Gemini 4 Carbon opens to developers with agent swarms, Qwen4-Max enters preview, and Mercor doubles to $20B

Nov 10, 2026

MODELS: Gemini 4 Carbon opens to every paying developer at $2/$10, and /swarm turns one prompt into an agent team

Google made Gemini 4 Carbon generally available on Nov 10, two weeks after it went only to cyber defenders in its Fairwind Program. Carbon is now on the paid tier of the Gemini API in Google AI Studio and Android Studio, in Gemini for business, and for Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Google Antigravity. Free API users and AI Pro subscribers don't get it yet. Until Nov 30 it costs the same introductory price as Gemini 4 Argon, $2 per million input tokens and $10 per million output tokens. After that it goes to $4/$20. Google says Carbon ranks #1 on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2x lower cost per task than the next best outside model. Those are Google's numbers. Google still recommends Argon for most coding and pitches Carbon for the hardest multi-step work, since it spends more tokens. The bigger news for AI coders is /swarm, now in preview in the Antigravity app and CLI. A coordinator agent spins up parallel workers in isolated git worktrees, plus adversarial verifiers that run tests and challenge every diff, and you can pin a different model to each role. For game devs, Google's showcase is a browser game in which a pelican rides a bike around its Mountain View campus. Google says a Carbon swarm built it, with Blender assets, Three.js and WebGL shaders, and Lyria music. Google also warns that swarms burn through quota much faster than a normal chat.

Source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/

Nov 11, 2026

MODELS: Alibaba puts Qwen4-Max in preview, aimed at coding and agent work, with no benchmarks yet

Alibaba's Qwen team announced on Nov 11 that "Qwen4 is coming" and that Qwen4-Max is now in preview on Alibaba's Token Plan, with early access in QwenWork, Qoder, the Qwen app and QwenStudio. The team says the model has deeper knowledge, richer world understanding and stronger coding and agentic skills, and is built for complex coding, front-end builds and professional software workflows. So far that is all there is. Alibaba has not published parameters, benchmark scores, context length, API pricing or a full release date, and AIBase reports a limited-time Token Plan discount of up to 90%. Treat any performance claims you see on X as unconfirmed until Alibaba posts numbers. Alibaba previously said Qwen4 uses a new architecture first tested in Qwen3.8-Flash-Next, and its roadmap lists Flash, Plus and a 27B model meant to run locally. If you already use Qoder, this is the cheapest way to try it on real tickets. As with any hosted model, keep sensitive code out until you have checked the provider's data terms.

Source: https://x.com/Alibaba_Qwen/status/2120335993643684315

Nov 10, 2026

DEALS: Mercor doubles to a $20B valuation on $750M, Graphon raises $35M for agent memory, and OpenAI buys Odessia

The money behind agents kept flowing on Nov 10. Mercor raised a $750 million Series D at a $20 billion valuation, double the $10 billion it reached in October 2025. Benchmark, Felicis, General Catalyst, NVIDIA, Salesforce, Sapphire Ventures and Zoom are among the investors. Mercor started as an AI hiring platform. It now pays a network of more than five million experts to train and evaluate AI models, and it will use the money to build evals, post-training and realistic environments where agents learn professional workflows. "Every job in the economy will need its own AI agents," co-founder Brendan Foody told Reuters. Graphon Labs raised a $35 million Series A led by RRE Ventures, bringing its total to $43.3 million less than six months after a seed round. Graphon's graph-based memory lets agents such as Claude or GPT query linked facts with sources. The company reports 93.9% on the LoCoMo memory benchmark using 58% fewer tokens than loading the whole conversation, which is its own result. And OpenAI acquired Odessia, an AI travel agent that founder Francis Davidson says he built by directing teams of AI agents. Terms were not disclosed, TechCrunch reads it as an acqui-hire, and the consumer service has shut down.

