A design platform shipping its first in-house image model. An enterprise chat product ending the choice between two interfaces. A 1M-token-context model whose open weights are due by November 4. Three things from the past week.
Claude Enterprise: Chat and Cowork Are Now One
Anthropic announced on October 23 that it has unified Claude Chat and Claude Cowork into a single interface for Enterprise customers. Chat and Cowork first merged for Pro and Max plan users in September. Enterprise is catching up now.
The change sounds minor. It is not. Before this, Enterprise teams chose their Claude interface before they started working. Chat for a quick question, Cowork for a full project. That split created a category error. Nobody knows at the start of a task whether it will turn into a long cloud job. Now they do not need to guess. A conversation starts as a chat. If Claude needs to run code or process files, it moves into a cloud session and keeps working, even after the user closes their laptop. The finished output comes back when it is ready: a document, an analysis, a deck.
Anthropic says each cloud session runs in its own isolated virtual machine with network egress off by default. Chats, projects, connectors, skills, and plugins all carry over.
The new admin control worth noting is Efficient mode. It locks the unified experience to Sonnet and Haiku models, keeping usage costs down for teams that do not need Opus-class output on every task. Admins also get model entitlements per role, spend limits at the organization, group, and member level, and pooled group budgets. Anthropic says usage-based pricing stays unchanged and that the newer models are cheaper to run: Opus 5.5 costs 40% less than Opus 5 on typical workloads, and Sonnet 5.5 costs up to 30% less than Sonnet 5.
Daniel Padilla, CTO of LabX, summed up what the merge fixed: “Bringing them together has made everything simpler, both for our users and administrators. The transition was seamless, nothing broke across our connectors, skills or MCPs.”
That quote is about operations, not outcomes, which is honest. The value is in the friction removed. Teams using Claude for content workflows already built around connectors get the most from this.
Why it matters: Every tool that forces users to pre-select a mode before they know what the task requires is adding cognitive overhead before the first keystroke. Removing that choice is not a small UX fix. It is the difference between a tool that fits the way people work and one that requires people to think about the tool before they can think about the work.
Find out more: Anthropic’s Claude Cowork and chat come together for Enterprise
Canva Ships Its First In-House Image Model
Canva released Canva Craft on October 22. It is the first image generation model built and trained from scratch by Canva Research, the company’s in-house AI lab. This is a bigger move than the headline suggests.
Craft uses a two-stage architecture. The first stage interprets a prompt and builds a compositional plan: which elements appear, their spatial relationships, scale, visual hierarchy, and any text that needs to appear. The second stage renders that plan into an image at up to 2K resolution across more than 30 aspect ratios. Canva says the two-stage approach is what lets users make targeted edits without changing unrelated parts of an image.
On DesignArena, the crowdsourced image model benchmark, Craft debuted as the fourth most advanced image generation model overall. It ranked first for Landscapes & Nature (out of 85 models). Canva says this makes Canva Research the world’s second most advanced AI lab for image generation as measured by DesignArena rankings.
Canva Craft is available to all Canva users globally. Access it from the Dream Lab tab in the bottom left panel. Developers can also use it through the Canva API. A Canva MCP Server connection is coming.
Canva says most image models are trained primarily on photographs, while Craft’s training places much greater emphasis on design. Its training data includes millions of human-made templates from Canva’s content library: posters, social posts, presentations, infographics, menus, and typography-rich layouts. Canva acquired Leonardo.ai in 2024. Two cost-optimized variants, Craft-fast and Craft-ultrafast, are coming.
Why it matters: Canva now owns a foundation image model it trained itself. That changes the economics. When your image generation runs on your own stack, you control quality, speed, and cost in ways you cannot when you are reselling someone else’s API. The strategic question is how far Canva pushes Craft across its own product, and how fast.
Find out more: Canva’s Introducing Canva Craft newsroom post
Kimi K3.1 Pro: Open Weights, 1M Tokens, Blender Scripts
Moonshot AI announced Kimi K3.1 Pro on October 23. The model is available now on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. Open weights ship by November 4.
The headline specs: 2.8 trillion total parameters, 104 billion active parameters, Mixture-of-Experts architecture, 1-million-token context window, multimodal. It is a post-trained version of Kimi K3 with substantially more post-training compute applied.
The practical pitch is 3D modeling and long-horizon coding. Kimi says K3.1 Pro can write and run Blender and CAD scripts directly from a text brief, sketch, or reference image, without MCP connectors. It also handles full-project builds in Unreal Engine, Unity, and Godot.
Finance is another stated priority. On a internal Moonshot benchmark designed to stress-test factual accuracy, K3.1 Pro’s claim-level hallucination rate came in at 8.3%, down from 23.4% for K3. Moonshot is clear that the examples are deliberately crafted to probe where models make errors, and that the rates do not represent overall error rates in daily use.
A Highspeed variant runs at 2-3x the speed of K3.1 Pro. Open weights release by November 4.
Teams comparing open-weight options can weigh it against other models with serious local-deployment numbers.
Why it matters: A 1M-token window matters most for exactly the use cases Kimi is pitching: full-project codebases, long financial documents, end-to-end 3D modeling workflows. Open weights at that window size mean teams can run those jobs on infrastructure they control. Whether the weights live up to the API numbers after November 4 is the question worth watching.
Find out more: Kimi’s Introducing Kimi K3.1 Pro on X, and the tech blog