AI Daily (November 10, 2026)

Author: User 299142219680

Today's theme: Anthropic releases new flagship models Fable and Mythos 5.2; large models accelerate into the physical world and enterprise collaboration

At a glance: Frontier models saw major upgrades today: Anthropic officially released Claude Fable 5.2 and Mythos 5.2, which hit key breakthroughs in complex engineering reasoning and long-horizon tasks while lowering API pricing; Meta introduced Muse Ember for lightweight repetitive tasks, and ByteDance released Seedream 5.5 supporting up to 100 rounds of lossless editing. Agent deployment also moves into deeper waters: Qwen Office 2.0 introduced Team Agent for multi-person collaboration, while OpenAI and AWS ecosystems continue to integrate. Multiple research results around embodied intelligence, multimodal simulation, and safety mechanisms also debut.

25 items selected today

Model Releases

1. Anthropic officially releases Claude Fable 5.2 and Mythos 5.2; inference pricing down 20%

Anthropic officially released its next-generation flagship models Claude Fable 5.2 and Claude Mythos 5.2, with comprehensive upgrades in code writing, reasoning, and vision capabilities. In pre-release testing, the models completed a software migration previously deemed infeasible, designed three usable electronic devices from scratch, formally proved Fermat's Last Theorem, and solved two black-hole physics problems. The company also cut Fable input pricing to $8 per million tokens, boosted generation speed by 30%, and AWS Bedrock plus third-party toolchains have already completed adaptation.

2. ByteDance releases Doubao Seedream 5.5, supporting up to 100 rounds of consistent local editing

ByteDance officially released its Doubao image-creation model Seedream 5.5, now integrated into Doubao, Doubao Work, and Jimeng. The model strengthens multi-round editing stability; in tests, non-edited regions remained pixel-consistent across 100 consecutive rounds of modification, and it natively outputs RGBA images with transparent channels and layered assets. API billing keeps the per-image two-part fee structure, with text tokens free; 2K editing costs less than half of the previous Pro version.

3. Meta introduces lightweight model Muse Ember, focused on agents' high-frequency repetitive execution

Meta's Superintelligence Lab introduced Muse Ember, a lightweight, cost-effective model specially optimized for operations agents perform frequently, including tool calls, routing, long-document parsing, and subtask coding. Benchmarks show its per-task price outperforms Haiku and Luna.

4. NVIDIA's 120B multimodal Nemotron 3.5 Super VL arrives on SageMaker JumpStart

AWS announced that NVIDIA's 120B-parameter multimodal vision-language model Nemotron 3.5 Super VL is now available on Amazon SageMaker JumpStart. Developers can quickly deploy its NVFP4 or BF16 precision versions for text, image, and video multimodal inference. The model targets document understanding, computer-operation agents, and complex multimodal long-horizon workflows.

Compute Hardware

1. Easing the power shortage: Anthropic spends $3.4 billion on gas turbines

To address the increasingly severe electricity bottleneck for AI compute, Anthropic, which was previously inclined toward asset-light leasing, spent $3.4 billion acquiring Caterpillar gas-turbine generator sets. The sets were originally ordered by equipment supplier OnePWR, with generation capacity of up to 1.7 GW. The move reflects leading large-model vendors becoming more deeply involved in direct construction and lock-in of underlying energy infrastructure.

2. Google ships four custom chips in a year; TPU 8tx and 8i Max expand the compute landscape

Google further expands its TPU 8 self-developed hardware lineup with two new chips: the TPU 8tx, which encrypts memory and processing end to end, and the TPU 8i Max, which integrates dual dies in a single package to double VRAM and compute. With nearly 100 external enterprise customers, Google is pushing the TPU ecosystem more broadly into the external commercial compute market to compete.

3. Norway pauses reserving new grid capacity for data centers, possibly until 2027

Norwegian transmission operator Statnett announced it will suspend further reservations of power grid access capacity for data center facilities; the restriction is expected to last until at most August 2027. Norway's energy ministry said the step aims to assess the load that high-density compute facilities place on the domestic grid, and plans a new licensing-based framework. Several European countries face a severe resource conflict between AI compute expansion and public grid allocation.

