AI News
Nov 11, 2026 · TuesdayPublic sources · Daily newsPublished daily at 08:00 Beijing time

AI News Daily

AI developments selected from public sources

Issue 20411Nov 2026Tuesday
11monthThis month 16 issues
12
major stories
25
sources
5
first-party stories
About 6 min read

Lead story

Anthropic Launches Roadshow for Proposed IPO

Anthropic announced that it has launched the roadshow for its proposed initial public offering (IPO). The company also stated that the registration statement for the relevant securities has been filed with the SEC but has not yet become effective, and that no securities may be sold nor offers to buy accepted before it becomes effective, and that this post does not constitute an offer to sell or a solicitation of an offer to buy.

First-party sourcesAnthropicAlso 9 other sources reported thisOriginal ↗
01

Model releases / updates

Ant Lingbo Open-Sources Tactile Foundation Model LingBot-Touch, Real-Robot Task Success Rate Up 14 to 26 Percentage Points

Ant Lingbo Technology has open-sourced the tactile foundation model LingBot-Touch, pretrained on nearly 5,000 hours of real tactile data and covering three sensing principles—magnetic tactile, visuotactile and piezoresistive—encoding different signals into unified tactile tokens. In 26 real-robot tests on the Sharpa Wave and XHand1 dexterous hands, after integration the average task success rates of four action models—LingBot-VLA, π0.5, Diffusion Policy and T-Rex—rose by about 14 to 26 percentage points compared with no tactile sensing. The technical report, model and code have been open-sourced on Hugging Face, GitHub and ModelScope.

Original

Gemini 4 Carbon Lands on Google Cloud and Agent Studio

Gemini 4 Carbon is rolling out on Google Cloud Platform and is already available in Agent Studio. The official description calls it the most capable model for deep work, with frontier performance in long-horizon agentic coding, enterprise knowledge work and scientific reasoning, and an output token limit of 1M. The publisher said it will begin hands-on testing.

Original

Google DeepMind Releases Gemini Robotics 2.5 Vision-Language-Action Model

Google DeepMind releases Gemini Robotics 2.5, calling it the most capable vision-language-action (VLA) model to date, with success rates improved by 30% to 100% over previous baselines across 200 tasks and task completion speed improved by 20% to 30%. In proof-of-concept trials at Mercedes-Benz factories in Hungary and Germany, the Apptronik Apollo 2 humanoid robot equipped with the model autonomously sorted automotive parts with a 96% success rate. The model is being rolled out to robotics partners, alongside the release of the open-source benchmark ASIMOV-Jailbreak, on which Gemini 4 Argon achieves a 99.5% jailbreak resistance rate.

Original

MiniMax releases image model H3-Image and opens model weights

MiniMax releases H3-Image, an image model based on H3, and opens its model weights. The company says it can generate 4K × 4K, or 16-megapixel, images in 10-20 seconds, and supports strong instruction following, image editing and up to 16 reference images. The model is now available via API, MCode and Hailuo AI, with 50% off generations above 16 megapixels during launch week.

Original
02

Product release/Updated

Arena Showcases 17 Generation Examples from Claude Fable 5.2

Arena showcased 17 generation results from Anthropic's Claude Fable 5.2 on Arena, covering detailed 3D worlds, playable games, website UIs, data visualizations and interactive animations generated from code. Arena previously said Claude Fable 5.2 (xHigh) topped the Code Arena WebDev leaderboard with 1,832 points, up to 30% faster than Fable 5.1, at $8/$40 per million tokens, down from the previous $10/$50, ranking ahead of Claude Opus 5.5 (Max) at 1,818 points, Gemini 4 Carbon (High) at 1,768 points and GPT-6 Astra (Max) at 1,767 points.

Original

NVIDIA open-sources Agent Substrate, providing a runtime for large-scale agents on Kubernetes

NVIDIA releases open-source Agent Substrate, which multiplexes large numbers of agent sessions onto a small number of hot Kubernetes worker Pods; idle sessions are checkpointed together with their memory and filesystem state, then restored to any idle worker, achieving 10x the container density of standard containers. The project has been donated to the CNCF Sandbox and targets three types of workloads: coding assistants, reinforcement learning post-training and short-lived GPU sandboxes. Working with NVIDIA Dynamo, Dynamo feeds back inference signals such as estimated time to first token and GPU pool saturation so Agent Substrate can decide whether to keep, suspend or resume an agent.

