Claude Fable 5.2 与 Mythos 5.2 齐发：同一模型，两扇门
Anthropic released new versions of its most advanced models on Monday, with the launch of Mythos 5.2 and Fable 5.2. The company says the new models “show improvements in areas like coding, reasoning, and vision” and communicate more clearly with users.
Notably, Anthropic is lowering the cost of Fable from $10 to $8 per million input tokens. As a result, Fable 5.2 is now cheaper than its primary competitor GPT-6 Astra. Anthropic says that the new version of Fable will return responses 30% faster across effort levels. Faster inference serving is available for double the price.
Fable 5.2 is also compatible with zero data retention (ZDR) policies through the Enterprise Frontier Safeguards system, long a sticking point for many enterprise clients.
It’s Anthropic’s first model release since the arrival of a third-party report from the assessment group METR, which suggested the lab might be approaching recursive self-improvement (RSI) — or the ability for AI models to carry on model training independently. The METR report found internal models were speeding up Anthropic’s AI progress by roughly 2x, although many uncertainties remain.
Still, Anthropic’s release notes significant ongoing challenges in accurately testing model capabilities. “Claude Mythos 5.2 and Fable 5.2 each voice suspicion that they are being evaluated more often than any of our recent models, which challenges our ability to assess how they will act in real-world settings. As these settings expand and capabilities increase with future models, we expect this challenge to grow.”
Anthropic maintains that the new version of Mythos does not have the ability to act covertly against its user, a verdict that’s detailed further in Mythos 5.2’s system card.
“Claude Mythos 5.2 shows stronger covert capabilities than prior Claude models on several of our evaluations, particularly when its reasoning is hidden from the monitor,” the system card reads. “That said, we believe that low levels of absolute performance on our new covert evaluations support claims that rely on detection of harmful actions in high-stakes situations.”
Fable 5.2 has been available to select parties in pre-release for some time, and Anthropic says the model has already “solved two open problems in theoretical physics, formalized Fermat’s Last Theorem, and supported numerous other research projects.”
Also on Monday, AI startup Physical Superintelligence announced a new proof of the Aaronson‑Ambainis conjecture, a pivotal step in setting theoretical limits for quantum computing. Notably, the lab said the result was achieved by running “a non-public Anthropic model” within PSI’s own system. Anthropic did not confirm or deny whether Fable 5.2 was the model in question.
字节 Seedream 5.5：图片能改 100 轮，透明底直接出
ByteDance Seed Introduces Seedream 5.5 for Multi-Round Image Editing
ByteDance Seed introduced Seedream 5.5 on November 9, 2026, describing it as its next-generation image creation model for professional visual production. The release focuses on stable editing across as many as 100 rounds, native transparent assets, and improvements to realism and small-text rendering.
The editing work targets the cumulative errors that can appear when each generated output becomes the input for another revision. ByteDance says targeted training preserves unedited regions at the pixel level, eliminates color shifts in those areas, and reduces detail loss and image degradation over successive edits. For designers, that puts continuity between revisions at the center of the release.
In the company’s editing demonstration, clothing, accessories and tabletop objects around the same person changed repeatedly. The model was still able to adjust small items such as rings and eyeglass frames while accounting for the subject’s pose, contact between accessories and skin, occlusion and lighting. A separate interior demonstration moved a box around a changing study, with the model adjusting perspective, cast shadows and newly exposed areas as the room evolved.
Transparent Assets for Compositing
Seedream 5.5 can generate RGBA images directly, using an alpha channel to represent full or partial transparency. Creators can produce these assets from text descriptions or reference images. ByteDance says the resulting elements can enter layout, compositing and production workflows without a separate background-removal and edge-cleanup step.
The examples concentrate on difficult edges and reuse. A transparent snowy-owl asset retains the feather tips, small gaps between feathers and claw outlines. Another demonstration reuses character assets across different settings, preserving their defining features and textures while changing poses, groupings and composition.
The official model page provides two more production-oriented examples: extracting a flamingo as a separate transparent RGBA layer, and generating a running-girl sequence with each image on its own transparent layer.
Realism and Small-Text Rendering
ByteDance says Seedream 5.5 renders lighting, materials and skin textures more naturally while reducing excessive gloss, overexposure and blotchy textures in some outputs. Its text-rendering changes address complex character shapes and densely arranged small type. Demonstrations include a compact Shakespeare newspaper page with English articles arranged around illustrations, and a Chinese blue-morpho infographic whose eight illustrated explanatory notes retain legible small characters.
