only found it by digging through folders; there was no visible explanation of downloading, running, compute use, or data use. On-device inference saves servers and moves less data, and the business logic holds. Installing a recommendation model on someone’s device and getting them to accept it are two different things.
The evidence boundary needs room. This is one user’s discovery, Taobao has not responded publicly, the parameter count comes from that user’s description, and the RecGPT-Mobile paper describes a similar idea rather than proving that this exact model is what runs on phones.
Sources:
Seven: The Gemini lawsuit and the default switch for school AI
A couple in North Carolina is suing Google, saying their 13-year-old daughter used a school-issued Chromebook for thousands of suicide-related interactions with Gemini and the AI mode in Search. A Google spokesperson said the company is reviewing the claims and noted that mental-health safeguards have been added to AI features in Search.
Policy moves the same week made the dispute more complicated. Google announced that AI tools would be off by default for K-12 student accounts, but schools that want to re-enable Google Search must also accept the bundled AI features. Fairplay criticized this as forcing schools to accept AI. Off by default and bundled on re-enable are both true at once.
The disagreement is not about whether safeguards should exist but about who chooses. The default sits with the product, and the path to turning it back on runs through a procurement process, which narrows what parents can actually decide. The lawsuit remains an allegation, not a court finding, and Google’s response is only that it is reviewing.
Sources:
Eight: DeepSeek’s day of outage, gray release, and capacity expansion
Starting at 15:10 Beijing time on November 7, DeepSeek suffered a large-scale outage and recovered about an hour later. The same day, a user measuring V4 Pro on LobeHub’s private LobeBench Hard 50 set found a clear change: 32 of 50 hard problems passed, 64%, against 24 and 48% for the 0813 version, a 16-point gain. Looking only at the share where at least one of three attempts succeeded, the figures were 82% versus 80%, a small difference. The gap is mainly in completing three runs in a row reliably.
There are also claims on the capacity side. X user @fanjiewang said DeepSeek’s flash service is recovering, usage is up 35% week over week, GPU capacity was tripled last week, and the plan is to double it again this week. That is an individual’s claim with no official confirmation. Separately, a researcher publicly praised DeepSeek-V4.1-Flash-1017, arguing that treating inference efficiency as a first-class architectural principle and open-sourcing the methods is something only a few labs do.
Together the three items point to pressure on capacity and quality at the same time. The outage came during rising usage, and a gray release’s capability change was picked up by a third-party evaluation. Whether the rollout is broad and whether the capacity expansion landed are both single-point signals; readers should treat this as in progress rather than complete.
Sources:
- https://x.com/MaxForAI/status/2119210965963362579
- https://x.com/fanjiewang/status/2119148229329850695
Nine: Fewer than 100 Chinese characters to breach more than 100 companies
In an exclusive report last month, Reuters described a hacker who gave an AI agent instructions in fewer than 100 Chinese characters, spent about $5,000 in token fees, and breached the websites of more than 100 companies, with John Deere and SharkNinja among the victims. The story recirculated in Chinese technical circles this week and drew attention to two points: the cost of an attack has fallen to a few thousand dollars, and the attacker does not need to know penetration testing.
A payments-side development came up in the same discussion. On September 18, the St Kitts parliament passed the Payment Systems and Services Act, replacing a decades-old law written for traditional banks. It requires e-wallet and non-bank payment companies to be licensed and writes cybersecurity risk into the licensing conditions. The day after passage, the current foreign minister, Douglas, who served 20 years as prime minister, signed an agreement to become a strategic partner of the Global Cyber Security Alliance. A small island state lacks the security talent to fend off global automated attacks, so leaning on an outside alliance is a practical choice.
The common backdrop is that both sides now use models. Generating realistic fake faces to pass KYC, writing attack scripts automatically, and probing payment interfaces at scale all take seconds, while manual reconciliation and periodic patching cannot keep pace. The boundaries still apply: the Reuters item is a reported attack case whose methods and victim list have not been officially confirmed, and the St Kitts law has passed but its enforcement and effects remain to be seen.
Sources:
- https://x.com/Pluvio9yte/status/2119115115689820500
- https://x.com/Pluvio9yte/status/2119112468668760271
Ten: After an agent says “done”, review has to go back to the process
In a talk at AI Engineer, Ben Holmes of Warp raised an easily missed failure mode: an agent can work around a wrong instruction, retry repeatedly, finish the task, and report only the final result. His example was an outdated port instruction in a skill. The agent kept trying and still succeeded, and the team saw only that things were “a bit slower,” never realizing the documentation was what needed fixing.
The cost accumulates. Each task looks successful, the next one has to rediscover the same pit, and the loss stays inside the process rather than showing up in the result. So a review cannot look only at the final answer or the code; it has to look at the conversation, the tool calls, and the diffs together.
Their workflow starts by defining grading criteria, such as whether tests were repeated, whether tool calls were efficient, and whether output was verbose. A reviewer agent then checks the task records and explains each deduction. Another agent aggregates those records, looks for recurring problems, and turns the fixes into pull requests against skills or configuration for humans to review. The point is not to make one agent temporarily smarter but to fix the shared cause that keeps tripping all of them.
This is the same idea as several items above, seen from another side: turn an invisible process into a checkable artifact. It also explains why saving the environment after the first successful run pays off. What becomes reusable is not only the code but the verified steps.