Original
03

Industry news

Anthropic Said to Plan Nasdaq Listing Next Week, Expected Valuation Range $1.1 Trillion to $1.75 Trillion

According to people familiar with the matter, Anthropic plans to list on Nasdaq next Thursday or Friday under the ticker ANT, will hold roadshow meetings with core investors this week, and is expected to file documents early this week disclosing the IPO price range. The market's expected valuation range for the company is $1.1 trillion to $1.75 trillion, corresponding to an expected fundraising size of $80 billion to $100 billion. Foreign media analysis noted that if the market environment deteriorates further, the company could still postpone the IPO.

Original

Anthropic Withdraws IPO Pricing Document; Draft Had Sought to Raise Up to $69.3 Billion

Reuters reported that Anthropic sent banks a securities offering draft on the evening of November 9, seeking to raise up to $69.3 billion through an IPO, and a follow-up email then said the pricing details were inaccurate, with the company stating it would update the document. The draft showed 866.7 million Class A shares priced at $70 to $80, of which Anthropic was selling about 843.3 million shares and existing shareholders were selling 23.3 million shares; there was also a $26 billion private placement of Class N common stock involving Google, Nvidia, AMD, Broadcom, Micron, Microsoft and TSMC. The report said Anthropic's revenue for the quarter ended September 30, 2026 was close to $17 billion, with quarterly growth having slowed markedly.

Original
04

Research papers

Google DeepMind Formally Verifies Signal Protocol Code with AlphaProof Code

Google DeepMind researchers used the AlphaProof Code agent framework to complete formal verification of key security properties of the actual Signal protocol code, generating more than six million lines of Lean proofs over the course of four weeks. The framework divides work among multiple LLM agents, with human experts responsible for determining the security goals to be proven and reviewing about one percent of the code. The team also reintroduced two real, already-fixed vulnerabilities into an isolated version of the Signal code, and the system produced machine-checked proofs of insecurity in both cases. DeepMind has also open-sourced the formal verification artifacts.

Original

Anthropic Fable 5.2 completes 41.3% of real freelance tasks on the Remote Labor Index

Anthropic's new model Fable 5.2 tops Scale Labs and CAIS's Remote Labor Index at 41.3%, a benchmark measuring the share of real freelance projects AI agents complete at a quality acceptable to clients. When the benchmark launched a year ago, the top score, Manus, was only 2.5%; in July, Fable 5 scored about 16%, and Fable 5.1 under 18%. Behind it on the leaderboard are Anthropic Opus 5.5 at 21.25% and OpenAI GPT 6 Astra at 20.83%, with Google Gemini Argon 4 fifth at 17.08%. The dataset contains 240 complete projects from 358 verified Upwork professionals, worth about $144,000 in total, with an average of $632.60 per project and nearly 29 hours of average human labor time; Scale Labs says Fable 5.2 improved most on audio and music editing, architecture and interior design, and 3D modeling, while industrial product design and animation remain difficult.

Original
05

Tips and perspectives

The Harness Wars: Why Model Companies Are Racing to Turn Agent Runtime Systems into Services

A Microsoft technical whitepaper shows that with the same Claude Opus 4.8, merely swapping the Harness behind Copilot Studio raised the pass rate on complex tasks by 22.1 percentage points and cut input token consumption by 21%. Anthropic, OpenAI, DeepSeek, Moonshot AI and others have successively opened up or hosted their Harness for developers, and Kimi proposed Harness as a Service on November 5. The article notes that hosting fees currently account for only a small part of total Agent costs; in an example of Anthropic Managed Agents running a one-hour coding task, the total cost was $0.58, with hosted runtime fees accounting for about 14%.

Original
06

Briefs

End of issue

AI News DailyAutomaticallycompiledfrom public reports, with source links · View archive