Today’s through-line runs from a flagship model’s general release, to a research lab stepping toward the public markets, to a consumer chat product opening a path for business discovery, to a robotics model being tested on a factory line, and to a new set of open image-model weights: Google on November 10 made Gemini 4 Carbon, its strongest model to date, available to developers, enterprises, researchers and consumers. Anthropic the same day launched the roadshow for its proposed initial public offering, set an expected price range of $70.00 to $78.00 a share, and applied to list its Class A common stock on the Nasdaq Global Select Market under the symbol “ANT.” OpenAI introduced Businesses in ChatGPT in beta, allowing companies to claim a profile, view aggregated insights, and, for eligible U.S. ecommerce and online-service businesses, launch a Business Agent. Google DeepMind also unveiled Gemini Robotics 2.5, its most capable vision-language-action model, and reported a 96% success rate in a Mercedes-Benz production-line proof of concept. MiniMax released and open-weighted H3-Image, an image generation and editing model built on H3, saying it can produce a 4K image in under 10 seconds.

1. Google makes Gemini 4 Carbon, its strongest model to date, generally available

On November 10, 2026, Koray Kavukcuoglu, senior vice president of Google DeepMind and Google’s chief AI architect, wrote on the company blog that Gemini 4 Carbon is now available to developers, enterprises, scientists and consumers. Google called Carbon its strongest model, with frontier performance in cybersecurity and in long-horizon agentic workflows across complex knowledge work, software engineering and scientific reasoning.

Two weeks earlier, Google had released Carbon to a set of cybersecurity defenders through its Fairwind Program. The company wrote that, after the recent launch of Gemini 4 Argon, Carbon pushes the frontier further. Argon remains the preferred choice for most software engineering and knowledge work, while Carbon is built for tasks that test the limits of a model’s intelligence, such as complex, multi-step coding and debugging and open-ended scientific research and discovery, using more tokens than Argon to explore thoroughly, reason deeply and run extended verification loops.

On independent benchmarks designed to test performance in real workflows, Google said Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench, at more than two times lower cost per task than the next best external model. The company also partnered with Artificial Analysis to launch a Multimodal Knowledge Agent (MKA) benchmark, arguing that existing tests do not adequately capture how agents reason natively over dense files that contain charts, images, blueprints, schematics and tables; Carbon ranks second on that benchmark while costing more than six times less per task, on average, than the top-performing model. In prerelease evaluations, Google said Carbon matched the best model tested on Hebbia’s end-to-end financial agentic evaluation and ranked first among 23 frontier-model configurations on Balyasny Asset Management’s General Financial Intelligence benchmark.

To put Carbon to work on larger and more complex projects, Google is previewing /swarm in Google Antigravity, turning a single prompt into a coordinated team of agents that can be tracked in real time. Developers can assign Carbon to roles that demand the deepest reasoning, such as architectural design, coordination and auditing, while pairing it with faster, lighter models in parallel isolated workspaces and with adversarial verifiers that test every change. In a post the same day, the Antigravity team wrote that /swarm hands a prompt to a family of agents: a coordinator spins up specialized parallel workers in isolated git worktrees so edits do not collide, and adversarial verifiers run tests and challenge diffs; the agents collaborate on a shared, append-only message board and park at zero token cost when idle.

On availability, Gemini 4 Carbon is offered through the Gemini API in Google AI Studio and Gemini for Business to all paying developers and enterprise customers, at the same introductory price as Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens. The introductory price runs through November 30, 2026, after which the rate becomes $4 per million input tokens and $20 per million output tokens, and applies only to the Carbon model intended for public access. When Google first announced the model on October 28, it also said cached input tokens would be priced at a 95% discount to the input rate. Carbon is also rolling out to all Google AI Ultra subscribers in Google Antigravity, the Gemini app and AI Mode in Search. The model card states that Carbon is based on Gemini 4 Argon, that inputs may include text, images, audio and video, that the combined input-output context window reaches up to about 1 million tokens, and that the output is text.

