Kimi K3.1 Pro, MiniMax M3.1 Flash weights, and Anthropic-managed Claude on Google Cloud goes GA

Kimi K3.1 Pro is a post-trained version of Moonshot AI's K3, a 3T-class mixture-of-experts (MoE) model. It has 2.8T total parameters with 104B active per token, a 1M-token context window, and takes text and images as input. A Highspeed variant that runs 2-3x faster shipped alongside it, and both are live today in the Kimi app, Kimi Code, and the Kimi API under the model name kimi-k3.1-pro.

Moonshot reports 65.7%Vendor claim on Terminal-Bench 4.0, ahead of GPT-6 Astra (58.2%)Vendor claim and just behind Claude Opus 5.5 (66.4%)Vendor claim. API pricing drops to $2 per million input tokens and $10 per million output tokens, cheaper than K3. Once the weights land, teams get another frontier-level coding agent model they can host themselves. Moonshot also says it stripped cyberattack-related data from training to keep the model's hacking ability from growing.

MiniMax M3.1 Flash is a multimodal MoE model aimed at coding, agentic work, computer use, and security tasks. It has about 428B total parameters with 23B active per token, accepts text, image, and video, and handles up to 1M tokens of context. It uses sparse attention throughout, was trained with quantization in mind, and ships with a DSpark draft head for speculative decoding. You can serve it with SGLang or vLLM.

hf download MiniMaxAI/MiniMax-M3.1-Flash --local-dir MiniMax-M3.1-Flash

MiniMax cites 61.7 on OSWorld 2.0Vendor claim, 73.7% on DeepSWE v1.1Vendor claim, and 76.2 on BankerToolBenchVendor claim. Its own table also shows 38.1 on Terminal Bench 4.0Vendor claim, below DeepSeek V4.1 Flash-1017 (42.3)Vendor claim. With only 23B active parameters it should be cheaper to serve than models in the same performance band, but the license is the MiniMax Community License, not open source, so read the terms before any commercial use.

Claude Models on Google Cloud (Anthropic managed) lets you call the Anthropic-operated Claude API directly from a Google Cloud project. Authentication uses Google Cloud IAM roles such as roles/claude.user instead of API keys, and usage is billed on your Google Cloud Marketplace invoice and draws down existing Google Cloud commitments. Features like Managed Agents (beta), code execution, Agent Skills, the Files API, and prompt caching arrive the same day they do on the native Claude Platform; in Python you use the AnthropicGoogleCloud client.

pip install -U "anthropic[google_cloud]"

Until now, running Claude on Google Cloud meant the Google-operated Vertex offering, where new features often showed up late. Now teams can keep their Google Cloud commitments and IAM setup and still get the latest betas right away. Fast mode for Opus 5.5 is also GA here, with up to 2.5x faster outputVendor claim at $8 per million input tokens and $40 per million output tokens. Anthropic, not Google, is the data processor, though, so if you have strict data residency requirements or need Model Armor, stay on the Google-managed offering.

Anthropic released Claude API version 2026-10-21, the first new version since 2023-06-01. Under it, sending temperature, top_p, or top_k returns a 400 error even at default values, and so do document blocks in assistant messages and citations in user messages. New major SDK versions that send this version by default (Python 2.0.0, TypeScript 1.0.0, Go 2.0.0, Java 3.0.0, and others) shipped alongside it, as did ant CLI 2.0.0.

pip install -U "anthropic>=2.0.0"

Upgrading the SDK alone changes your API version, so teams with automatic dependency updates are affected immediately. Code that pins output with sampling parameters, proxies that read the X-Api-Key header, and scripts that use old Admin API field names all need changes. The SDK license also moved from MIT to Apache 2.0. On Amazon Bedrock, the new version works only with Claude 4.7 and later models.