- AI
- Business
- Hardware
- Voice
TAG44AI Daily Digest · November 11, 2026
1. Anthropic Expands Claude into an Office Suite (Announcement on the First Day of the IPO Roadshow)
Anthropic announced on November 10 at an event in San Francisco a suite of tools that expands Claude from a chat-based assistant into a comprehensive office environment. First, Claude Sheets is entering beta for paid plans, with support for web, desktop, and mobile. Teams simply describe what they need, and Claude generates a spreadsheet containing multiple sheets, formulas, formatting, and drop-down menus; from then on, it can continue editing cells directly alongside the team and clearly indicate the changes it makes. Claude Whiteboards is also in beta, handling team brainstorming and turning it into organized boards with assignees and deadlines, and can convert the results into documents or slide presentations.
The most aggressive offering is Claude Meetings, which is initially available in beta for Team and Enterprise plans and supports Google Meet. Claude joins video calls as a visible participant, drafts an agenda plan as the discussion progresses, retrieves information from connected business apps, and answers questions in the meeting chat. If an external guest joins, Claude automatically leaves unless an administrator has granted prior permission. Meanwhile, the former Claude Dashboards has been renamed Claude Data and is now generally available (GA). It connects to Snowflake, Salesforce, Amazon Redshift, Google BigQuery, and ClickHouse, and provides data-team-approved metric definitions, real-time sharing, and version tracking. Interactive connectors have also been enhanced, allowing Adobe, Atlassian, Box, and Notion apps to be operated directly within Claude, while the marketplace now features more than 9,000 connectors and plugins based on the Model Context Protocol. Enterprise plans run each cloud session in an isolated VM, keep connector credentials outside the VM, and also provide a Compliance API that integrates with more than 100 security tools, including Splunk and CrowdStrike.
The timing of the announcement is unmistakable. Earlier that morning, Anthropic began its IPO roadshow, revealing its plan to list 825 million Class A shares on the Nasdaq Global Select Market under the ticker symbol ANT at $70–78 per share.At this price point, the scale of procurement amounts to approximately $57.8 billion to $64.4 billion. In customer examples, Advantage Solutions reduced the time required to prepare financial materials for each cycle from about 10 hours to within 30 minutes, while Anthropic’s own legal team compressed marketing reviews from three days to 24 hours. Following the announcement, Microsoft’s stock price briefly fell by more than 3%. Nicholas Lin, the product lead, candidly acknowledged that Sheets is still a basic feature and is not intended for complex financial modeling—an honest stance in competing with Excel. The strategy is to start with workflows where AI is genuinely strong and expand from there.
— Anthropic · Bloomberg
2. OpenAI launches “Businesses” and “Business Agents” in ChatGPT
OpenAI has begun beta testing Businesses in ChatGPT, enabling companies to authenticate their own profiles, manage information, and understand how users are engaging with them on ChatGPT, which has 1.2 billion weekly active users. In Business Hub, verified companies can update basic information, register product feeds—including prices and inventory—on a self-service basis, and view aggregated insights such as mention counts, site clicks, and discovery topics extracted from conversations. OpenAI’s policy is not to share users’ conversations or Business Agent chat content with companies, and insights are displayed only when they exceed a certain aggregation threshold.
The agentic component comes in the form of Business Agents. Agents representing verified brands converse directly with users. When a company appears in a model’s response, users can open that company’s ChatGPT profile and ask the agent questions in a separate, clearly labeled conversation. They can receive answers based on the company’s own information, such as product comparisons and specification checks. The initial rollout is limited to U.S. e-commerce and online service companies that meet certain criteria. Shopify merchants can connect their stores through Business Hub or the Shopify admin dashboard, and items covered by the Shopify Catalog have already been integrated into ChatGPT. Companies can sponsor Business Agents and use them as landing destinations for ChatGPT Ads; sponsored placements are clearly labeled and kept separate from ChatGPT’s own responses.
The strategic context is also important. Last year, OpenAI ventured into direct e-commerce with a cart feature and a commission-based model, but ended it in March of this year because of sluggish growth. The current direction marks a pivot from a “mall inside ChatGPT” to a “search-and-discovery business on ChatGPT.” Benji Shomair, the VP of business development, explained that the method of business verification varies by market and industry and that details would not be disclosed. Since the entire product is underpinned by trust in the question of whether an agent really is from that brand, this lack of transparency is likely to draw attention. The point of comparison is Google’s business-profile ecosystem, and the magnitude of the stakes is clear given the inclusion of advertising.