On cybersecurity, Gemini 4 Carbon for cyber, accessed through the Fairwind Program, is described as Google’s strongest cybersecurity capability to date, able to inspect software, trace how flaws can be chained across systems, and generate verified root-cause patches. Google said the model leads on key benchmarks for finding (SEC-bench Pro), exploiting (ExploitGym) and patching (CWE-bench v1) vulnerabilities; scores the company published on October 28 were 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. Early internal and external tests cited by the company include: on an internal benchmark for Project Fortify, Google’s autonomous system for scanning, triage and remediation, Carbon reached 100% recall on scanning and 95% precision on triage, and in tests on open-source repositories improved by more than 25% on Argon’s already high recall; Google’s Cloud Vulnerability Research team used Carbon to show that two bugs in a popular open-source package were in fact exploitable, comparable in severity to the recent PHP2Shell exploit, and, by building a proof of concept, helped maintainers reclassify the issues as critical and issue a patch, with details to follow after responsible disclosure; and in Wiz’s Scan for Good program, the model was used to scan two large software-as-a-service providers, uncovering zero-day vulnerabilities in under two hours, after which the team worked with the affected organizations on remediation.

On science and mathematics, Google said researchers built an agentic formal-verification framework called AlphaProof Code, used Carbon alongside other models to translate security goals in the code underlying the Signal cryptographic protocol into mathematical theorems, and had agents generate millions of lines of Lean code, helping prove key security properties in weeks—work the company said has historically taken expert teams years. In protein-structure prediction, researchers used a multi-agent setup to have Carbon search for remote homologs across petabase-scale unannotated sequence archives, environmental and viral sequence catalogs, and three-dimensional structure databases; across seven low-homolog protein families in the CASP17 blind test set, Carbon increased the number of homologs retrieved by about 40 times on average, without retraining structure-prediction models. Four weeks earlier, Google had opened applications to give 10,000 scientists six months of free access to Argon in Antigravity; the company said thousands have already joined, and it is now extending Carbon to current and future participants, with applications still open and onboarding continuing on a weekly rolling basis.

On safety, Google said that in simulated cyberattack environments seeded with ExploitBench and CyScenarioBench, Carbon blocked 99% of attack trajectories from an adaptive red-teaming agent, lowering the per-attempt breach rate to 0.8%, compared with 3.1% for Argon; in Gray Swan testing, the success rate of automated cyberattack queries fell to 2.17%, compared with 9.86% for Argon; and external testing by Trajectory Labs found no new universal jailbreaks across 132 hours and more than 13,000 requests. On a 375-task agentic honesty benchmark, Carbon passed 92% of tasks, compared with 70% for Argon, declining to fill in missing experimental data 98% of the time and flagging methodological flaws 68% of the time. The company had said two weeks earlier that Carbon reached a 0% attack-success rate on the Gray Swan IPI benchmark for indirect prompt injection, while also noting that, in some distributions, the model still has room to improve against more sophisticated attacks.

Why it matters: Google is moving Carbon from a limited cybersecurity preview into the paid API and Ultra subscribers, but it is still leaving everyday work to Argon and spending the extra reasoning tokens on longer verification loops; the lower per-task costs and stronger safety figures are meant to show that this is not only a chase for peak benchmark scores.

2. Anthropic launches its IPO roadshow at $70 to $78 a share

On November 10, 2026, Anthropic, PBC said it had launched the roadshow for its proposed initial public offering of Class A common stock. The company said the offering consists of 825,000,000 Class A shares, of which 801,351,354 are being offered by Anthropic and 23,648,646 by certain existing stockholders; Anthropic will not receive proceeds from the selling stockholders’ shares. It also expects to grant the underwriters a 30-day option to buy up to an additional 123,750,000 Class A shares. The offering price is expected to be between $70.00 and $78.00 a share. Anthropic has applied to list the Class A shares on the Nasdaq Global Select Market under the symbol “ANT,” and stressed that the registration statement has not yet become effective and that the announcement is not an offer to sell or a solicitation of an offer to buy the securities.

