AI News: The AI Agent Race Just Exploded

HWHunyuan3D WorldClawTencentUnverified

Tencent Hunyuan research that generates an entire 3D world from a prompt, with every element an editable separate asset. It uses GPT Image 2, SAM 3 segmentation and Hunyuan 2D-to-3D conversion; not yet available, GitHub is empty.

GRGrokbotxAIUnverified

xAI's desktop and mobile agent app where you create named agents, each with its own cloud computer. It has plugins, a teach-a-task recording feature and scheduled routines. Matt demoed research, email triage and chief of staff bots.

S2Seedance 2.5ByteDanceUnverified

Video model now available on Artlist. It generates up to 30-second videos in one go and accepts up to 50 reference images for consistency. This was the sponsored segment.

CAClaude AI Content WatermarksAnthropicLimited in EU

Anthropic is embedding imperceptible machine-readable watermarks in text generated by Claude across its products. Detection is not conclusive, and Matt expects workarounds and stripping tools.

SUSunoSunoAvailable in EU

Suno will add audio watermarking and fingerprinting and a download cap from September 3rd. The cap is 20 songs per month on Pro and 60 on the more expensive plan. Matt has mixed feelings but says Suno is still the best music generator.

SASpotify AI Persona BadgeSpotifyUnverified

Spotify will add an AI persona badge to artist profiles from mid-September and let artists self-disclose. AI personas will be excluded from editorial and algorithmic recommendations. Matt supports the idea.

G4Grok 4.6xAIAvailable in EU

Near the top on DeepSWE and leads GDPval and Harvey Lab. It costs $2 per million input and $6 per million output tokens. On Matt's BusyBench it ranked 43rd, below Grok 4.5. Grok 4.7 is expected in 3-4 weeks.

G3Gemini 3.7 FlashGoogleAvailable in EU

Cheap, fast Flash model priced at 75 cents per million input and $3.75 per million output tokens. It scored 65.3% on DeepSWE and led Matt's BusyBench leaderboard.

MGMuse GlimmerMetaAvailable in EU

Meta's small open-weight agentic model for local use. It beats similar-size models such as Gemma 4 31B and Qwen 3.6 27B on many benchmarks. Full precision needs about 55 GB, and a quantized version is available. Its BusyBench output was poor.

N3Nemotron 3.5 LightningNVIDIAAvailable in EU

NVIDIA's 30B-parameter open-weight model for long-running agents, in the same class as Glimmer. It did poorly on BusyBench SVG generation, but cheaply and quickly.

DEDeepSeek-V4-ProDeepSeekUnverified

Generally available release following the preview. It scored 62.7 on the benchmark shown, ahead of GLM 5.2 and Opus 4.8 but behind Fable, Kimi K3 and GPT 5.6 Sol. On BusyBench it cost about 8 cents and took 3.5 minutes.

Microsoft's new coding model. Few benchmarks or details were released, and it isn't on OpenRouter, so Matt couldn't test it.

GSGPT-5.6 SolOpenAILimited in EU

Ultra-fast mode of GPT-5.6 Sol running on Cerebras hardware at about 750 output tokens per second. It is limited to trusted testers. Matt saw only a side-by-side demo, and he calls GPT-5.6 Sol his favorite coding model.

CDChatGPT Desktop App for LinuxOpenAIAvailable in EU

The ChatGPT desktop app, which includes ChatGPT, ChatGPT Work and Codex, is now available on Linux.

CCClaude Code Auto ModeAnthropicAvailable in EU

From August 14th, new Claude Code sessions on Pro, Max and Team plans default to auto mode. Most steps are auto-approved unless Claude judges them potentially harmful.

CCClaude Chrome Extension with CoworkAnthropicAvailable in EU

The Claude Chrome extension now uses Claude Cowork and can take more control. In the example it screenshots all open tabs and then generates a report in the Claude app.

CCClaude Code Session MessagingAnthropicAvailable in EU

Claude Code sessions can now message each other. One session sends a summary and the other picks it up mid-task, so you don't have to re-explain context.

L2LTX 2.5LightricksAvailable in EU

Open-weight video model that makes a 10-second clip from an image in about 6.8 seconds on NVIDIA superchips. It runs locally on RTX GPUs.

WAWan3.0AlibabaUnverified

Alibaba's video model that generates decent-looking clips up to 30 seconds long.

MAMAI-Image-2.6MicrosoftUnverified

Microsoft's image model, ranked number two on the text-to-image Arena leaderboard behind GPT Image 2. The ranking comes from blind voting.

A DeepMind model that reads sign language from video and transcribes it as captions. Matt saw it in demo videos on the website.