Wednesday, November 11, 2026

Anthropic’s IPO Enters a New Stage

Anthropic’s IPO Enters a New Stage

Today’s Overview

Good morning, Anthropic is moving its proposed IPO toward investors, while Claude is adding tools meant to keep spreadsheets and team planning inside its workspace. Meanwhile, researchers are finding ways to help AI agents check past lessons before acting. A lot is changing in how AI enters work and how autonomous systems learn. Let's dive in.

Top Stories

Anthropic Opens Its IPO Roadshow

Anthropic has opened the roadshow for a proposed public offering of 825 million Class A shares, with a price range of $70 to $78. It has applied to list on Nasdaq as ANT, and most shares are being offered by Anthropic itself.

  • Anthropic is offering 801,351,354 shares, while existing stockholders are selling 23,648,646.
  • The underwriters are expected to receive a 30-day option to purchase up to 123,750,000 additional shares.
  • The registration statement has been filed with the SEC but is not yet effective, so the shares cannot yet be sold.

Claude Adds Sheets and Whiteboards

Claude Sheets and Whiteboards are in beta for paid plans across web, desktop, and mobile, bringing collaborative spreadsheets and planning boards into Claude. Claude Dashboards is now Claude Data, generally available for data-connected dashboards and broader data storytelling.

  • Whiteboards can turn a finished plan into a document or deck, or continue the work in tools such as Lucid and Miro.
  • Sheets can be exported as XLSX or CSV when users need more spreadsheet functionality.
  • Claude Data’s live-sharing viewers can see current numbers without connector access, but can only view results from approved queries.

Gemini 4 Carbon Reaches Developers

Google has opened Gemini 4 Carbon to paying developers and enterprises through its API, while it rolls out to AI Ultra subscribers in Antigravity, the Gemini app, and Search’s AI Mode. Google says it leads three coding, research, and automation benchmarks at less than half the cost per task of the next-best external model.

  • Carbon’s listed price matches Gemini 4 Argon at $2 per million input tokens and $10 per million output tokens.
  • The release adds frontier cybersecurity capabilities through the Fairwind Program.
  • Antigravity is also previewing /swarm, a multi-agent mode.

Research & Analysis

Can Individual Neurons Explain Model Behavior?

Goodfire frames neuron-level analysis as a route to inspect language-model computation. It argues that researchers may not need to rely on sparse autoencoders in every interpretability study.

  • Goodfire calls its agents Neuron Scientist.
  • The agents generate candidate descriptions of neuron functions.
  • They also test those descriptions rather than leaving them as unverified labels.

dNaViT Builds One Tokenizer for Vision Tasks

dNaViT encodes images as discrete visual codes designed to support understanding, generation, editing, and reconstruction through a shared representation. Its authors report results matching or exceeding specialized baselines across a range of visual benchmarks.

  • At 512-by-512 resolution, it reports 28.12 dB PSNR and 0.81 SSIM with a spatial downsampling factor of 32.
  • Paired with Qwen3-8B, dNaViT scores 885 on OCRBench and rivals Qwen3-VL-8B on key document and reasoning benchmarks.
  • Its reported scores include 49.5 on GenEval2 and 52.2 on Qwen-Image-Bench, alongside strong bilingual text rendering.

WikiTrace Lets Agents Check Their Memories

On five agentic benchmarks, WikiTrace outperformed existing memory systems and a baseline given the same lessons without source links. It links each insight to a diagnosis of its source run and the raw trajectory, which agents can inspect when a lesson is incomplete, conflicting, or drawn from a failed task.

  • The authors traced 52% of memory-induced failures to an item the agent followed; 20% involved a wrong item.
  • Across eight benchmark suites, WikiTrace improved average performance by 10.7 points over using no memory.
  • Memory-related regression dropped from 6.5% to 4.5% with WikiTrace.

Weight Arithmetic Can Curb Code Reward Hacking

Researchers subtract weights from code models trained with loose and strict reward rules to isolate a direction associated with inefficient solutions. Extrapolating beyond the strict-reward model cut resource-limit failures from 14.5% to 8% without another training run, but too much extrapolation can produce malformed outputs.

  • The strict-reward model had already lowered resource-limit failures from 22.8% to 14.5% before weight extrapolation.
  • On harder-to-run problems, 71% of commonly solved programs ran faster after extrapolation.
  • Stricter rewards also brought a correctness tax: correctness failures rose on hard problems while overall solve rate barely changed.

Trending AI Tools

  • Muse Wonder-powered models Specialized models are rolling out to users of the Muse agent app and Meta AI.

  • Businesses in ChatGPT Companies can claim profiles, update details, share product feeds, and offer beta Business Agents to eligible US customers.

  • MAI-Transcribe-2 Speech recognition with diarization, configurable transcription styles, and word-level timestamps, priced at $0.10 per audio hour.

Quick Hits

  • OpenAI acquires Odessia , an AI travel agent, in what appears to be an acqui-hire; no travel-specific product plans have been announced.

  • Anthropic IPO plans remain on track despite AI-sector jitters, with trading potentially starting as soon as Thursday next week.

  • Arete-27B is an Apache 2.0 fine-tune of Qwen3.8-27B; on Nous Research’s 803-prompt evaluation, propaganda-tagged answers fell from 22% to 0.9% while general capability scores barely changed.

  • New York’s AI safety law has a regulator in place; OpenAI and Anthropic must register by January 1, and violations carry penalties.

  • Qwen4-Max is being previewed with an emphasis on stronger coding and agentic capabilities.

  • Bonds chat is multiplayer AI that ends when users and their friends fail to show up.

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.