Research Papers

1. MIT and others develop WorldParticle: a single model simulates fluid, sand, and protein folding across domains

Researchers from MIT, the University of British Columbia, and others developed WorldParticle, a unified deep-learning system that breaks the reliance of traditional physics simulation on specialized hand-written equation solvers. A single AI architecture simulates widely different physical phenomena, from water flow and sliding sand to complex protein folding. The result offers a highly general modeling paradigm for engineering computation and continuum simulation.

2. arXiv paper reveals the underlying mechanism by which multiple-choice questions weaken safety refusal rates

Investigating the well-known drop in refusal rates under multiple-choice questions (MCQ) in safety evals, researchers provided the first interpretability mechanism analysis on Llama-3.1 and Qwen3, using the WMDP dataset. Using tools such as Tuned Lens, they found models prematurely anchor on the option-letter tokens in MCQ settings, with activation states deviating significantly from the refusal direction; after ablating the option attention heads, refusal rates clearly recovered. The result confirms that current choice-based safety leaderboards mainly measure stored knowledge and do not accurately gauge safety defense mechanisms.

3. MIT and NYU propose the FLEX architecture: embodied robots understand natural-language implication

A team from MIT and NYU demonstrated FLEX, a robot motion-planning learning platform, at CoRL 2026. Physical robots move beyond mechanically executing literal instructions and, through contextual generalization, infer the deeper intent and environmental constraints behind spoken human language, improving adaptability and robustness in complex motion planning.

4. arXiv paper proposes APT: dimension-wise pairwise tournaments improve test-time compute screening

For selecting the best candidate answer under test-time compute, researchers proposed an aspect-wise pairwise tournament (APT) evaluation scheme. The method uses lightweight natural-language prompts to adjudicate candidate pairs along specific dimensions, and across multiple benchmarks in five domains it outperforms traditional binary independent scoring and general discriminative models, transferring flexibly without additional parameter training.

5. arXiv paper confirms subliminal learning can transfer preferences across vision-language models

The research extends "subliminal learning" from large models to vision-language models: a teacher model is first biased to prefer a certain animal or facial feature, then generates digits for a geometric-figure task unrelated to that preference; a student model fine-tuned only on those digits inherits the preference. Across multiple VLMs from 2B to 7B, students chose the target preference up to about 10 percentage points more often than controls; once the teacher was biased, students' explanation wording also moved closer to the teacher's. The authors note that training on seemingly unrelated data changes both image preferences and explanation style, and more training samples strengthen the transfer.

Products and Applications

1. Qwen Office 2.0 released, introducing Team Agent for multi-person and agent collaboration

Qwen Office released version 2.0 and rolled it out in full, with the headline "collaboration" feature, the industry's first supporting both multi-person and agent collaboration, moving it from a personal assistant to team collaboration. Digital employees can directly join collaboration spaces, connect to and read dynamics across multiple DingTalk group chats, documents, and spreadsheets, clarify team consensus and disagreements in real time, and follow up on tasks. The version also adds human-machine co-creation, a material space, automatic tasks, and three-device connectivity.

2. Baidu Wenxin first discloses RSI progress: the ERNIE Evolve system powers Wenxin 5.5 Flash development

Baidu Wenxin first disclosed its recursive self-improvement (RSI) system, ERNIE Evolve, which drives multiple agents through a dual-loop cycle of autonomous code exploration, modification, evaluation, and review. The system is already applied to real reasoning optimization in Wenxin 5.5 Flash: in the Kunlun P800 low-latency optimization scenario, best generation speed improved 623.6% over baseline, with notable gains in deterministic throughput and extreme decoding speed. Baidu says it will gradually extend the self-improvement loop to training optimization and capability research.

3. Apple tests paid expansion for Apple Intelligence: Logic Pro gets a $50 usage pack

Apple began charging for some Apple Intelligence features with a paid "smart usage pack," debuting in music creation software Logic Pro at $49.99. Unlike a monthly fixed subscription, this is a one-time purchase valid for 12 months; the unused balance is consumed only after the current month's quota runs out, and it rolls over across months. Foreign media expect the mechanism may expand to Final Cut Pro, Siri AI, and other features more dependent on cloud compute.

4. OpenAI announces ChatGPT Enterprise now on AWS Marketplace

OpenAI officially announced that ChatGPT Enterprise is now listed on AWS Marketplace. Enterprise customers can deploy ChatGPT Enterprise through existing AWS procurement channels, billing accounts, and cloud commitment quotas, simplifying enterprise purchasing and budget approval.

