Microsoft's Decision-1 and Claude Code Projects hand agents a referee and a project plan

Microsoft-Decision-1 is a tiny, very cheap model built to make yes or no calls for your agents

If you build agents or automations, a surprising share of your model bill goes on small judgement calls: is this email urgent, which tool should handle this request, did that step actually succeed? Microsoft-Decision-1 exists for exactly that job. Rather than writing open-ended text, it takes a fixed set of options and returns a calibrated probability for each one, covering yes or no questions, multiple choice, ratings and rubric-based grading of other agents' work.

It is post-trained from Alibaba's open Qwen3.5-9B, costs $0.042 per million input tokens with output free, and is available in Microsoft Foundry and on OpenRouter with a 32K context window. Microsoft says it topped a 36-benchmark comparison of nearly 150,000 questions, runs around 35 times faster than GPT-6 Sol and changes its answer on just 1.3% of reworded inputs, while Xbox Research reports similar labelling quality at 200 times lower cost. Those are Microsoft's own numbers, so test it on your own routing tasks before retiring a frontier model. It also lands in a fast-growing category: Liquid AI opened its d1 decision models last week and Inception's Mercury Decide reached OpenRouter on Thursday.

Office in Copilot starts rolling out to Frontier users

Microsoft first described putting Word, Excel and PowerPoint inside Copilot in late September, and on Friday the official Copilot account confirmed it is now rolling out in Frontier, the company's early-access programme. In practice you can ask Copilot to create or edit a real document, workbook or deck and keep refining it in the same conversation, rather than copying text between apps. The same account also promoted the new Copilot Home, which merges Chat with Cowork, Microsoft's mode for delegated multi-step tasks. If your organisation is enrolled in Frontier, now is the time to try it; everyone else should expect wider news around Microsoft Ignite in mid-November.

Claude Code Projects opens to every Pro and Max user on the waitlist

If you signed up for the Claude Code Projects waitlist, check your account: Anthropic says it has now let in every Pro and Max user who joined it. A project is an ongoing conversation in which Claude scopes a piece of work, splits it into tasks, runs each one as its own thread in parallel and then pulls the results together, and you can steer it from your phone while it keeps going in the background.

Until now it had been limited to a select group since its beta launch in mid-September. For solo developers and small teams it is the closest thing yet to handing Claude a whole project rather than one prompt at a time, and the developer team has posted a four-minute walkthrough to get you started.

Claude Managed Agents can now write and run their own multi-agent workflows

Developers building large jobs on the Claude Platform can now let a lead agent write its own plan and fan it out across many agents. With dynamic workflows the agent writes a small program that runs agents in phases, for example reviewing hundreds of documents, and the server runs it in the background before combining the results. Each run can use up to 64 threads at once and 1,000 agents in total.

There is no separate fee: you pay normal token rates plus the existing $0.08 per session-hour, so a full 24-hour run adds about $1.92 in runtime. It builds on the Managed Agents automations guide published on Thursday. One early write-up warns that a run can report "completed" even when some of its work failed, so build your own checks into anything that matters.

ChatGPT dots can now be set up on your phone and hand work to Codex

Dots, the personalised ChatGPT assistants OpenAI unveiled at DevDay, are a lot more useful if you live on your phone. You can now create a dot directly in the ChatGPT app on iOS and Android rather than needing a desktop, and a dot can start new Codex tasks and follow up on existing threads, drawing on your ChatGPT conversations and Codex history for context. It can also review and edit your scheduled ChatGPT automations, and cloud Codex tasks keep running when your computer is off.

The catch for UK readers: on Pro plans the rollout currently excludes the UK, the EEA and Switzerland. Business Premium and Enterprise are rolling out worldwide, with Enterprise switched off until an administrator enables it.

Codex starts predicting your next message

Codex can now suggest what you are likely to say next, based on the conversation so far and how you usually talk to it. When it finishes a response, a suggested follow-up appears in the message box; press Tab to accept it, edit it if you like, then send it yourself, as nothing goes automatically. It works in the Codex desktop app with GPT-6 Astra or GPT-6.1 Sol for adults on personal Pro accounts, and it is on by default, with a toggle under Settings, General, Composer if you would rather not see it. Over a long coding session it could shave real time off the back-and-forth.

Codex on Windows gets a faster, safer sandbox built on Microsoft Execution Containers

Running a coding agent on Windows has always meant trading safety against convenience. Codex now has a sandbox mode built on Microsoft's Execution Containers that needs no administrator approval, separate Windows accounts or firewall rules, enforces network rules more strictly and gives finer control over which files a command can touch. It needs a compatible, up-to-date Windows 11 PC, and the desktop app will prefer it automatically for personal accounts where supported. Watch for background development servers stopping when the command that launched them exits.

Qwen-Image-2.1-Turbo brings fast, open-weight image generation and editing

For anyone running image generation locally or on their own servers, speed is usually the bottleneck, and this release tackles it head on. Qwen-Image-2.1-Turbo is an accelerated checkpoint of Qwen-Image-2.1 that creates and edits images in just eight denoising steps, which means far quicker renders on the same hardware. The weights are on Hugging Face and load in Diffusers, with output up to 2752 by 1536 for widescreen images.

As with the original, it ships under the Qwen Research Licence rather than a permissive commercial licence, so check the terms before using it in client work. Separately, Qwen is offering a week of free access to Qwen3.8-Max, Qwen3.8-Flash and Wan3.0 through GMI Cloud, a cheap way to try its latest models.

Grok Bot now gets its own email inbox

Agents keep getting stuck at the same point: signing up for anything needs an email address and a verification code. Grok Bot now has its own inbox on mail.grokbot.com, separate from your personal email, which it can use to sign up for services, receive codes and receipts, contact businesses for you or arrange meetings. It is rolling out to users now, and on team accounts an administrator has to switch it on first. SpaceXAI has not said which plans include it or whether outgoing mail needs your approval, so keep it away from anything sensitive until those details are clear.

Brand Kit keeps every ElevenCreative image and video on brand

If you produce marketing content with AI, you know the pain of re-explaining your colours and fonts every time. ElevenLabs' new Brand Kit stores your logos, colours, fonts, reference images and brand rules as one workspace object, and you tag it with @ in Image and Video, Flows or the ElevenCreative agent so every generation follows the same guidelines. Because kits live in the workspace, a whole team can generate from them and agencies can keep one per client. ElevenLabs says brand voice and sound will follow.