AI Daily Digest · November 11, 2026

1. Anthropic Turns Claude Into an Office Suite, Hours Before Its IPO Roadshow

Anthropic used an event in San Francisco on Tuesday to push Claude well beyond chat. The company announced Claude Sheets, a collaborative spreadsheet in beta on paid plans across web, desktop and mobile, where teams describe what they need, Claude generates a working workbook with multiple tabs, formulas, formatting and dropdown menus, and then keeps editing cells alongside the team while annotating every change. Claude Whiteboards, also in beta, turns a brainstorm into an organized board with owners and deadlines, and can convert the result into a document or a slide deck. Both run on the same account as Claude Code and Claude Data, so work started in one place carries context into the next.

The most aggressive piece is Claude Meetings, entering beta on Team and Enterprise plans starting with Google Meet. Claude joins a video call as a visible participant, drafts the plan being discussed while people talk, answers questions in the meeting chat using connected work apps, and leaves automatically if an external guest joins unless an administrator allows it. Anthropic also made Claude Data generally available, the product formerly known as Claude Dashboards, which now connects to Snowflake, Salesforce, Amazon Redshift, Google BigQuery and ClickHouse with certified metric definitions, real-time sharing and version tracking. Interactive connectors were upgraded so Adobe, Atlassian, Box and Notion apps can run inside the Claude window, and the marketplace now lists more than 9,000 connectors and plugins, all built on the Model Context Protocol. On enterprise plans, each cloud session runs in its own isolated virtual machine with connector credentials kept outside it, and a Compliance API works with more than 100 security tools including Splunk and CrowdStrike.

The timing was not subtle. The launch landed the same day Anthropic kicked off its IPO roadshow, with 825 million Class A shares priced at 70 to 78 dollars under ticker ANT on the Nasdaq Global Select Market, a range that raises roughly 57.8 to 64.4 billion dollars. Advantage Solutions says Claude cut its per-cycle financial presentation work from about 10 hours to under 30 minutes, and Anthropic's own legal team cut marketing review from three days to 24 hours. Microsoft investors noticed: the stock slid more than 3 percent intraday after the announcement. Product lead Nicholas Lin was candid that Sheets is still basic and not aimed at complex financial modeling, which is the honest way to compete with Excel: start with the workflows AI actually does well and expand from there.
— Anthropic · Bloomberg

🔗 Anthropic · 财联社

2. OpenAI Opens Businesses in ChatGPT With Business Agents

OpenAI announced Businesses in ChatGPT in beta, giving companies a way to claim their profile, manage their information and see how people engage with them in the app, which now reaches 1.2 billion weekly users. Through Business Hub, a verified business can update key details, upload self-service product feeds with pricing and availability, and view aggregated insights including mentions, website clicks and discovery topics drawn from conversations where its site is cited. OpenAI states that user conversations and Business Agent chats are never shared with businesses, and insights only unlock above aggregate thresholds.

The agentic piece is Business Agents, an agent that represents a verified brand and talks directly with users. When a business appears in a model response, people can open its ChatGPT profile and choose to start a separate, clearly labeled conversation with its agent to compare products or ask follow-up questions, with the agent grounded in the business's own information. The first rollout covers eligible ecommerce and online service businesses in the United States, Shopify merchants can connect their store from Business Hub or the Shopify admin, and businesses can sponsor their Business Agent as the destination for ChatGPT Ads, with sponsored placements labeled and kept separate from ChatGPT's independent answers.

The strategic context is a pivot. OpenAI tried direct commerce last year with carts and commission splits, killed it in March after growth fell short, and is now building something closer to a search-and-discovery business model on top of ChatGPT instead of a storefront inside it. Vice president Benji Shomair said verification methods vary by market and business type and that OpenAI will not publish the specifics, an omission that will draw scrutiny since the whole product rests on users trusting that an agent really represents the brand it claims. The comparison business profiles imply is with Google's merchant ecosystem, and the advertising angle makes the stakes obvious.
— OpenAI · Reuters

🔗 OpenAI · Economic Times

3. Google Ships TPU 8tx and 8i Max, Four Chips in One Year

Google published two new configurations of its eighth-generation TPU on Monday, TPU 8tx and TPU 8i Max, and with them the family now covers four distinct jobs in a single year: 8t for scale-up training, 8i for low-latency reasoning, 8tx for confidential computing and 8i Max for massive context windows. Chip chief Amin Vahdat framed the philosophy plainly: rather than one chip trying to cover every market, specialized versions create far more value, and the agentic era is what forced the split.

