The model race moved toward faster, cheaper agent execution: Google introduced Gemini 3.8 Flash and a cyber-defense variant, while Meta’s Muse Spark 1.3 posted strong coding-agent results. The releases put price, latency, and deployability alongside raw capability as the competitive battleground.
Agents are also moving beyond foreground chat. Claude can now operate desktop applications in the background, and Perplexity open-sourced Lily, an Apple-silicon inference engine for hybrid local/cloud tasks. Security is becoming part of the product itself, with Google positioning autonomous vulnerability detection and patch generation for critical-infrastructure defenders.
Anthropic says Claude can click, type, and open desktop apps while the user works elsewhere. This turns computer use from a supervised interaction into a background workflow, increasing the practical reach of desktop agents while raising the importance of permissions, observability, and safe interruption.
Google DeepMind introduced Gemini 3.8 Flash with gains across software engineering, agentic tasks, and multi-step reasoning, plus Gemini 3.8 Flash Cyber for vulnerability detection and patching. The paired release shows general-purpose agent capability and specialized security automation advancing together.
Lily is the local inference engine behind Perplexity Computer, specialized for Qwen3.6-35B-A3B on Apple silicon. Perplexity says it prevents local execution from bottlenecking hybrid tasks, making on-device compute a more viable component of cloud-connected agent workflows.
Artificial Analysis reports that Muse Spark 1.3 (max), currently in limited partner preview, scored 68 on its Coding Agent Index in the Muse Code harness, second only to Claude Opus 5 (xhigh). The available xhigh variant scored 64, adding another competitive option to the coding-model market.
Alibaba Qwen says its updated model rose from 1669 to 1691 on CodeArena’s WebDev leaderboard, reaching first place for agentic web-development workflows. The result highlights continued pressure from open-model providers in multistep reasoning, tool use, and full-application generation.
Selected posts from the edition, with links to the originals.
“Claude can now use your computer in the background in Claude Cowork and Claude Code.”
View post on X ↗“Two new Gemini models are here to help scale your AI agents and secure code: 3.8 Flash and 3.8 Flash Cyber.”
View post on X ↗“Today we’re open-sourcing Lily, the local inference engine we built for hybrid compute in Perplexity Computer.”
View post on X ↗“Meta's Muse Spark 1.3 (max), which is in limited preview for Meta's partners, scores 68 on the Artificial Analysis Coding Agent Index.”
View post on X ↗“#1 on CodeArena: WebDev leaderboard. Qwen3.8-Max-0902 jumps from 1669 to 1691.”
View post on X ↗“Qwen3.8-Flash can now run 1.7× faster locally with MTP.”
View post on X ↗“Gemini 3.8 Flash (High) has reshaped the Pareto frontier for Agent Arena.”
View post on X ↗“Exciting new research on coding agents that keep building for days without a human.”
View post on X ↗“One /goal and it's been working for ~15hrs now and has simplified, unified, optimized, removed, or cleaned up the repo enough to shed 375,000 lines of code.”
View post on X ↗“How do you make sure your vendor selling an AI solution is incentivized to serve the good (and therefore expensive) models?”
View post on X ↗