AI NewsSeptember 3, 2026

The persistent, cloud-based AI agent platform expands from desktop to mobile.
Grok Bot, the AI teammate platform from SpaceXAI, rolled out on Android on September 2, letting users download the app to manage persistent cloud-based agents. These bots run 24/7 on their own computers, handling conversations, notifications, and workflows across mobile and desktop with the same shared workspace.
Early users are already building multi-agent teams, including master agents that delegate tasks and only ping humans when needed. The team is also planning a marketplace for pre-built bots and features like voice support.
Key Capabilities:

• Always-On Agents: Bots run 24/7 on their own compute, independent of whether your phone is active.
• Cross-Platform Sync: Shares the same workspace and context across both mobile and desktop.
• Multi-Agent Teams: Users are building master agents that delegate subtasks to other bots automatically.

 A new hybrid compute system starts research in the cloud, then shifts private work to local AI.
Perplexity launched hybrid compute for its Mac app, starting complex research in the cloud before shifting sensitive file work to local Apple Silicon models for full privacy. A built-in Privacy Gate scans for personal info on-device, offering local processing, masking, or user approval when needed.
Perplexity also open-sourced Lily, their fast local inference engine optimized for models like Qwen3.6 on M-series chips. This lets iPhone prompts route securely to a Mac without using cloud credits for local steps.
Key Capabilities:

• Privacy Gate: Automatically scans for personal information on-device before deciding how to handle it.
• Open-Source Engine: Lily is now open-sourced, built specifically for local AI inference on Apple Silicon.
• Hybrid Routing: Starts research in the cloud, then shifts sensitive tasks to local processing automatically.

 A new technique called MTP lets the model run 1.7x faster locally with no accuracy loss.
Unsloth announced Qwen3.8-Flash can now run 1.3 to 1.7 times faster locally using Multi-Token Prediction, or MTP, which lets the model predict multiple upcoming tokens at once instead of generating one at a time. GGUF quantized versions can reach 170 tokens per second on an RTX PRO 6000.
This speedup comes with no accuracy degradation according to Unsloth's testing. Gains are even larger on devices with lower memory bandwidth, like older Macs.
Key Capabilities:

• Faster Inference: Delivers 1.3x to 1.7x speed improvements through multi-token prediction technique.
• No Accuracy Loss: Maintains the same output quality despite the significant speed increase.
• Consumer Hardware Ready: Hits 170 tokens per second on an RTX PRO 6000 GPU.

Quasar 438B is externally validated as the strongest-performing European model evaluated so far.
Multiverse Computing introduced Quasar 438B, calling it the top European AI model. Independent benchmarking firm Artificial Analysis validated it as the strongest performing European model in their latest comparison.
It's the first large model released by Multiverse Computing, built with 438 billion parameters supporting English and Spanish, targeted at enterprise-scale agents, coding, and complex multi-step workloads.
Key Capabilities:

• European Leader: Externally validated by Artificial Analysis as the top-performing model from Europe.
• Bilingual Support: Built natively for English and Spanish enterprise use cases.
• Enterprise Focus: Designed specifically for enterprise-scale agents and complex multi-step coding workloads.

Muse Spark 1.3 sustains longer work sessions while using 20% fewer tool calls than its predecessor.
Meta released Muse Spark 1.3 with improved performance on agentic and coding tasks, focused on real world usability. It sustains longer horizon work across multiple workflows in a single thread and more actively collaborates with users by asking clarifying questions when stuck.
It's better calibrated on its own limits instead of hallucinating outcomes, using about 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2. Zuckerberg called it the biggest jump Meta has made so far on coding and agentic work.
Key Capabilities:

• Smarter Collaboration: Asks clarifying questions and confirms before taking consequential actions.
• More Efficient: Uses roughly 20% fewer tool calls and 25% fewer tokens than the previous version.
• Better Self-Awareness: Calibrated to recognize its own limits rather than hallucinating confident answers.

 It landed #1 in the Code Arena WebDev with a huge 77-point lead over the next closest model.
Claude Fable 5.1 debuted at #1 in the Code Arena WebDev leaderboard with 1765 points, breaking away from the pack with a 77 point margin. It's a major leap from Fable 5, which sat at #8 overall with 1628 points.
It now sits 77 points above Qwen3.8-Max-0902 in second place and 78 points above Opus 5 in third. At a blended $40 per million tokens, it also reshapes the performance ceiling in the higher priced model range.
Key Capabilities:

• Dominant Lead: Scores 1765 points, a 77 point margin over the next best model on the leaderboard.
• Big Jump: Improves sharply over Fable 5, which previously ranked #8 overall.
• Premium Pricing: Priced at a blended $40 per million tokens, reflecting its top tier performance.

Gemini 3.8 Flash and a Cyber variant launch just three weeks after 3.7 Flash, at the same price.
Google introduced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, its best reasoning and coding model yet, built using long-running agentic loops. It's Google's third Flash release in six weeks, with pricing held at $0.75 per million input tokens and $3.75 output.
The Cyber variant focuses on vulnerability discovery and automated patching, restricted to vetted defenders through Google's new Fairwind Program. Google's Chrome Security team reports it produced 2.6 times more correct patches than the best commercial rivals.
Key Capabilities:

• Rapid Release Cycle: Third Flash-tier model in six weeks, showing Google's accelerating update pace.
• Cyber Defense Focus: The Cyber variant found a critical vulnerability in under 2 hours during internal testing.
• Stable Pricing: Holds the same $0.75/$3.75 per million token rate as its predecessor through year end.