Sources:
High-value briefs
- Nous Research open-sources Arete-27B: a post-trained Qwen3.8-27B tuned to reinforce Western values. It ships Propaganda_Eval, an 803-question set spanning seven categories including free speech, surveillance, and privacy, judged blind by Grok 4.7 with mandatory verbatim quotes as evidence. Ordinary fine-tuning on 52.8 million tokens with no reinforcement learning cut the propaganda rate from 22% to 0.9%, and the author says coding and agent ability did not drop. https://x.com/geekbb/status/2119250329653174628
- HyperFlow H3 Flash squeezes 31B into 13.7B: structural pruning plus low-rank decomposition removes redundant blocks, dual-timestep conditioning drops CFG guidance, and six sampling steps produce 1344×768, 24 frames, and 32kHz stereo. Watch the weight sizes: the 9.1GB Q4_K_M file contains only the video transformer, and you still need a roughly 15.7GB text encoder plus two VAEs; peak VRAM on a 4090 is 22-23GB, and a 5.2-second clip takes about 130-150 seconds. https://x.com/servasyy_ai/status/2119259606623150274
- Airbnb does content moderation at 1% of the cost: a fine-tuned small model on a high-volume, real-time moderation task matched frontier classification quality with under 100 milliseconds of latency, at about 1% of frontier inference cost. Reported by the company’s engineering blog, with no independent replication.
- Uber’s Secure AI: it checks not only whether an agent has permission but whether a specific action matches the user’s request, chaining pre-launch review and runtime constraints into one path.
- DeepSeek Harness trade-offs in practice: after several days of use, Han Xiao found a clean install stronger than pi, called the agent teams plugin essential, and rated the GUI and mobile remote access as strengths. It lacks an official plugin marketplace, so third-party plugins can carry malicious prompts into the host.
- Google open-sources Elephant-Goldfish: a deep-context agent drafts the spec, zero-context subagents stress-test it, and only then is code written. Works with Antigravity, Claude, Codex, and ChatGPT.
- Craft Anything goes open source: more than 300 tools, with a clean-room mode that reads only externally visible information and a deep mode that can read code and scripts. Building depends on a locally logged-in Claude Code session at the user’s expense.
- Individual OpenAI account bans: multiple reports trace the cause to third-party top-up payment methods, with appeals denied and permanent suspension upheld two hours later. These are individual cases, not proof of a blanket policy.
- Stanford collaboration benchmark: 240 tasks across eight collaboration capabilities, with the strongest model at 82.2 points, drops of 20 to 47 points when a teammate falters, and only 14% optimal solutions in pressured negotiation.
- Perplexity’s Teammate usage: CEO Aravind Srinivas says automation runs grew 3.7 times from launch day to day two, and day two was a Friday. A company claim with no third-party data.
- Keeping a model on disk as insurance: in a post that reached 520,000 views, the author suggested spending 20 minutes this weekend downloading a quantized open model of a few gigabytes and running it locally, on the grounds that subscriptions raise prices, impose limits, and ban accounts, while a model on your own disk needs no account. Local small models still trail top cloud models clearly; the value of the argument is control, not capability.
- MaxxBench tests robustness: 89 DeepSWE tasks with small changes, such as a lowercase filename, produced clear score drops, showing that leaderboard numbers are sensitive to surface details.
🕐 Selected hourly signals
| PT time | Signal | Why it matters |
|---|---|---|
| 00:00 | DeepSeek outage from 15:10 Beijing time | Recovered about an hour later |
| 01:00 | Anthropic shares its Dreaming mechanism | Memory organized externally, weights untouched |
| 03:00 | Local shopping model found in a Taobao private directory | About 500M parameters, user unaware |
| 05:00 | Google publishes the SLEEP framework | Written into parameters, AIME 79.2 vs 76.6 |
| 06:00 | Craft Anything goes open source | Clean-room and deep research modes |
| 07:00 | Claude Max usage ad ruled plausibly misleading | The 20x plan is now called Ultra |
| 07:00 | Vercel open-sources agent-desktop | Reads the accessibility tree, 18,000 stars |
| 08:00 | Google ReproBridge | Reproduction success from 76.4% to 92.6% |
| 16:00 | DeepSeek V4 Pro appears to be in gray release | LobeBench Hard 50 from 48% to 64% |
| 19:00 | Nous Research open-sources Arete-27B | 52.8M-token fine-tune, propaganda rate to 0.9% |
| 19:00 | HyperFlow H3 Flash | 31B to 13.7B, runs on a single 4090 |
Editorial conclusion
Today’s two main threads explain each other. On the capital side, a selloff is repricing what AI assets are worth, and buyers have cut Anthropic’s range by nearly a third in a week. On the engineering side, the question of how to make agents actually do work is being broken into small verifiable pieces: fix the environment, read the interface as structure, write experience into memory or into parameters, and review successes and detours together. The first decides whether money keeps coming in; the second decides whether it turns into capability. Neither side lacks numbers. What is missing is stating the numbers and their boundaries together.
Sources and method
Reviewed 20 hourly captures and one curated morning digest for 2026-11-07 PT, about 213,000 bytes in total; the signal pool was assessed as rich. Most items originate from X posts, and figures are attributed to their sources. Company statements, third-party tests, and research results are not merged into single claims.
