@GoogleDeepMind
Gemini 4 Carbon rolls out to developers
The model targets long-running tasks. Cursor added it on launch day and published its benchmark results.
Google is rolling out Carbon through the paid Gemini API and to Google AI Ultra subscribers. It is designed for multistep coding, complex knowledge work and research that calls for repeated verification. Introductory API prices are $2 per million input tokens and $10 per million output tokens for an introductory period ending November 30.
@GoogleDeepMind
Gemini Robotics 2.5 broadens the range of physical tasks
DeepMind introduced a vision-language-action model with demos of folding laundry, serving pastries and sorting car parts. The company says the model works across different robot bodies, including humanoids and dual-arm gripper platforms.
@MiniMax_AI
MiniMax releases H3-Image weights
MiniMax says H3-Image can generate a 4096 × 4096 image in under 10 seconds, accepts up to 16 reference images and can edit existing images. The open-weight model is also available through the MiniMax API, MiniMax Code and Hailuo AI.
@OpenAI
OpenAI launches Businesses in ChatGPT in beta
Businesses can claim their profiles, update details and view aggregate metrics such as mentions and website clicks in Business Hub. Eligible U.S. ecommerce and online service businesses can also create a Business Agent that answers questions from information they provide.
@AnthropicAI
Anthropic begins roadshow for proposed IPO
Anthropic announced the roadshow for a proposed offering of Class A shares and applied to list on Nasdaq under ANT. The company said its SEC registration statement had yet to take effect at the time of the announcement.
@sundarpichai
DeepMind verifies key properties of compiled Signal protocol code
DeepMind introduced AlphaProof Code, a framework where researchers define the properties to prove and agents help construct formal proofs. Its Signal protocol study checked selected cryptographic properties against compiled machine code.
Cursor adds Carbon and reports a 60.4% benchmark result
Cursor says Carbon scored 60.4% at high effort, its highest result on CursorBench at the time.
llm-gemini 0.36 adds Carbon support
Simon Willison updated the Gemini plugin for his command-line tool and tried the same SVG generation task at low, medium and high effort.
Ripple 65-B Preview comes to Ollama
Ollama offers Ricursive AI’s 65-billion-parameter model for local use. The announcement describes it as an agentic coding model with up to a million tokens of context.
ARC-AGI-3 reports Carbon scores under two setups
ARC Prize reports 99.1% with the provider adapter at medium reasoning and 65.1% with the standard harness at low reasoning. The high-reasoning standard run was incomplete after some game sessions expired.
Perplexity brings Carbon to Computer
Perplexity added Carbon for Pro and Max subscribers. It can run persistent Teammates or one-off jobs in Computer. Perplexity reports a WANDR score of 0.857 at $5.86 per task.
Cursor Projects lets teammates’ agents collaborate
Teammates share a task board and project context while keeping their own chats and agents. Team members can be invited to the same project.
ChatGPT Business gains access to more third-party tools
OpenAI expanded Sign in with ChatGPT to let Business subscribers connect to tools including Notion, v0, Amp, Kilo Code, Warp and OpenClaw.
OpenAI publishes MCP extensions for ChatGPT
The public repository shows how plugins can use the ChatGPT sidebar, render custom file viewers and let users find and reference resources from the composer.
Claude expands support for MCP Apps
Claude expanded its MCP Apps integration to support app tools, side panels and embedded sites in the conversation interface.
Anthropic updates its Fable 5.2 prompting guide
The updated guide covers code ownership, approvals, progress reports and subagents. Simon Willison highlighted its suggestion to clarify when the model has authority to improve existing code.
Using several devices to locate a mystery beep
Simon Willison built a tool with Claude Code that recorded from several devices and used timing differences to estimate the source. He says it narrowed the search to about 40 centimeters, where he found an old detector with a low battery.
Claude adds collaborative whiteboards and spreadsheets
Whiteboards and Sheets are in beta for paid plans. Dashboards becomes Claude Data and is generally available, with certified queries and live sharing. Enterprise admins control feature access.
Vercel reports early Security Review figures
CEO Guillermo Rauch says the tool identified 57,897 issues in under a week, classed 6,029 as high or critical and patched 14,665 with an agent. The figures come from Vercel’s product reporting.
Hugging Face reports 20 million builders
CEO Clément Delangue says the Hub now stores 300 PB of models, datasets, agent traces and reinforcement learning environments, including 9 PB added by users the previous week.
vLLM moves its CUDA build to the LibTorch Stable ABI
vLLM detailed the migration of its CUDA extensions. PyTorch upgrades within the stable ABI’s supported range can reuse an existing vLLM build.