Vendor Evaluations and a Staged Rollout
ByteDance’s evaluation used MagicArena, with experienced image-generation evaluators scoring quality and usability across e-commerce, marketing, film and television, gaming, office productivity and fashion design. The company reports gains over Seedream 5.0 Pro in both measures across every category, with usability rates in film and television and office productivity more than twice those of the previous version.
In ByteDance’s separate expert pairwise evaluation, reviewers scored Good/Same/Bad preferences for text structure and realism in text-to-image and image-to-image tasks. The reported scores were positive against the predecessor in all four comparisons, reaching +55.62% for image-to-image realism. Comparisons with GPT Image 2.5 Sunburst and Nano Banana 2.1 also favored Seedream 5.5 in text structure across both tasks and showed a substantial text-to-image realism advantage.
Text-rendering tests covering Chinese and English, including dense small text, produced company-reported severe and minor text-corruption rates of 0.4% and 1.7%, respectively. Corruption covers wrong or missing characters and disordered character structure.
The Chinese-language announcement says Seedream 5.5 is gradually rolling out through Jimeng’s web creation and canvas tools, Doubao’s app and web image generator, Doubao Work skills, and Volcengine Ark’s visual-model experience center.
ByteDance plans to support more complex contextual inputs and creative instructions, expand the model’s use in agent workflows, and release a Pro version. The current release’s production significance lies in the work between generated images: preserving approved regions during revision and supplying transparent assets that can move directly into compositing.
英伟达 GR00T 2：机器人看人类视频学干活
NVIDIA 推出 GR00T 2，让机器人通过人类视频学习
NVIDIA 于 2026 年 11 月 9 日推出 GR00T 2，发布地点是奥斯汀举行的机器人学习大会（CoRL）。NVIDIA 报告称，这款部分通过人类视频训练的模型，在目标机器人的训练数据中未出现的 50 项任务上，实现了 40% 的零样本成功率。
GR00T 2采用一个基于 Cosmos-Predict2.5-2B 构建、拥有 20 亿个参数的 Transformer。它结合近期的相机视图、机器人状态和书面指令，生成预期的未来视图和下一步动作序列。一个预训练检查点即可支持平行夹爪机械臂、灵巧手和人形机器人。尽管 NVIDIA 将该模型介绍为开放模型，但其研究页面称，代码将于 12 月开放。
预训练数据包括 402,730 小时的第一视角人类视频、10,184 小时涵盖 2,897 项任务的手持夹爪演示，以及 23,500 小时的机器人遥操作数据。Human-1M-100K 数据集开放其中 100,000 小时的人类数据，包括视频、语言标注以及头部和手部运动数据，并采用 CC BY 4.0 许可，允许商业和非商业用途。
跨机器人形态测试展示了技能如何在不同机体之间迁移，以及如何利用有限的机器人数据进行适配。NVIDIA 报告称，通过人类演示学会的滤纸布置任务，在没有针对该任务进行遥操作的情况下，成功率达到 65%；从手持夹爪迁移到 YAM 机械臂的玩具自卸卡车组装任务，成功率达到 60%。在 Dexter 评测中，GR00T 2 在 40 项任务上至少各完成了一次试验。每项任务提供一次演示时，平均成功率达到 71%；提供五次演示时，平均成功率达到 95%。这五次演示为每项任务提供的机器人数据不到三分钟，随后在一台 RTX 5090 上进行约 30 分钟的适配器微调。
