A weekly AI roundup covers a flurry of releases: xAI's Grok 4.6 (matching top models at half the price, with model card confirming Cursor Bench data was removed from training) and the new Grok Bot, an always-on multi-agent swarm product from SpaceXAI/Cursor with its own compute per bot, iOS/macOS apps, and granular permission controls. DeepSeek V4 Pro 0813 launched in general availability under MIT license (1.6T params, 49B active, 1M context) alongside its open-source harness, which hit 23K GitHub stars in a day. Meta re-entered open source with Muse Glimmer 30B (Apache 2.0, runs on a single 24GB GPU) and promised the larger Muse Spark 1.2. Mid-show breaking news included Google's Gemini 3.7 Flash (fast, multimodal, near cost/performance Pareto frontier) and OpenAI's ultrafast GPT-5.6 Sol preview on Cerebras. Additional notes cover Qwen3.8-Max, Nvidia Nemotron 3.5 Lightning, Anthropic's new global Claude output watermarking under the EU AI Act, and a paper on API keys leaked via hidden reasoning traces.
Table of contents
Grok Bot and Grok 4.6 from SpaceXAI/CursorOpen Source AI and other newsThis weeks buzzGemini 3.7 Flash - breaking in the middle of the showThat’s a wrapThursdAI - Aug 13, 2026 - TL;DRQuestions this post answers
What is Grok Bot and how does it differ from other autonomous AI agent tools like OpenClaw or Claude Cowork?
Grok Bot, released by SpaceXAI/Cursor, is an always-on swarm of independent agents rather than one agent with multiple personalities. Each bot gets its own isolated computer, agent-to-agent communication is transparent but read-only to the user, there's no model picker (Grok 4.6 decides), and it works across macOS and iOS with notifications and task handoff between devices. Track fast-moving agent tooling releases like this one on daily.dev before choosing your next always-on assistant.
How does Grok 4.6 compare to other frontier models in price and performance?
Grok 4.6 lands at #4 on intelligence and #5 on speed on the Artificial Analysis leaderboard while costing half as much as the models ranked above it, at $2/$6 per million tokens. It scored 61.3 on Frontier Code, beating GPT-5.6 Sol and trailing only Opus 5, and its model card confirms Cursor Bench data was removed from its training weights, where it now ranks #1. Compare frontier model cost and performance trade-offs on daily.dev before picking a coding model.
Is DeepSeek V4 Pro open source and what license does it use?
DeepSeek V4 Pro 0813 is released in general availability under the MIT license, with 1.6 trillion total parameters and 49 billion active parameters, a 1 million token context window, and pricing of $0.435/$0.87 per million tokens. It improved sharply on DeepSWE, jumping from 12.8 points in the preview to 62.7 in this release, and its companion harness gained 23,000 GitHub stars within a day. Follow open-weight model licensing changes on daily.dev when picking a model for production use.