TPU 8tx targets regulated enterprises and governments running frontier models next to sensitive data. It pairs the same silicon as TPU 8t with fifth-generation AMD EPYC CPU hosts and builds encryption into the hardware layer, so data and model weights stay protected and isolated even while being actively processed, avoiding the usual pattern where memory must expose plaintext during compute. Using the low-latency Boardfly network built for real-time reasoning, Google claims up to 2 times lower latency than TPU 8t with a near-zero performance penalty for encryption, and because the system carries training-grade compute, teams can fine-tune on private data, distill large models and train agents with reinforcement learning without moving that data between clusters. TPU 8i Max attacks the memory wall from the other side: two TPU 8i dies on one package deliver double the compute, double the SRAM and double the HBM bandwidth of a single 8i, and because the pair fits the existing rack footprint, data centers get roughly double the compute per square foot and per watt.

The software story matters as much as the silicon. Google open-sourced TorchTPU last month so PyTorch code and Hugging Face models run natively without rewrites, vLLM and SGLang have out-of-the-box TPU backends, and a new open-source TPU agent suite in preview plugs into Claude Code, Cursor and Google Antigravity to automate onboarding, diagnose bottlenecks and autotune kernels in plain language. Google says its native vLLM-TorchTPU stack set price-performance records on the InferenceX and AgentX benchmarks for agentic coding and long-context work. The market backdrop explains the urgency: Reuters reports Alphabet shares rose about 3 percent on the day, TPU capacity is effectively sold out, Bloomberg Intelligence projects TPU sales growing from roughly 12 billion dollars this year to over 135 billion by 2028 excluding cloud rentals, and Anthropic's S-1 discloses around 161 billion dollars in TPU-related lease commitments.
— Google · Reuters

🔗 Google

4. Gemini Robotics 2.5 Reaches 96 Percent on a Mercedes-Benz Production Line

Google DeepMind announced Gemini Robotics 2.5, its most capable vision-language-action model for robots outside the lab, and the demonstration that matters is not a benchmark video. Apptronik's Apollo 2 humanoid ran automotive kitting on active Mercedes-Benz production lines in Kecskemét, Hungary and Sindelfingen, Germany, picking metal brackets from bulk bins and arranging them in the exact sequence each vehicle required, with a reported 96 percent success rate. Two models split the work: Gemini Robotics ER tracks the workflow while the 2.5 actions model controls the hands and arms. DeepMind's robotics head Carolina Parada has been explicit about the trade-off her team is making: they push the generalization axis even though it costs perfection on any single application, and director Kanishka Rao called the remaining 4 percent gap a steep road.

On the lab side, the model was evaluated across more than 200 activities focused on dexterity, including serving food with tongs, pipetting liquids and tying a bow, which is the canonical nightmare for robots because ribbon deforms unpredictably. DeepMind reports 30 to 100 percent higher success rates than previous baselines depending on task type, with 20 to 30 percent faster completion. The same software spans hardware: Apollo 2's SharpaWave hand offers 22 degrees of freedom, while the Franka Duo platform runs on simple parallel grippers, so the model is not tied to one robot design.

Safety got its own release track, and this is where the announcement may age best. Gemini Robotics 2.5 uses inference-time reasoning guidance to steer around physical hazards, improving safe task completion around obstacles by 33 percent and around nearby people by 43 percent. DeepMind also open-sourced ASIMOV-Jailbreak on Hugging Face, a benchmark of 1,500 prompts that disguise unsafe requests across dozens of scenes, where Gemini 4 Argon achieved 99.5 percent jailbreak resistance versus 82 percent for the previous ER 2 baseline. Underneath all of it, Gemini 4 Argon acted as a simulation engine, generating thousands of varied practice runs from a single camera scene and building digital twins that let engineers diagnose failures in simulation before touching real hardware, which is a large part of why post-training improvements carried over to production with, in Apptronik's words, very little drop-off.
— Google DeepMind · Ars Technica

🔗 Google DeepMind · Ars Technica

5. TSMC's Board Approves Up to $1B Stakes in Anthropic and OpenAI

TSMC's board approved two investments on Tuesday that would have been unthinkable for a foundry a few years ago: up to 500 million dollars to buy Anthropic common shares at its initial public offering price, and up to 500 million dollars for preferred stock in OpenAI Group PBC. The resolutions appeared in the company's official board meeting announcement, alongside a separate approval of up to 500 million dollars to establish an investment company for strategic stakes across the semiconductor ecosystem. Reuters framed it accurately: the world's largest contract chipmaker wants exposure not just to the chips that run AI, but to the model developers that keep demand for those chips high.

