AI Daily

Anthropic launches Claude Fable and Mythos 5.2 with different access requirements

2026-11-09 · New

Anthropic released Fable and Mythos 5.2 using the same underlying model. Fable serves general users; Mythos requires its cybersecurity or life-sciences verification program. Standard API input/output pricing is $8/$40 per million tokens. Fable applies cyber and bio safeguards; teams must choose access appropriate to their tasks and eligibility. The optional faster mode costs more.

Anthropic: Claude Fable and Mythos 5.2

ByteDance releases Seedream 5.5 with iterative editing and transparent images

2026-11-09 · New

ByteDance Seed released Seedream 5.5, demonstrating consistency across 100 successive edits and native RGBA transparency. Its official page links to Dola and BytePlus Playground. Transparency and localized editing help create reusable design assets. The 100-round result is a company demonstration, not a guarantee for every image. The higher-specification Pro version is still forthcoming.

ByteDance Seed: Seedream 5.5

Seedream 5.5 official access

Qwen Office 2.0 adds team spaces, event triggers and mobile remote control

2026-11-09 · New

Qwen Office desktop 2.0.0 adds team spaces connected to DingTalk documents, AI tables and Yida. Messages, approvals and tasks can trigger digital employees. The mobile app is released with remote control of enabled computers. Teams can turn one-off chats into event-triggered workflows. Existing document permissions remain in force; remote control requires the same account and organization, and administrators can disable it.

Qwen Office desktop 2.0.0 release notes

The Beijing News: Qwen Office 2.0

Baidu presents ERNIE Evolve for automated model-engineering experiments

2026-11-09 · New

Baidu presented ERNIE Evolve, combining isolated experiments with a separate evaluation stage, and released selected code and logs. It reports performance reaching 6.236 times baseline after four optimization rounds in a low-latency Kunlunxin P800 setting. The records let developers inspect changes and evaluation procedures. The figure is Baidu’s configuration-specific result, not external replication. This is an automated research method, not a new launch of the model under development.

Baidu: ERNIE Evolve automated research

ERNIE Evolve experiment code and records

OpenAI deploys Control Room to centrally stop research-model outputs

2026-11-09 · New

OpenAI described Control Room, deployed across its research-inference infrastructure for agent training, evaluations and data processing. It can stop model outputs within 35 seconds of an intervention decision. Central control provides a stop mechanism across research clusters and fails closed if the control service fails. The 35 seconds measures post-decision output cessation, not risk-detection time or termination of every underlying workload.

OpenAI: Control Room

Google researchers use Lean to verify AI-generated quantum programs

2026-11-07 · Seven-day watch

Google DeepMind and Quantum AI researchers published a method in which model agents propose quantum programs against Lean specifications and verify correctness and resource costs, collaborating through shared knowledge repositories. The approach offers checkable proofs for quantum algorithms that are difficult to simulate directly. Reported improvements concern algorithms and cost models, not measured hardware speedups. The original November 6 UTC submission falls on November 7 in Beijing.

Discovering formally verified quantum programs with AI

Quantum-program paper affiliations and method

Claude Fable · Anthropic

5.2 · 2026-11-09 · stable

This issue: general-user access; standard API input/output pricing is $8/$40 per million tokens.

Anthropic: Claude Fable and Mythos 5.2

Checked 2026-11-10T03:22:26.443259+00:00

Claude Mythos · Anthropic

5.2 · 2026-11-09 · stable

This issue: verified-access variant of the same model; cybersecurity or life-sciences verification is required.

Anthropic: Claude Fable and Mythos 5.2

Checked 2026-11-10T03:22:26.443259+00:00

Seedream · ByteDance

5.5 · 2026-11-09 · stable

This issue: iterative editing and RGBA transparency; Dola and BytePlus Playground links are offered. Pro is forthcoming.

ByteDance Seed: Seedream 5.5

Seedream 5.5 official access

Checked 2026-11-10T03:22:26.443259+00:00

Qwen Office · Alibaba

2.0.0 · 2026-11-09 · stable

This issue: team spaces, event triggers and mobile remote control, subject to account, organization and administrator permissions.

Qwen Office desktop 2.0.0 release notes

Checked 2026-11-10T03:22:26.443259+00:00

Bailian CLI · Alibaba Cloud

2.8.4 · 2026-11-06 · stable

Recent verified baseline: fine-tuning parameter and metrics updates; not a new release in this issue.