At that price range, the 825 million shares imply an offering size of about $57.75 billion to $64.35 billion. TechCrunch reported that, at the top of the range and on a fully diluted basis, the company would be valued at about $1.404 trillion, above the $965 billion valuation reached in a May fundraising but below a target of about $2 trillion reported earlier. CNBC put the fully diluted figure at up to about $1.405 trillion and, citing people familiar with the matter, said the company plans to price on November 18 and begin trading on November 19, with a retail allocation of about 20%.

Both outlets, citing the updated filing, also reported that Anthropic has agreed to sell $26 billion of non-voting Class N stock in a private placement at the IPO price to AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC; TechCrunch said Google and Nvidia are investing $10 billion each, and that Anthropic estimates net proceeds of about $84.6 billion to $93.7 billion, the upper figure including the underwriters’ option. The filing, as reported, put run-rate revenue at about $81 billion at the end of October, up about $10 billion from the end of September. CNBC said revenue for the nine months ended September 30 was $33.24 billion, with a net loss of $104 billion in the same period, most of it from changes in the fair value of convertible notes and derivative liabilities. TechCrunch also noted that, as a public-benefit corporation, the founders hold a Class F share with 50.1% of the voting power and that the Long-Term Benefit Trust will elect a board majority, limiting the influence of Class A buyers.

Why it matters: The price range moves the laboratory to the final steps before a public listing, but the implied valuation has already come down from earlier reports of about $2 trillion to around $1.40 trillion; trading still depends on the registration becoming effective, and the roadshow is offering a range, not a final price.

3. OpenAI introduces Businesses in ChatGPT and Business Agents in beta

On November 10, 2026, OpenAI said it was introducing Businesses in ChatGPT in beta, describing it as a new way for companies to connect with the 1.2 billion people who use ChatGPT each week. Through Business Hub, businesses can claim their organization, manage their information and see how people engage with them in ChatGPT. The company is also adding Business Agents, starting with eligible ecommerce and online-service businesses in the United States.

In Business Hub, verified companies can update key details and add self-service product feeds with pricing and availability. OpenAI said any eligible business in a place where ChatGPT is available can claim its profile through Business Hub; existing OpenAI Enterprise customers and advertisers can work with their representative; and Shopify merchants can connect a store from Business Hub or the Shopify admin to streamline verification, with eligible Shopify products already integrated into ChatGPT through Shopify Catalog.

On insights, Business Hub provides aggregated figures such as mentions and website clicks, as well as discovery topics—common themes in conversations where a business’s website is cited. OpenAI wrote that it does not share users’ conversations or Business Agent chats with businesses, and that insights are available only in aggregate and only once they meet certain thresholds.

After claiming a profile, eligible organizations can build, test and launch a Business Agent. When a business appears in a model response, people can open its ChatGPT profile; if a Business Agent is available, a customer can start a separate conversation to compare offerings or ask follow-up questions, grounded in the information the business has provided. That conversation is clearly labeled, kept separate from the original ChatGPT thread, and distinct from ChatGPT’s independent answers. Businesses can also sponsor their Business Agent and use it as the destination for ChatGPT Ads; sponsored placements are clearly labeled and kept separate from ChatGPT responses. The Verge reported the beta launch the same day.

Why it matters: OpenAI is opening a path from its weekly audience of 1.2 billion users to businesses, but only behind verification, aggregated insights and clearly labeled, separate conversations; the first step is a claimed profile and a branded agent, not a general opening to every industry and every market.

4. Google unveils Gemini Robotics 2.5, with a 96% success rate in a Mercedes-Benz line trial

On November 10, 2026, Google DeepMind, in a post by Carolina Parada, introduced Gemini Robotics 2.5, calling it the company’s most capable vision-language-action (VLA) model to date. The model is meant to widen the range of skills a robot can learn and the speed at which it can acquire them, and to improve dexterity, reliability and safety so that robots can move out of the laboratory and into more unstructured real-world settings. The company said it had earlier this year brought whole-body intelligence to robots of many types, including humanoids, and that the latest progress was due in large part to Gemini 4 Argon’s ability to accelerate robotics research. Gemini Robotics 2.5 is rolling out to Google’s robotics partners.

