Kimi releases K3.1 Pro open-weight model

Kimi releases K3.1 Pro, an open-weight model, while Canva launches Craft and Kevin Weil raises $215 million for Warp Sciences. Also inside: Anthropic IPO stakes, OpenAI's Decisions API, Google Cloud's Agent Compute Engine and updates to MCP and Copilot tools.

AI lab releases

Kimi releases K3.1 Pro open-weight model and a faster Highspeed version

Kimi launched K3.1 Pro, a 3T-class open-weight model whose full weights, the trained parameters anyone can download, are due by Nov 4. Kimi says it scores 65.7% on Terminal-Bench 4.0, above GPT-6 Astra at 58.2% and just under Claude Opus 5.5 at 66.4%. It can read 1 million tokens, or word pieces, at once, and a Highspeed version runs two to three times faster.

Why it matters: Developers may soon be able to run a model that rivals closed systems on coding tests on their own hardware.

Sources: Kimi

Canva launches Craft, its first in-house image generation model

Canva says Craft plans layout, scale and text before it generates an image. It outputs up to 2K resolution in more than 30 aspect ratios and was trained mainly on design templates rather than photos. It is open to all users through Dream Lab and ranks fourth on the DesignArena leaderboard.

Why it matters: Canva users get a native image tool aimed at text-heavy designs such as posters and menus.

Sources: TechCrunch, The Verge, The Decoder, CNBC

Acquisitions + deals

Reports detail who stands to gain from Anthropic's planned IPO

Eric Schmidt reportedly owns about 1% of Anthropic, half the 2% held by CEO Dario Amodei, according to Bloomberg. No venture firm holds more than 2%, Newcomer reports, partly because Anthropic wanted to limit any single investor's control. At the latest $965 billion valuation, Iconiq's stake is worth $17 billion and Menlo Ventures' $14 billion, Bloomberg says.

Why it matters: Spread-out ownership could shape who holds influence over Anthropic once its shares trade publicly.

Sources: TechCrunch, Newcomer, 9to5Google

Former OpenAI science chief Kevin Weil launches Warp Sciences with $215M

Warp Sciences has raised $215 million in a round led by Lightspeed Venture Partners. It will run a US chemistry contract lab that makes molecules for drug makers' programs. Data from its experiments will train its AI models, while robots learn lab tasks from chemists.

Why it matters: Weil's bet is that AI for science needs physical lab data, not just more software.

Sources: SiliconANGLE, Bloomberg

OpenAI and Poland's NASK agree to work on AI safety research

OpenAI signed an agreement with NASK, whose AI Safety Research Center acts as Poland's AI safety institute. The two will study ways to test advanced AI systems, find weaknesses and strengthen safeguards. CERT Polska, part of NASK, earlier used OpenAI models to help find six flaws in MikroTik RouterOS.

Why it matters: Poland joins the US and UK in giving a national institute a role in a frontier lab's safety work.

Sources: OpenAI

MCP + agent tooling

MCP TypeScript SDK 2.4.0 adds a Skills extension

Version 2.4.0 of the TypeScript SDK for MCP, a standard for linking AI apps to tools and data, adds Skills support. Servers can now list and serve reusable skill files, and clients check each file's size and digest when reading it. The method that attaches Skills to a server is marked experimental, and a validator update rejects some invalid schema IDs.

Why it matters: Developers building MCP servers can share packaged agent skills through a common interface.

Sources: MCP on GitHub

Copilot in Visual Studio adds pull request summaries and model choice

A Git agent in Copilot Chat can now summarize a pull request, with links to its comments and files. Bring Your Own Model, in preview, lets teams use providers such as OpenAI, Anthropic, Ollama and Microsoft Foundry. That feature now requires the new Agent (Preview) mode, and Ollama models must be added again after updating.

Why it matters: Visual Studio users can pick their own AI provider but should expect some setup work after updating.

Sources: The GitHub Blog

LangChain outlines how agents can search without overloading context

LangChain says agents struggle when they search clumsily, pull in too much at once, or pick up stale material. Its Deep Agents harness, the software wrapped around a model, matches file names and lines before reading a page at a time. For custom search tools, it suggests saving results to files and letting subagents read them.

Why it matters: Teams building agents get a concrete pattern for keeping a model's limited working memory focused.

Sources: langchain.com

Bayer builds multi-agent assistant for crop breeding workflows on AWS

Bayer Crop Science built the Agentic Decision Engine to answer plain-language questions about its machine learning workflows. A supervisor agent sends each question to a specialist agent that pulls context from the right system. Bayer and AWS say the first phase answers about 80 percent of questions without an engineer.

Why it matters: Large firms are starting to use agents to explain their internal data systems, not just to run them.

Sources: Amazon Web Services

SDKs + APIs

OpenAI makes its Decisions API generally available

The Decisions API, built on GPT-6 Luna, answers questions with set answer options and returns a probability, choice or score. OpenAI says it is up to 10 times faster than calling Luna through the Responses API. New Python and Node SDK releases change Decisions types, which can break code that assumes instructions are plain text.

Why it matters: Apps that make many small routing or labeling calls could cut delay and cost.

Sources: OpenAI, OpenAI on GitHub

OpenAI, Anthropic, Hugging Face and Alibaba donate tools to AI foundation

The Linux Foundation's Agentic AI Foundation took on four new projects, bringing its total to 10. OpenAI gave Open Responses, a shared spec for how apps call AI models, and Anthropic gave Agent Skills. An API Evangelist analysis finds only 26 of 120 OpenAI-compatible providers declare the Responses endpoint.

Why it matters: Developers could switch model providers with less rewriting if the spec gains wide adoption.

Sources: SiliconANGLE, API Evangelist

Google Cloud unveils Agent Compute Engine for hosting AI agents

The service runs agents in microVMs, small virtual machines meant to pair container efficiency with stronger isolation. Login credentials stay outside the sandbox, and a service called Agent Gateway submits them for the agent. Google says idle sandboxes are shut down and revived when needed, with rollout planned in the coming months.

Why it matters: Cloud customers running agents get a hosting option designed around containment and cost.

Sources: SiliconANGLE

AWS shares fix for failed SageMaker endpoint updates with auto scaling

Changing a SageMaker AI endpoint's setup fails in CloudFormation while auto scaling is still registered, AWS explains. Manual workarounds can leave endpoints without auto scaling or stacks stuck in a failed rollback. AWS offers a reusable construct for CDK, its code-based infrastructure tool, that removes and restores auto scaling around each update.

Why it matters: Teams serving models on SageMaker can update endpoints in one deployment without manual steps.

Sources: Amazon Web Services, The Information

Research

AWS guide shows multi-turn reinforcement learning on SageMaker using Wordle

Multi-turn reinforcement learning trains a model on one reward given after a full sequence of actions. In AWS's Wordle example, gpt-oss-20b rose from a 61% to an 84% solve rate on unseen words. The training run cost about $339 in tokens.

Why it matters: Companies can tune small, cheaper models for their own multi-step tasks without labeling each step.

Sources: Amazon Web Services, The Information, Reuters, The New York Times, Business Insider

Thanks for reading, and we will be back with more tomorrow.

---

About ScrutifAI: ScrutifAI is an independent third-party agency that tests software for AI agents, building a single access point where agents and people can access and use tested tools without learning each one individually, at a fraction of the cost of buying them directly.