— OpenAI · Reuters
🔗 OpenAI · Economic Times
3. Google announces TPU 8tx and 8i Max (four-chip lineup within one year)
Google announced two new configurations of the eighth-generation TPU—TPU 8tx and TPU 8i Max—on its official blog on November 10. This brings the eighth generation to a four-system lineup in a single year. The 8t is for scale-up training, the 8i for low-latency inference, the 8tx for confidential computing, and the 8i Max for massive context windows. Amin Vahdat, head of the chip business, clearly stated the policy: “Rather than trying to target the entire market with a single chip, dedicated versions create far greater value,” adding that the requirements of the agent era prompted this split.
TPU 8tx is intended for regulated industries and government agencies that run cutting-edge models alongside sensitive data. It pairs the same silicon as the TPU 8t with a fifth-generation AMD EPYC CPU host and integrates encryption at the hardware layer. Data is normally encrypted at rest and in transit, but becomes plaintext in memory while being processed; the 8tx keeps data in use and model weights protected and isolated. It adopts the low-latency Boardfly network, developed for real-time inference, and is said to reduce latency by up to two times compared with the TPU 8t, with virtually no performance degradation from encryption. It also has training-grade compute, enabling fine-tuning with private data, large-model distillation, and agent training through reinforcement learning on the same system, without moving sensitive data to a separate cluster. The TPU 8i Max, meanwhile, is an answer to the memory wall. It houses two TPU 8i chips in the same package, doubling compute, SRAM, and HBM bandwidth. Because twice as much silicon fits within the footprint of existing racks, data centers can nearly double their compute density per square meter and watt.
Software support is also advancing. TorchTPU, which was open-sourced last month, allows PyTorch code and Hugging Face models to run on TPUs without modification, while vLLM and SGLang include TPU backends as standard. The open-source TPU agent suite, currently in preview, connects to Claude Code, Cursor, and Google Antigravity to automate onboarding, diagnose bottlenecks, and automatically tune kernels using natural language. According to Google, its proprietary vLLM-TorchTPU stack recorded the best price-performance ratio on the InferenceX and AgentX benchmarks, which measure agentic coding and long-context processing. Market conditions are also favorable. Reuters reported that Alphabet shares rose by about 3% on the announcement date and that TPU supply capacity is effectively sold out. Bloomberg Intelligence estimates that TPU sales will grow from approximately $12 billion this year to more than $135 billion in 2028, excluding cloud-based rentals. In addition, Anthropic’s S-1 lists approximately $161 billion in TPU-related lease-payment commitments.
— Google · Reuters
4. Gemini Robotics 2.5 achieves a 96% success rate on Mercedes-Benz’s real production line
Google DeepMind announced Gemini Robotics 2.5. It is the company’s most powerful VLA (Vision-Language-Action) model to date for robots operating outside the lab, and its notable achievement is not benchmark videos but a deployment track record with actual machines.Apptronik’s humanoid Apollo 2 handled automotive kitting on live Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany, removing metal brackets from bulk bins and arranging them in trays in the order required for each vehicle with a 96% success rate. The division of labor uses a two-model setup: Gemini Robotics ER tracks the workflow, while the 2.5 action model controls the hands and arms. Carolina Parada, who leads robotics at DeepMind, stated plainly, “Specializing in a particular use case can yield high precision, but at the cost of generalization. We will keep pushing on generalization,” while director Kanishka Rao calls the remaining 4% gap a “difficult road.”
In lab evaluations, more than 200 tasks were conducted, focusing on dexterity. These included serving with tongs, pipetting, and tying ribbon bows, with deformable objects in particular being a classic challenge for robots. According to DeepMind, success rates improved by 30–100% over the previous generation depending on the task type, while completion speed improved by 20–30%. The software is hardware-agnostic. Apollo 2’s SharpaWave hand has 22 degrees of freedom, while the same software also runs on the Franka Duo platform with its parallel gripper, making it independent of any particular robot morphology.
Safety is being handled as an independent release track, and this may be the most important part in the long term. Gemini Robotics 2.5 avoids physical hazards through inference-time guidance, improving successful completion of safety tasks by 33% around obstacles and by 43% when people are nearby. DeepMind also released on Hugging Face the open-source benchmark ASIMOV-Jailbreak, which contains 1,500 prompts that disguise unsafe requests across dozens of scenarios; Gemini 4 Argon demonstrated 99.5% robustness against them (the previous generation’s ER 2 scored 82%). At the foundation, Gemini 4 Argon serves as a simulation engine, generating thousands of diverse practice runs from a single camera feed to build a digital twin and diagnose the causes of failures in simulation before the real hardware is touched. According to Apptronik, post-training improvements carried over into production with “almost no degradation.” Agile Robots, Apptronik, and Mercedes-Benz are the initial partners.