Google said it evaluated the model’s task breadth, performance and complexity across more than 200 distinct tasks, such as serving food with tongs, tying a bow on a gift box and pipetting liquid, combining tool use, manipulation of deformable objects and the careful handling of delicate items. Compared with prior baselines, Gemini Robotics 2.5 achieved a 30% to 100% higher success rate depending on the type of task, and completed tasks 20% to 30% faster. The model works natively with multiple robot types, including Apptronik’s Apollo 2 humanoid, which uses the five-finger, 22-degree-of-freedom SharpaWave hand, and the Franka Duo platform with two-fingered parallel grippers.

To test reliability, Google partnered with Apptronik and Mercedes-Benz to deploy Apollo 2 on active automotive production lines in Kecskemét, Hungary, and Sindelfingen, Germany. In the proof of concept, the robot used the embodied reasoning model Gemini Robotics ER to track the workflow and the Gemini Robotics 2.5 action model to control its arms and hands, picking metal automotive brackets from supply bins, placing them into sequencing containers and correcting for gaps, with a 96% success rate. Jeff Cardenas, co-founder and chief executive of Apptronik, said the models improved substantially in a few weeks of post-training and carried over to the Mercedes-Benz lines with very little drop-off.

On safety, Gemini Robotics 2.5 uses inference-time reasoning guidance to steer around physical hazards in real time. Google said the capability raised safe task completion by 33% around obstacles and by 43% around nearby people. To test adversarial jailbreaks aimed at robots, the company is also releasing ASIMOV-Jailbreak, an open-source benchmark with hundreds of prompts that disguise unsafe requests across dozens of scenes; it said Gemini 4 Argon achieved 99.5% jailbreak resistance, up from 82% for the previous baseline, Gemini Robotics ER 2.

Why it matters: This is not only another laboratory demonstration of tying a bow—Google is putting a 96% success rate on working Mercedes-Benz production lines, and it is publishing a new jailbreak benchmark and steerability results alongside it, treating reliability and safety as conditions for entering a factory.

5. MiniMax open-weights H3-Image, saying it can generate 4K images in under 10 seconds

On November 11, 2026, MiniMax released MiniMax H3-Image, an open-weight image generation model built on H3, with support for text-to-image generation, reference-based generation and image editing. The company said the model can generate a 4K (4096×4096) image in under 10 seconds, with richer detail and stronger prompt adherence, that it can understand and use up to 16 reference images, and that its 4K generation speed is up to about five times that of mainstream alternatives. The weights have been published on Hugging Face as MiniMaxAI/MiniMax-H3-Image, and the model is also available through the MiniMax API, MiniMax Code and Hailuo AI.

MiniMax said Standard mode is about 1.5 to 2 times as fast as mainstream alternatives, while Highspeed mode reaches 3 to 4 times; pricing is about 50% to 70% of comparable mainstream offerings, and for launch week the company is offering 50% off image-generation requests at 4K resolution and above. The model is released under the MiniMax-H3-Image Community License Agreement, dated November 10, 2026. The agreement states that the applicable territory is worldwide, excluding the European Union, the United Kingdom, the Republic of Korea and the United States of America.

Why it matters: MiniMax is putting both 4K generation speed and open weights into the market, lowering the barrier to obtaining and running the model; the community license, however, expressly carves out markets including the European Union, Britain, South Korea and the United States, so open weights are not the same as unrestricted commercial use everywhere.

---

Today in one line: Google made Gemini 4 Carbon, its strongest model to date, generally available at Argon’s introductory price and previewed /swarm in Antigravity; Anthropic launched its IPO roadshow with a $70 to $78 price range and applied to list as ANT; OpenAI introduced Businesses in ChatGPT and Business Agents in beta; Google DeepMind unveiled Gemini Robotics 2.5 and reported a 96% success rate in a Mercedes-Benz line trial; and MiniMax open-weighted H3-Image, saying it can generate 4K images in under 10 seconds.