ERNIE 5.5 Flash Preview builds a full Unreal game, Antigravity agents get a browser, and Fractile raises $720M at $7B

Nov 11, 2026

MODELS + GAME DEV: Baidu's ERNIE 5.5 Flash Preview wrote a full Unreal Engine game in one 11-hour run, and the weights are due by Nov 25

Baidu released ERNIE 5.5 Flash Preview on Nov 11, the first post-trained model on top of the ERNIE 5.5 Flash Base checkpoint it opened on Nov 5. It is a mixture-of-experts model with 280B total parameters and 14B active per token, a 1M-token context window, and text-only input. Baidu pitches it at coding, 3D modeling, knowledge work and content creation, and says it trained it specifically on game development: game mechanics, gameplay logic, runtime state and asset pipelines, with support for Unreal Engine, Unity and Godot. Its headline demo is a third-person wuxia game written in C++ for Unreal Engine 5.8 in a single 11-hour agent run. Because the model can't see the screen, it turned playtest bugs such as a character floating above the ground into code tests. Other demos include an airport and airliner built in Blender by two agents. The API is live on Baidu's Qianfan platform and on OpenRouter at $0.28/$1.13 per million tokens, with a free variant listed for now. All benchmark numbers are Baidu's own, and its own chart shows it well behind the top closed models on Terminal-Bench. Baidu calls this an unfinished checkpoint and says the Apache 2.0 weights come by Nov 25, so treat the open-weight part as a promise until the files appear.

Source: https://ernie.baidu.com/release/ernie-5-5-flash-preview/

Nov 11, 2026

CODING TOOLS: Google Antigravity's agent now opens its own browser to test your app, with logins and downloads left to you

Google Antigravity said on Nov 11 that browser use is rolling out in its app, "closing the loop between researching, writing code, and verifying it live." According to the Antigravity docs, from app version 2.30.0 on personal Google accounts (Free, Plus, AI Pro and AI Ultra) you can just describe what to check, and the main agent starts a browser subagent in the built-in browser. It can open your localhost app, click through flows, read console errors and failed network requests, take screenshots and verify a fix before it calls the task done. Several browser subagents can run at once, each in its own tab and context window, and they run on a Gemini Flash model by default to save higher-tier quota. You choose what it signs into, and Google says downloads pause until you approve them. Business and enterprise accounts, Remote Agents and SSH workspaces don't get the built-in version yet. There, and in the Antigravity CLI, /browser drives your local Chrome 144+ through the Chrome DevTools MCP server. For anyone building web games or front ends with an agent, this is the step that catches the "it compiles but the button is broken" bugs.

Source: https://antigravity.google/docs/browser/

Nov 10, 2026

CODING TOOLS: Chrome, Edge, Firefox and MDN ship Modern Web Guidance 1.0 to stop coding agents writing legacy web code

Google's Chrome team announced the stable 1.0 release of Modern Web Guidance on Nov 10. It is a free, open-source set of agent skills that puts web platform best practices and browser compatibility data into your coding agent's context, so it reaches for native browser features instead of bloated JavaScript and heavy polyfills learned from old training data. Version 1.0 adds a modernization workflow that scans an existing codebase for safe upgrades, a larger database of expert-reviewed guides (the release notes count 203 guides covering about 250 features), and a public eval dashboard. Firefox and Edge now help run it alongside Chrome and MDN. Apple is not listed. Install it with npx modern-web-guidance@latest install, or through the plugin systems of the major coding agents, including Claude Code, Copilot CLI, Antigravity CLI and Vercel's npx skills, per the GitHub repo. Search runs offline with no API keys. The improvement figures so far come mainly from Google's own testing, so check the results on your own project.