技术白皮书对五次演示的结果作了限定：受评任务是从模型能够通过五轮训练、以至少 80% 的成功率学会的任务中选出的，因此，这一平均值在一定程度上是由任务选择方式决定的。
仿真基础平台是 Isaac Lab 3.0，自 2026 年 10 月 20 日起已全面开放。它通过统一的任务 API 支持物理、渲染和可视化后端；Newton 支持刚体、可变形物体和粒子，而且许多工作流程无需安装完整的 Isaac Sim 即可运行。在一项基于视觉的灵巧操作基准测试中，NVIDIA 报告称，Isaac Lab 3.0 使用 Newton 时的训练吞吐量最高可达 Isaac Lab 2.3 使用 PhysX 时的 2.5 倍。
AlphaNIST将大规模仿真实践用于 NIST Assembly Task Board 的 20 项任务，其中包括尺寸小至 4 毫米的零件。一个共享策略仅通过是或否的成功奖励进行训练。NVIDIA 报告称，该策略的训练成功率为 98.2%，而针对每项任务分别训练的策略为 81.4%；前者使用的 GPU 时间约少 39 倍。这些结果来自仿真训练。
AlphaNIST 背后的方法是由华盛顿大学和 NVIDIA 开发的 成功引导采样。该方法将练习重点放在接近策略能力边界的配置上，尤其是策略大约有一半概率能够解决的配置。该方法的作者报告称，这种方法可有效扩展至超过 100 万个仿真环境，并能将基于 RGB 图像的装配策略零样本迁移至真实的 UR5e 机械臂。
Isaac Lab-Arena提供评测层。这个开源框架通过模块化组件构建可重复的评测，并在 GPU 加速的仿真环境中运行；故障分析还能指出哪些条件需要再次训练。NVIDIA 表示，其 0.4 版开发者预览版即将在 CoRL 发布。
DexBench v1.0由 RLWRLD 与 NVIDIA 开发，包含零售、物流、装配线以及酒店餐饮场景中的 15 项任务。来自加州大学伯克利分校和 NVIDIA 的 VABAR涵盖工具分类与悬挂、齿轮啮合、注射器包装，以及电缆插入和布线，并配有一个包含 33 项技能的库。VABAR 项目页面称，目前的两个难度等级构成验证集，更难的测试集仍在开发中。
加州大学伯克利分校的 Graph-as-Policy框架使用多个智能体组合出可解释的机器人技能有向图，在仿真中演练，并在失败后调整图的结构和参数。该项目报告称，经过十轮演练，其 Make Popcorn 任务的表现从最初的 33% 提升至仿真中的 94%，并在真实机器人上达到 90%，即 20 次运行中成功 18 次。目前的基准测试侧重于准静态操作；可变形物体、动力学和移动目标仍属于未来工作。
GR00T 2 使用推理智能体安排技能顺序、监控进度，并在零件掉落或工作停滞时调用恢复程序，而模型则负责接触丰富的操作任务。NVIDIA 报告称，在使用纠正数据和该智能体的情况下，机器人在每种玩具各进行 100 次试验、且每次均从随机倾倒的零件开始时，成功组装了 98 架玩具飞机、97 辆玩具拖拉机和 97 辆 Hot Wheels 汽车。仅进行监督微调时，相应数量分别为 69、51 和 23。
白皮书还单独报告称，在其长时程后训练基准测试中，平均成功率约为 70%。在通过视觉引导进行垃圾分类的任务中，机器人在一个回合中等待智能体的时间中位数占 38%，这揭示了智能体监督带来的延迟成本。
NVIDIA表示，Nexuni正在测试GR00T 2在洗衣处理任务中的应用，Foxconn正将其用于服务器制造自动化，Siemens则在评估工业应用场景。目前披露的工作仅处于测试和评估阶段，尚未宣布投入生产。
白皮书总结指出，可靠性分别依赖于预训练所提供的广度、纠正性示范和强化学习所带来的精确性，以及智能体所具备的记忆、验证和恢复能力。因此，对工业用户而言，有参考价值的衡量标准是围绕GR00T 2的完整工作流程，以及每项任务所需的监督。部署仍面临一项挑战：在保持已报告可靠性的同时，减少NVIDIA自身结果中记录的每项任务适配时间和智能体等待时间。
谷歌 TPU 8tx 与 8i Max：一年出齐四款 TPU
Google adds TPU 8tx for confidential AI and a two-chip TPU 8i Max, its third and fourth TPU systems this year
- TPU 8tx uses the same silicon as the TPU 8t training chip, paired with 5th Gen AMD EPYC CPU hosts, and keeps data and model weights isolated during processing [1].
- TPU 8i Max packs two TPU 8i chips into one package, doubling compute, SRAM and HBM bandwidth versus TPU 8i in the same rack footprint [1].
- Both are reconfigurations of existing silicon, giving Google four TPU systems announced in 2026 [1][3].
- Google says TPUs now serve nearly 100 customers, including frontier labs, trading firms, retailers and universities, but named none [1].
- Google gave no general-availability date, pricing or regions for either system [1][5].