The timing reads as a confidence signal placed deliberately inside a nervous market. Anthropic started its IPO roadshow the same day with a 70 to 78 dollar price range targeting as much as a 1.4 trillion dollar valuation, and is separately raising 26 billion dollars in a private placement with Google, NVIDIA, AMD and Microsoft named as backers. Meanwhile NVIDIA's stock has been swinging on governance drama and broader fears that AI capital spending is outrunning returns. TSMC itself just posted a sixth consecutive quarter of record profit, describes AI demand as extremely robust with orders still exceeding capacity, and yet has taken a cautious tone on 2027, so putting its own equity on the line is the strongest available statement that the demand it sees in orders is real.

The two tranches carry very different risk. The Anthropic purchase is anchored to a listed price range, so the exposure is transparent even if the deal only makes sense at pricing that holds; the OpenAI preferred stock is private, with no disclosed valuation or terms, which means the public record contains no way to mark it. Neither investment comes with a completion timeline. What TSMC gets, beyond upside, is demand visibility: reading its biggest customers' roadmaps from the inside lowers the risk of building too much leading-edge capacity just as the AI cycle turns, and that visibility is worth more than the 1 billion dollars it is spending to get it.
— TSMC · Reuters

🔗 TSMC · CNA

6. Gemini 4 Carbon Rolls Out to AI Ultra Subscribers

Google began rolling Gemini 4 Carbon out to AI Ultra subscribers on Tuesday across the Gemini app, AI Mode and Antigravity, moving last month's tightly scoped preview into a real product tier. Introductory API pricing is 2 dollars per million input tokens and 10 dollars per million output tokens through November 30, and Gemini 4 Argon remains available for AI Pro plans and above, so the two models now sit at deliberately separated price and capability rungs rather than replacing each other.

The interesting part is what Carbon's graduation says about Google's model strategy. Carbon started life as a restricted security-oriented variant, and its path from preview to paid tier in about a month is the same cadence OpenAI and Anthropic are running with their specialized models: ship a narrow, high-capability variant to a vetted audience first, watch how it behaves, then widen access once the safety data supports it. Third-party comparisons have been mixed in a useful way, reporting strong capability scores alongside a cost per task several times higher than cheaper alternatives, which puts the burden on Ultra-tier users to know when the extra capability pays for itself.

For developers the practical takeaway is the pricing window. The 2 and 10 dollar rates expire November 30, after which standard pricing applies, so anyone benchmarking Carbon against Argon, Claude or GPT-6 family models for agentic coding or long-context work should run those comparisons this month while the introductory rate makes large test runs affordable.
— Google · 9to5Google

🔗 Google · 9to5Google

7. ElevenLabs Ships Scribe v3 With Big Error Reductions and Live Diarization

ElevenLabs released Scribe v3 for recorded audio and Scribe v3 Live for realtime transcription. The company claims 27 percent fewer word errors than Scribe v2 across 102 languages, and for the realtime variant, 43 percent fewer errors than Scribe v2 Realtime. Scribe v3 Live also adds speaker diarization for the first time, meaning the live stream now separates who said what rather than producing an undifferentiated wall of text.

Speech-to-text is one of those layers where improvements compound quietly across an entire product stack. Agent voice pipelines, customer support transcription, meeting assistants and accessibility tools all sit on top of an ASR model, and a 27 to 43 percent error reduction changes what downstream logic can safely assume: fewer correction loops, fewer misrouted support tickets, more reliable command parsing for voice agents that act rather than just transcribe. Realtime diarization specifically unlocks multi-party use cases, since an agent that joins a sales call or a standup needs to know which participant made a commitment, not just what words were spoken.

ElevenLabs built its reputation on voice generation and has been expanding into the perception side, and Scribe v3 is the clearest sign it wants to own the full loop of hearing and speaking for agents. Competing against Whisper-class open models and hyperscaler ASR on price is hard, but the differentiator on display here is accuracy in live conditions, which is exactly the metric that matters once a transcript drives downstream actions instead of sitting in an archive.
— ElevenLabs

🔗 ElevenLabs

KD Agentic · AI Daily Digest