Bailian CLI release notes

Checked 2026-11-10T03:22:26.443259+00:00

Google announces TPU 8tx and 8i Max with new hosts, networking and packaging

2026-11-09 · New

Google announced two TPU system configurations. 8tx pairs 8t silicon with fifth-generation AMD EPYC hosts and Boardfly networking. 8i Max packages two 8i chips together, doubling compute, SRAM and HBM bandwidth versus one 8i, according to Google. 8tx targets confidential inference and private fine-tuning; 8i Max targets very large context windows. No general-availability date was established in the announcement reviewed here, and doubled chip resources do not guarantee doubled application throughput.

Google: TPU 8i Max and 8tx

Meituan shares MR0 robot research and training code across 12 embodiments

2026-11-06 · Seven-day watch

Meituan’s robotics team shared MR0 vision-language-action research with training and evaluation code. The paper uses 147 million vision-language samples and 15,060 hours of robot trajectories spanning 12 embodiments. Physical demonstrations involving blocks, clothes and cables distinguish it from digital-only agents. Code and evaluation interfaces support scrutiny, but these remain research results and demonstrations, not deliveries of general-purpose robots or unattended production capability.

Meituan Robotics 0 paper

Meituan MR0 training and evaluation code

NVIDIA, Rebellions and SK Telecom plan hybrid inference infrastructure; R200 targets 2028

2026-11-08 · Recent, within 72 hours

NVIDIA, Rebellions and SK Telecom announced a collaboration using Rebellions R200 NPUs, NVLink Fusion and MGX. The plan assigns prefill to NVIDIA and decode to Rebellions, with SK Telecom pursuing data-center deployment. The design assigns inference stages to different accelerators, offering an integration path for Korean infrastructure. R200 systems are expected to be deployment-ready in 2028; integration, delivery and real-workload efficiency remain to be demonstrated.

NVIDIA, Rebellions and SK Telecom collaboration

Six countries and ten AI companies sign Paris declaration on evaluations and risk reporting

2026-11-09 · New

Six countries including Canada and France, ten AI companies and four civil-society organizations signed the Paris AI safety declaration, committing to shared risk methods, third-party evaluations, incident information sharing and a repository of non-sensitive findings. Company commitments also address fail-safe and shutdown mechanisms for training and internal models. Developers now need to turn commitments into inspectable evaluation and reporting processes. The declaration is not a newly binding law or proof that the repository already operates.

Canadian PM: Paris AI safety and security declaration

Adobe and Disney expand multiyear alliance for custom Firefly creative models

2026-11-09 · New

Adobe and Disney expanded their multiyear strategic alliance, bringing Firefly Foundry custom models into creative and experience development and involving StudioLAB and ESPN Edge. The companies had already worked together in Imagineering. The alliance combines proprietary content with engineering support for workflows such as concept design. Some projects are under way and further activations are forthcoming. No financial value was disclosed, and the custom models are not thereby available to the public.

Adobe and Disney expand multiyear alliance

NVIDIA board opens independent inquiry into reports Huang was targeted by foreign intelligence

2026-11-08 · Recent, within 72 hours

NVIDIA’s board announced an inquiry by independent directors and outside counsel into reports that Jensen Huang was targeted by foreign intelligence. It says it has no reason to believe national or company secrets were compromised, and Huang will cooperate. The development is the opening of a corporate-governance inquiry. Its scope and formal findings are the next disclosures to watch. The statement neither establishes a leak nor turns reported allegations into proven facts.

NVIDIA board announces independent inquiry

Phison reports NT$86.076 billion Q3 revenue and an AI-ecosystem share above half

2026-11-06 · Seven-day watch

Phison reported third-quarter consolidated revenue of NT$86.076 billion, up 374.6% year on year, and a 59.7% gross margin. Its broadly defined AI ecosystem accounted for more than half of revenue, including enterprise SSDs, AI PCs, networking and servers. This is realized storage-supply-chain revenue, rather than a funding valuation or order target; the broad category is not pure AI-model revenue. Inventory turnover days rose from 224 a year earlier to 274, making subsequent shipment absorption relevant.

Phison third-quarter 2026 results