Developer publishes a two-step H3-Image LoRA
The Hugging Face release offers a standalone LoRA and merged weights. Its author says it can produce native 4096 × 4096 images with two denoising evaluations.
Hedra launches Model Factory for post-training
Hedra combines model post-training and inference in one platform, targeting teams in areas such as robotics and manufacturing that work with multimodal data and simulation environments.
Hugging Face shares a guide to training mixture-of-experts models
The handbook starts with a dense baseline and uses Megatron-LM code to explain expert selection, routing, load balancing and communication across GPUs.
Raschka tracks hybrid attention in open models
While surveying open language models, Sebastian Raschka observed that more of them are combining full attention with other sequence-processing mechanisms.
Marin team explains design choices behind its 535B model
Percy Liang shared the rationale behind Marin 535B-A23B, including activating eight of 384 experts and choices around GatedNorm, sconv, XSA and the learning-rate schedule.
Adobe introduces Firefly Design in beta
Adobe brings vector, raster, motion, 3D and print work onto one canvas, with precise editing tools and features for client review and delivery.
Pika uses Seedream 5.5 for layer-based edits
Pika says Image Studio can split an image into more than 100 layers, target areas for iterative edits with pins or paint and export transparent PNGs.
Synthesia demos a digital likeness you can talk to
Synthesia says users can record a short clip to make a digital likeness, speak with it in real time and share it by link.
Ideogram 4.5 leads a facial-identity editing test
Ideogram cited a Rapidata benchmark in which its model placed first among 20 for preserving facial identity after five successive edits.
Premiere beta adds object removal for video
An Adobe MAX demo showed a mask tracking an object through a shot, then Generative Remove filling the area left behind. The feature is in public beta.
Adobe and Disney announce a multiyear creative technology alliance
The companies plan to bring tools including Firefly Foundry, Premiere and Photoshop into Disney’s production workflows, with engineering and research support for proprietary models.
Adobe previews Firefly Builder
An Adobe MAX preview showed a tool that turns descriptions and reference material into app prototypes. Firefly Builder is in private beta.
Maestro 4.1 adds H3-Image
Maestro’s developer says version 4.1 adds MiniMax H3-Image for image generation and editing in the local creative studio.
NVIDIA shows GR00T 2 running on a robot
Jim Fan says GR00T 2 runs entirely on the robot’s onboard compute and can chain dexterous manipulation with whole-body control over longer tasks.
Alpamayo 2 Edge targets real-time reasoning in vehicles
NVIDIA showed the model planning driving trajectories in real time on DRIVE AGX Thor and producing explanations when needed. Examples include yielding to pedestrians and making room for an approaching vehicle.
Flide reuses shared prefixes for decision models
The paper keeps branching states in a single inference stream to reduce recomputation while matching independent forward passes. The authors report speedups of up to 9.5 times on their evaluated workloads.
Longtail-SWE tests repository repairs in 13 languages
The authors assembled 202 tasks from real GitHub issues and pull requests, with executable environments and tests covering languages that are less common in coding benchmarks.
Three researchers launch STABL
Sayash Kapoor, Rishi Bommasani and Arvind Narayanan launched a lab to study open-ended agent research, AI capability measurement and the bottlenecks that shape adoption.
Mercor raises $750 million in Series D
Mercor says the round values it at $20 billion and will fund enterprise agent evaluation, model post-training and simulated work environments.
Jason Wei tracks which benchmarks frontier labs report
He tallied evaluations reported in 2026 model releases. HLE, SWE-Bench Pro, DeepSWE and Terminal-Bench were widely used in his sample.
Hassabis shows Carbon porting an old game
He says Carbon analyzed the original Theme Park binaries, added a DeepMind-themed ride and produced a browser version.
Google uses multiple agents to build a pelican game
Sundar Pichai demonstrated Antigravity’s /swarm preview. Builder and reviewer agents iterated across Blender and Three.js and checked the rendered results.
Mollick publishes a fantasy transit simulator
Ethan Mollick used Fable 5.2 and several agents to make the playable Dread & District. Players run a fantasy transit network that connects villages and carries passengers.
Creator builds an interactive ink demo with Fable 5.2
The demo turns typed characters into simulated ink. Moving the pointer disturbs the fluid and triggers audio feedback.
Creator shows a chain-reaction animation built with Carbon
The creator says Carbon wrote a CUDA/C++ path tracer and combined it with MuJoCo rigid-body simulation for a roughly 60-second chain-reaction sequence.