— Google DeepMind · Ars Technica
🔗 Google DeepMind · Ars Technica
5. TSMC Board Approves Up to $1 Billion in Investments in Anthropic and OpenAI
On November 10, TSMC’s board of directors approved two investments that would have been unthinkable for a foundry just a few years ago: acquiring up to $500 million worth of Anthropic common stock at its IPO offering price and acquiring up to $500 million worth of preferred stock in OpenAI Group PBC. The two resolutions are listed in the company’s official announcement of board resolutions, which also approved establishing an investment company of up to $500 million for strategic investments in the semiconductor industry and its ecosystem. Reuters’ summary is apt: the world’s largest contract manufacturer is seeking exposure not only to the chips that power AI, but also to the model developers that sustain demand for those chips.
The timing can be read as a vote of confidence in an uneasy market.Anthropic began an IPO roadshow on the same day, targeting a valuation of up to approximately $1.4 trillion at a price range of $70–$78 per share, while also pursuing a $26 billion private placement backed by Google, NVIDIA, AMD, and Microsoft. Meanwhile, NVIDIA shares have been highly volatile amid governance concerns and uncertainty over the outlook for AI spending, fueling warnings that AI capital expenditures may be outstripping the returns they generate. TSMC itself has posted record earnings for six consecutive quarters through last month. It describes AI demand as “extremely robust, with supply capacity falling short of demand,” while taking a cautious outlook for 2027. Putting its own capital on the line through shareholder exposure is arguably the strongest affirmation that the demand visible in order data is real.
The two positions have very different risk profiles. The Anthropic position is tied to the IPO price range, so the exposure is transparent assuming the price holds. OpenAI’s preferred shares are privately held, and neither their valuation nor their terms have been disclosed, making them impossible to value from public information. The completion date for neither has been disclosed. What TSMC gains is more than upside potential—namely, visibility into demand. Being able to read its largest customer’s roadmap from the inside directly reduces the risk of overbuilding cutting-edge capacity at an inflection point in the AI cycle. The value of that visibility should exceed the $1 billion spent to acquire it.
— TSMC · Reuters
6. Gemini 4 Carbon begins general availability for AI Ultra
On November 11, Google made Gemini 4 Carbon generally available to AI Ultra subscribers. It is available in the Gemini app, AI Mode, and Antigravity, marking the promotion of a preview previously offered exclusively last month into an official product tier. Introductory API pricing of $2 per 1 million input tokens and $10 per 1 million output tokens applies through November 30. Gemini 4 Argon remains available to AI Pro subscribers and above, with the two models deliberately separated into price and capability tiers rather than serving as replacements for each other.
What stands out is the model strategy indicated by Carbon’s promotion. Carbon began as a security-focused, limited variant and reached paid tiers roughly one month after its preview. This is the same cadence OpenAI and Anthropic follow with specialized models: first ship a narrow, high-performance variant to vetted users, observe its behavior, and broaden availability once safety data supports doing so. Third-party comparisons offer a mix of useful praise and criticism, with high capability scores but a reported cost per task several times that of cheaper alternatives. Ultra-tier users are responsible for identifying the situations in which the additional cost can be recouped.
For developers, the practical implication is the duration of the introductory pricing. The $2/$10 rate ends on November 30, after which standard pricing will apply. Developers looking to compare Carbon with Argon, Claude, and the GPT-6 family would be well advised to run large-scale tests this month, while introductory rates make doing so inexpensive.
— Google · 9to5Google
🔗 Google · 9to5Google
7.ElevenLabs Releases Scribe v3 and Scribe v3 Live
ElevenLabs has released Scribe v3 for prerecorded audio and Scribe v3 Live for real-time transcription. According to the company, Scribe v3 reduces word error rates across 102 languages by 27% compared with Scribe v2, while the real-time version, Scribe v3 Live, achieves a 43% reduction in errors compared with Scribe v2 Realtime. Scribe v3 Live also introduces speaker diarization for the first time, making it possible to distinguish “who said what” from live audio.
Speech recognition is a layer whose effects quietly ripple across an entire product stack. Agents’ voice pipelines, customer support transcription, meeting assistants, and accessibility tools all sit on top of ASR models, and a 27–43% reduction in errors changes the range that downstream logic can safely assume. Examples include fewer correction loops, fewer misrouted support tickets, and improved reliability in command interpretation by voice agents that take action rather than merely transcribe speech. Real-time speaker diarization, in particular, unlocks use cases involving multiple participants. Agents taking part in sales calls or stand-ups need to know not only the words spoken, but also which participant made a commitment.
ElevenLabs has built its reputation in speech generation and has been expanding into perception, and Scribe v3 is the clearest expression of its ambition to control the entire “listening” and “speaking” loop for agents. It is difficult to compete on price with open Whisper-based models and hyperscaler ASR offerings, but the differentiator demonstrated here is accuracy under live conditions—a metric that matters at the stage where transcription drives downstream actions rather than disappearing into an archive.
— ElevenLabs
KD Agentic · AI Daily Digest November 11, 2026