Google added two systems to its eighth-generation Tensor Processing Unit line on Monday: TPU 8tx, a confidential-computing system that runs Google's training silicon alongside AMD processors, and TPU 8i Max, which combines two inference chips in one package for long-context work [1]. With the TPU 8t and TPU 8i unveiled at Google Cloud Next in April, Google now has four TPU systems announced in one year, which it called a first [1].
Amin Vahdat, Google's chief technologist for AI and infrastructure, said TPUs now serve nearly 100 customers across frontier labs, trading firms, retailers and universities [1]. Google did not name them and did not give an availability date, pricing or regions for the new systems [1][5].
TPU 8tx: training silicon in a secure enclave
TPU 8tx uses the same silicon as TPU 8t, the training chip, paired with 5th Gen AMD EPYC CPU hosts [1]. Google built it for governments and regulated enterprises that want to run frontier or proprietary models and agents on sensitive data, with hardware-level protection that keeps data and model weights isolated while they are processed [1][2]. Data in use normally has to be decrypted in memory, which is the gap confidential computing closes [3].
Google pitches the system for inference and small-scale training on one box: customers can fine-tune on private data, distill large models into smaller ones and train agents with reinforcement learning without moving data between clusters [1][2]. It runs on Boardfly, the low-latency network Google built for TPU 8i, and Google claims up to 2x lower latency than TPU 8t with a near-zero performance penalty from the security features [1][3]. Google published no absolute latency or throughput figures.
TPU 8i Max: two inference chips per package
TPU 8i Max puts two TPU 8i chips on a single package, doubling compute, on-chip SRAM and high-bandwidth-memory bandwidth relative to TPU 8i and fitting twice the silicon into the existing rack footprint [1]. Google aims it at workloads where memory is the bottleneck, such as many agents working across a large codebase or hours of video [3], and at the million-token context windows its Gemini models support [2].
Google's April specifications give the baseline: each TPU 8i carries 288 GB of HBM, 384 MB of SRAM and 8,601 GB/s of HBM bandwidth [4]. Google did not publish absolute figures for the Max package.
Packaging over new chips
Neither system is a new chip. Google is changing packaging, networking, memory and security around existing silicon, and Vahdat described TPUs as modular building blocks within its AI Hypercomputer that let it reconfigure systems quickly [1][3]. Google said it uses its AlphaChip system and Gemini to speed up chip layout [1].
The April chips have not yet shipped broadly. Google said then that TPU 8t and TPU 8i would be generally available later in 2026, claiming up to 2x better performance per watt than the previous Ironwood generation [6].
The software push against Nvidia
Much of the announcement targets the software lock-in that keeps customers on Nvidia GPUs. Google open-sourced TorchTPU last month so PyTorch and Hugging Face models run natively on TPUs, with TPU backends for the vLLM and SGLang serving frameworks [1]. It also previewed a TPU agent suite that plugs into Claude Code, Cursor and Google Antigravity to handle onboarding, bottleneck diagnosis and kernel tuning, and released an open-source XLA Serving Engine that serves models across TPUs, GPUs and CPUs [1]. Google said Cloud Pathways and dynamic slicing are now generally available on Google Kubernetes Engine, claiming up to 5x faster job startup and 4.5x faster fault recovery for the latter [1].
The new hardware closes gaps rather than opening a lead. Nvidia already offers confidential computing on its GPUs, and it has strengthened its inference line through a licensing and acqui-hire deal with Groq [3]. Google is also spending heavily on the data centers that will house its TPUs, including a $17 billion campus in Cheyenne and a $15 billion Malaysian commitment, and last week it backstopped Anthropic's Kentucky lease with TeraWulf.
Adobe 牵手迪士尼：用 Firefly 定制迪士尼专属模型
Disney and Adobe Announce AI Alliance with Custom Firefly Models
Adobe announced a multi-year strategic alliance with The Walt Disney Company on November 9, 2026, becoming an Official AI Innovation Partner of Disney. The agreement includes custom Firefly Foundry models and technical collaboration across Disney’s theme-park, studio and sports operations.
Announced at Adobe MAX, the agreement gives Disney early access to Adobe’s latest AI, creative and media technologies, including Firefly, Firefly Foundry, Premiere and Photoshop. Dedicated Adobe engineering and research resources will support Disney’s production pipeline, pairing the software with direct technical assistance.
