Anthropic released Claude Fable 5.2 and Mythos 5.2, improving coding, reasoning, and vision capabilities. Fable 5.2 delivers up to 30% faster output. API pricing is $8 per million input tokens and $40 per million output tokens, down 20% from 5.1. Fable is available to paid users, while Mythos is available to organizations conducting cybersecurity or biomedical research. In Artificial Analysis tests, Fable 5.2 scored 65 at the highest reasoning effort.

Google introduced two configurations, TPU 8i Max and 8tx, targeting long-context inference and data-sensitive workloads, respectively. 8i Max packages two 8i chips into a single unit to boost compute performance and memory bandwidth. 8tx uses 8t chips and AMD EPYC hosts for confidential computing in areas such as government and finance. On the software side, TorchTPU supports running PyTorch and Hugging Face models, while the open-source TPU Sync can reduce KV transfer latency by more than 90%.

Meta introduced Muse Ember, a small model designed for agent tasks. It supports a 1-million-token context window and native multimodal understanding, and can be used for tool calling, coding subtasks, information extraction, and long-document processing. Meta Model API pricing is $0.10 per million input tokens and $0.50 per million output tokens; the model is also available on OpenRouter. Meta says Ember was trained using Muse Wonder Full.

ByteDance released the image model Seedream 5.5, with a focus on improving image-generation realism, text rendering, and multi-turn editing stability. It also supports direct generation of assets with transparent RGBA backgrounds. In successive editing tests, the model retained image details after 100 rounds of revisions. It is now available on platforms including Jimeng and Doubao, and has expanded to Dreamina, CapCut, and BytePlus. The API charges for image inputs and outputs, with a 2K image costing about RMB 0.28.

Google introduced Gemini 4 Argon, a model for robot control and physical reasoning that can be used to build simulation scenarios, annotate robot videos, and iteratively improve control strategies. After simulation practice, the model can guide Boston Dynamics' Spot through tasks, and it can also control robotic arms using visual coordinates. In one test, researchers generated simulation data with domain randomization from a single RGB image, raising the real-world success rate of a vision-language-action policy from 56% to 75%.

OpenAI has deployed Control Room, a centralized control system in its research infrastructure, to track model inference workloads and terminate tasks when safety or security risks arise. The system covers agent sampling and data processing during model training and evaluation, and works alongside monitoring systems. Once a decision to intervene is made, it can stop model outputs within 35 seconds. The system uses a fail-closed architecture and rejects inference requests if the control service itself fails.

NVIDIA is expanding its collaboration with South Korean technology agencies and companies. The plan is to connect Rebellions' R200 neural processing unit to the NVIDIA platform via NVLink Fusion and develop rack-scale systems using the MGX architecture, with deployment targeted for 2028. The partnership is also exploring disaggregated inference, with GPUs handling prefill and NPUs handling decoding. NVIDIA is also providing Dynamo deployment solutions for several South Korean open models, with inference performance on B200 GPUs improving by up to 2x.

AWS introduced TorchNeuron, a PyTorch-native backend that lets developers run existing PyTorch code on Trainium chips with minimal changes. It supports distributed training interfaces including DDP, FSDP2, and DTensor, and can use the Neuron Kernel Interface to optimize custom operators. AWS says it has validated 100,000 popular models and more than 800 machine learning architectures, and that training and inference workloads can share hardware capacity.

Microsoft is extending Copilot Chat usage restrictions to Microsoft 365 customers who have not purchased a separate Copilot license. The new restrictions will apply to organizations with up to 2,000 seats starting Nov. 2, and will extend to large enterprises on Nov. 9. Microsoft has not disclosed specific quotas, and enterprise administrators cannot view or adjust the limits.

Visa tested 22 AI models on 15 shopping tasks, with an overall pass rate of 57.2%. In 54% of failed attempts, the model still recommended buying an item even though the user's needs had not been met or recognized. Follow-up tests are planned to cover checkout and order placement.