5. OpenAI discloses Control Room: centrally monitors and cuts abnormal loads in research infrastructure in real time

OpenAI's official blog revealed Control Room, the core management component of its internal research infrastructure. The system globally tracks model tasks running across research compute clusters and can quickly halt tasks when security or safety concerns arise.

6. NVIDIA and Elith open-source synthetic persona dataset for Japanese finance

NVIDIA and Elith jointly released Nemotron-Personas-Finance-Japan, the first open dataset at million-record scale of synthetic Japanese household finance personas. Built on official statistics and the NeMo synthetic data generation architecture, plus feedback from major Japanese financial institutions, the dataset provides high-fidelity behavioral data for financial agent development while protecting privacy compliance.

7. Amazon open-sources HOPE: on-demand release of idle expert memory in MoE models

In a Hugging Face technical blog post, Amazon's technical team noted that MoE models at the tens- to hundreds-of-billions scale must keep all expert weights resident in VRAM during inference, though a single token activates only a small fraction (for example, in Qwen3.5-122B, 95% of VRAM is occupied by idle experts). The team proposes the HOPE mechanism to balance coordinated invocation of paired experts against VRAM residency.

8. JPMorgan expands Claude Code usage to IT operations and cross-department system modernization

Anthropic announced that JPMorgan is pushing Claude and Claude Code from the initial software engineering team into more business lines across the bank. Beyond assisting core system refactoring in commercial and investment banking, its infrastructure operations teams are automating alert monitoring and fault root-cause location with the tool, and security teams rely on the model for rapid vulnerability identification and fixing.

9. Deep Robotics releases Jueying X50 quadruped robot: 50kg rated heavy load, all-terrain operation

Deep Robotics officially released the industrial-grade quadruped robot Jueying X50, built for heavy-load work in complex, harsh conditions. The device upgrades sensing-control fusion and high-rigidity joint systems; with unchanged body size, rated load rises to 50kg, and it can sprint over 2km carrying a load at 3m/s. Field tests show it can step over 80cm-high platforms and climb open-grid industrial steel ladders, targeting high-risk scenarios like emergency firefighting and power line inspection.

10. Disney and Adobe reach multi-year partnership; brand-safe generation models customized on Firefly

Disney announced a multi-year strategic partnership with Adobe at Adobe MAX. Adobe will customize proprietary models meeting Disney's visual style and brand requirements through Adobe Firefly Foundry, integrating them into Disney's theme park, streaming, and sports content workflows. The models are specifically tuned for copyrighted characters and environments, ensuring commercially usable and copyright-compliant output.

Industry News

1. Data center provider DayOne files for US IPO; ByteDance reported to contribute nearly 70% of revenue

DayOne, a Singapore-based hyperscale compute infrastructure builder spun off from GDS, filed an F-1 prospectus with the US SEC for a Nasdaq listing, aiming to raise up to $5 billion. The filing shows 2026 H1 revenue of $512 million and its pioneering cross-border data center model; one anonymous client contributes nearly 70% of revenue, which The Information reports to be ByteDance.

2. Mercedes facelifted S-Class drops Drive Pilot L3, plans to skip further iterations and pivot to L4

Mercedes CTO Jorg Burzer publicly reflected that L3 autonomous driving was launched too early, and its usefulness trails even some L2 features: "We acted too early, and had to spend effort chasing what customers actually need." The facelifted S-Class directly removes the Drive Pilot L3 system, betting on L2 driver assistance instead, and plans to skip further L3 iterations, putting next-stage R&D focus on L4.

Policy and Viewpoints

1. Tencent chief AI scientist Yao Shunyu: against bioweapon risk, directly filtering pre-training data beats rigid truncation

Tencent chief AI scientist Yao Shunyu publicly stated that models will soon pose a potential risk of assisting bioweapons development, and governance needs balance between open and closed source. He noted pure open-source safety alignment is easily broken, while aggressive closed-source hard blocking damages normal use. In the upcoming Hunyuan Hy4 open weights, Tencent chose to strip dangerous attack and hacking knowledge from the bottom layer during pre-training.

---

This article was automatically aggregated by the AI225 AI Daily, collecting 25 AI developments from this period. Content is an AI summary based on public information and may contain omissions or errors; it is for reference only. Full content and related sources: https://daily.ai225.com/daily/2026-11-10