Custom Models Built Around Disney Properties
Firefly Foundry, Adobe’s offering for developing proprietary brand-specific models, is the technical center of the alliance. Disney will work directly with Adobe on generative AI models tailored to its properties.
Adobe describes the models as human-centered and artist-led, customized on Disney properties and trained on character attributes, environments and visual styles. Its stated goals are commercially safe, production-ready outputs, faster creative work and the capacity to develop more creative projects.
Imagineering’s Existing Firefly Work
Adobe first announced the Imagineering collaboration on June 16, 2026, for design and pre-production visualization across Disney Parks and Experiences. Walt Disney Imagineering Research & Development is already using Firefly Foundry in theme-park development. A sketch-to-image model turns rough hand-drawn concepts into fully rendered 2D concept art, while a custom image model generates on-brand, franchise-accurate assets for properties including Mickey & Friends, Frozen, Moana, Lilo & Stitch and Cars.
The sketch workflow is designed to help designers compare and refine different creative directions. Adobe’s June announcement also listed a 3D modeling capability as forthcoming, intended to convert 2D renderings into detailed prototypes for planning builds, estimating materials and coordinating with engineering teams.
Disney described the training approach on June 18, 2026: the underlying model uses licensed and proprietary assets, followed by customization with more than 70 years of Imagineering concept art and technical designs. Kyle Laughlin, senior vice president of Walt Disney Imagineering Research & Development, said the project was intended to accelerate iteration while respecting Disney’s creator-driven process. “What’s most exciting is that it allows months of work to be compressed into days and weeks,” he said.
StudioLAB and ESPN Edge
Adobe is also joining The Walt Disney Studios’ StudioLAB program as an innovation partner. StudioLAB brings creative and technical staff together to prototype ideas and turn experiments into proven tools and storytelling opportunities; Adobe will work with Disney creatives and technologists to explore and expand internal creative workflows.
At ESPN Edge, ESPN and Adobe will examine AI applications for sports-fan engagement and content delivery, as well as new ways to use ESPN’s archive. Co-innovation projects with both StudioLAB and ESPN Edge are already underway, with additional activations and experiences planned during the alliance’s term.
The agreement also has a promotional dimension. Adobe will increase its presence across Disney’s entertainment and sports media platforms, while global marketing activations, industry-event collaborations and customer storytelling will involve Walt Disney World Resort, Disneyland Resort and Disney Cruise Line.
For Adobe, the alliance creates a broader production role and a prominent showcase for Firefly Foundry; for Disney, it adds dedicated engineering support and models built around its own properties. The practical test is whether Adobe can adapt the property-specific approach used at Imagineering to the different needs of StudioLAB’s internal workflows and ESPN’s content delivery.
Mistral Large 4 开放权重：1 万亿参数，编码基准领跑开源
Mistral于2026年11月9日开放Mistral Large 4完整模型权重，该1万亿参数多模态模型在FrontierCode 2.0基准取得19.7%得分。
根据官方发布，Mistral Large 4此前已进入公开预览阶段，经过三周强化学习后，编码和防御性网络安全能力显著提升。模型在Cognition的FrontierCode 2.0基准上得分19.7%，在Terminal-Bench 4.0上从预览版的28%提升至41%，在Artificial Analysis Cyber Index上达到51%。它支持低、中、高三种推理模式，默认高模式用于高精度场景。
模型通过持续强化学习实现token效率提升，在Artificial Analysis Coding Agent Index中得分48%，但平均每任务token消耗比同类领先模型低47%。三种推理模式允许开发者根据任务复杂度选择，降低推理成本。欧洲基础设施从头训练且 lineage清晰，为自托管部署提供基础。
模型在Surge Code盲测中被偏好超过GLM-5.3、Kimi K3和DeepSeek V4.1 Flash，仅次于Claude Opus 5.5。
开放权重结合修改版MIT许可证，使商业使用成为可能，但MaaS服务需单独授权。这降低了开发者运行复杂编码代理的门槛，同时在防御性网络安全任务中减少不必要拒绝。Hugging Face平台提供下载，支持vLLM和SGLang部署。
从已有基准数据看，Mistral Large 4在开放模型编码和网络安全领域形成局部领先，但token效率优势能否转化为大规模采用，仍取决于实际部署成本与生态支持。欧洲训练背景可能吸引注重数据主权的机构，但与闭源前沿模型的整体差距仍需持续观察。