Source: https://developer.chrome.com/blog/modern-web-guidance-v1

Nov 11, 2026

DEALS: Fractile raises $720M at $7B for memory-first inference chips, and Rox becomes a $1.3B AI agent unicorn

Two big bets on the agent era landed on Nov 11. London chip startup Fractile raised a $720M Series C at a $7B post-money valuation, co-led by Thrive Capital, Founders Fund, Lightspeed and Redpoint, with CapitalG, BlackRock, T. Rowe Price, DST Global and the UK's Sovereign AI fund also in, according to backers Balderton and Georgian. That is up from roughly $1B in May. Fractile stacks DRAM directly on top of its logic so large models can run fast at very long contexts, the exact load that long agent runs create. TechCrunch notes it has not shipped a production chip yet, and plans its first chips and systems next year. Separately, Rox raised a $100M Series B at a $1.3B valuation, co-led by General Catalyst and March Capital with Sequoia, Factory, Databricks Ventures and Prosperity7. Rox builds revenue agents that research accounts, generate pipeline and prepare outreach, and says its annual recurring revenue grew more than 10x in its first year in market (company figures).

Source: https://www.balderton.com/news/fractile-raises-720m-to-build-the-hardware-that-pushes-ais-frontier-forward/

Nov 11, 2026

AGENTS: Gemini API Managed Agents can now run their tools in your own sandbox on Vercel, Modal, Cloudflare and more

Google AI Studio announced on Nov 11 that developers building with Managed Agents on the Gemini API can now "bring your own sandbox" with self-hosted environments. Until now a single API call gave the agent a Google-provisioned Linux sandbox to run code, manage files and browse. Now Google's agent harness still plans and reasons in Google's cloud, while the code execution and file work happen on infrastructure you control. Google is launching it with seven partners, Vercel, Modal, Daytona, Fly.io, Google Cloud, Cloudflare and E2B, as AI Tracker summarizes. ORIENS reports that the connection is an outbound HTTPS link through the open-source Gemini API CLI, so you don't open inbound firewall ports, and that the feature is in preview and may change. For AI coders, that means custom images, GPU access, your own network rules and data residency without giving up Google's managed agent loop. You pay for the compute in your own environment.

Source: https://x.com/GoogleAIStudio/status/2120610219235660183

Nov 10, 2026

AGENTS: n8n Labs releases Nod, an open-source personal agent with its own computer that asks before it acts

n8n released Nod on Nov 10 as a research preview from its new n8n Labs team. It is a standalone app for Mac and Windows, separate from n8n and needing no n8n account, that gives an agent its own local virtual machine and browser. You ask in one chat, and it works through your connected apps (Gmail, Calendar, Drive, Slack) and on websites with no integration at all: cancelling a trial and asking for the refund, downloading supplier invoices each month, or pulling numbers from a dashboard that has no export. It is open source under Apache 2.0 and runs with the model you choose, through an API key, your ChatGPT plan or a local open model. n8n says a separate check confirms you asked before it sends, pays, posts or deletes anything, saved logins and cards are encrypted on your machine and entered through forms the model can't read, and every task ends with a cost receipt and a $5 check-in by default. n8n also reports a 93.2% score on the LongMemEval memory benchmark (company figure). It stops when your laptop sleeps, and n8n warns to expect rough edges.

Source: https://blog.n8n.io/meet-nod/

Nov 11, 2026

OPEN WEIGHTS: Runway opens Praxis-1, a 27B robot model trained on video, under Apache 2.0

Runway made Praxis-1 available to everyone on Nov 11 with open weights under Apache 2.0, including commercial use. Praxis-1 is a 27B-parameter mixture-of-experts diffusion transformer that predicts video and robot actions together, built on the same video pretraining as Runway's world models. Runway says it tops NVIDIA's RoboLab-120 leaderboard among world action models at 46.2% success across 1,200 episodes, ahead of FLUX 3 Action at 42.9%, and that fine-tuned on Google's RT-1 robot, which it never saw in training, it reaches 80.0% on SimplerEnv versus 52.4% for the original RT-1 policy. Those are Runway's numbers. The release includes inference and fine-tuning code with full and LoRA fine-tuning, and it is on GitHub and Hugging Face with LeRobot support, in vLLM-Omni, and on hardware from RTX 5090 workstations to Jetson Thor. The takeaway for builders: Runway's research says ordinary web video works as well as robot teleoperation footage for pretraining, which puts capable robot policies within reach of small teams.

Source: https://runway.com/news/company-news/praxis-1-is-now-available

Ship faster this week

Grab the free Agentdex starter kit with prompts and agent configs for Cursor, Claude Code and Codex.