Source: https://www.businesswire.com/news/home/20261110410031/en/Mercor-Raises-%24750-Million-Series-D-at-%2420-Billion-Valuation-to-Help-Enterprises-Own-Their-Intelligence

Nov 10, 2026

AGENTS: Vorflux 2.0 launches AutoBots, always-on Slack coworkers that pick up engineering work on their own

Vorflux co-founder Prasanna Sankar launched Vorflux 2.0 on Nov 10. Its AutoBots are AI coworkers that live in Slack and own ongoing jobs rather than one-off tasks. According to Vorflux, you give each bot a role, a KPI, the tools it needs and a daily budget with chosen models per task. It works 24/7 in its own channel and brings you decisions that need approval. A directory offers ready-made bots for on-call triage, cloud cost, CI health, code cleanup, bug hunting, project tracking and release notes, or you can define your own goal. Sankar, who previously co-founded Rippling, argues that models now deliver predictably and that the bottleneck is starting, tracking and approving parallel work. You can bring your own Codex or Claude Code subscription and pay for machine hours. Vorflux also says it recently closed a $15 million seed round backed by Y Combinator and Peak XV Partners. Its claims about output, such as a bot merging more PRs than a human, are company-reported. Early reaction on daily.dev is curious but cautious: checking the work gets harder once agents start assigning themselves tasks. Start small, with tight budgets and required review.

Source: https://x.com/myprasanna/status/2120205408371634624

Nov 10, 2026

GAME DEV: House House says LLMs touched under 0.1% of Big Walk's code, after 19 IGF jurors refuse to vote for any AI-assisted game

The question of where AI help ends hit indie awards this week. On Nov 9, 19 Independent Games Festival jury members posted a statement saying they would decline to vote for any game that uses generative AI "in any capacity", and that code is "not an exception." On Nov 10, House House updated Big Walk's FAQ. The Untitled Goose Game studio says it used LLM chatbots sparingly in the final stretch to research bugs in third-party code, and that snippets it heavily modified make up "well under 0.1%" of the game's code. It gives an example: Unity URP couldn't cull shadow casters behind the camera, so the game ran 50% slower for about 20 seconds at sunrise and sunset. A Unity engineer reportedly concluded the fix wasn't possible, and an LLM suggested a workaround that became about 10 lines inside URP. The studio says generative AI was not used anywhere else, and that it disclosed the use in its IGF submission but not on Steam, because Steam doesn't ask about code or chatbot use. It also calls its lack of an AI policy "a failure." Whatever your view, the practical lesson for indie game devs is that awards and storefronts now draw different lines. Keep a record of exactly where and how you used AI, and be ready to disclose it.

Source: https://bigwalk.game/faq/

Nov 11, 2026

OPEN WEIGHTS: MiniMax releases H3-Image, a 16-megapixel image model that turns a game map into a first-person view

MiniMax released H3-Image on Nov 11, an open-weight model for text-to-image, reference-based generation and editing. MiniMax says it renders a full 4096x4096 (16-megapixel) image in under 10 seconds, at up to 5x the speed of mainstream alternatives, with no upscaler, and accepts up to 16 reference images. Those are company claims. The Hugging Face model card lists a 33.1B-parameter diffusion transformer distilled to six sampling steps. It is a heavy download, about 69 GB for the model plus a 67 GB text encoder, and it runs on SGLang, vLLM-Omni, diffusers and ComfyUI. The showcase is aimed at game art: environment concepts, characters, creatures and props, an untextured 3D render turned into a film still, and a game map turned into a first-person view with a minimap. The weights come under MiniMax's own community license, not Apache or MIT, and some uses need separate written permission, so read it before you ship assets commercially. If you don't have a large GPU, the MiniMax API and Hailuo AI host it, and MiniMax is offering 50% off 4K requests during launch week.

Source: https://www.minimax.io/blog/minimax-h3-image

Nov 10, 2026

WORLD MODELS: IFM's open PAN 2 shows agents what happens before they act, with weights and training code under Apache 2.0

The Institute of Foundation Models released PAN 2 on Nov 10. IFM was set up by Abu Dhabi's Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), and PAN 2 is an open, general-purpose world model. Give it a starting scene and an action in words, and it generates a video of what happens next. A language model first writes a time-stamped plan of events and camera moves, and a video diffusion model trained from scratch on footage picked for its physics then renders it. IFM says PAN 2 completed more full multi-step plans than any other model it tested, and ranks third on the Physics-IQ Verified leaderboard and first among open-weight models. IFM also says long-term world consistency is still an open problem. Unusually, IFM is releasing the video model weights along with the full training and inference code, all under Apache 2.0. For AI coders building agents, this is a free simulator for checking "what if I do this?" before acting. For game devs, it is an open base for experimenting with physics-aware generated scenes. Expect research-grade tooling for now.

Source: https://ifm.ai/pan2/press-release/

Ship faster this week

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