Claude now runs on your Mac in background

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🤖 Claude now runs on your Mac in background

🇨🇳 Alibaba's Qwen rivals top US AI

⚔️ Meta says its new AI rivals OpenAI

🔐 Gemini 3.8 Flash hunts security bugs

🔒 Anthropic lets firms keep AI safety data

Plus: 💡 5 strategies & tactics, 🎁 6 other news you might like, 🧰 6 tools, and 📚 5 papers.

🤖 Claude now runs on your Mac in background LINK

Anthropic today rolled out background computer use for Claude Code, letting the agent operate on your Mac without commandeering the cursor, so you can keep working while it runs tasks in parallel.

The feature turns computer use into a multitasking unlock rather than an interruption, following OpenAI's Codex-turned-ChatGPT, which shipped background computer use on the Mac earlier this year.

Access is limited to Pro and Max subscribers, and the background capability runs only on macOS.

🇨🇳 Alibaba's Qwen rivals top US AI LINK

Alibaba's Qwen-3.8-Max-0902 debuted at the top of the Code Arena WebDev leaderboard, matching Anthropic's Claude Fable 5 roughly two months after that model launched.

The 2.4-trillion-parameter model scores 1,629 on Code Arena WebDev at $2 per 1M input and $6 per 1M output tokens, undercutting Claude Opus 5 (Max), which scores 1,688 but costs $20 per 1M.

It also leads the Data & Analytics and Consumer Product categories, though critics have raised distillation concerns given Fable 5 shipped with guardrails against exactly that.

⚔️ Meta says its new AI rivals OpenAI LINK

Meta released Muse Spark 1.3, its most powerful LLM yet, claiming it matches OpenAI and Anthropic's latest models and leads on coding and agentic automation, available now via the paid Meta Model API.

Independent testing by Artificial Analysis scored Muse Spark 1.3 at 62 on its Intelligence Index, trailing only Anthropic's Fable 5.1 and Opus 5 while beating OpenAI's GPT-5.6 Sol, and it uses ~25% fewer tokens than version 1.2 at the same price.

The model handles multiple concurrent workflows, retains context across tasks, and asks for confirmation before irreversible actions, though Meta has dropped its open-source approach and hasn't decided whether to release Muse Spark 1.3's weights.

🔐 Gemini 3.8 Flash hunts security bugs LINK

Google shipped Gemini 3.8 Flash in two variants, including Flash Cyber, a version tuned for autonomously finding and patching security vulnerabilities that already exceeds much larger frontier models on defensive coding tasks.

Flash Cyber hit 86.2% on CyberGym and 47.2% on CWE-Bench, produced 2.6x more correct Chrome patches than bigger models, and Wiz measured up to 9.7% higher vulnerability recall at 2.3-5.2x lower cost.

The standard 3.8 Flash carries a 1M-token input window and $0.75/$3.75 per million input/output tokens, but Flash Cyber ships only to vetted "trusted defenders" via the Fairwind Program due to its more permissive misuse mitigations.

🔒 Anthropic lets firms keep AI safety data LINK

Anthropic launched Enterprise Frontier Safeguards, letting eligible customers store AI misuse-monitoring data in their own cloud environments while combining zero data retention with automated detection of abuse patterns, rolling out in phases later this fall.

EFS stores activity in customer-controlled infrastructure like Amazon S3, Azure Blob Storage, or Google Cloud Storage using their own encryption keys, with automated systems flagging signals such as offensive cyber attempts or stolen credentials, and no Anthropic human review.

It works across Claude Code, Claude Enterprise, Bedrock, Google's Agent Platform, and Microsoft Foundry, and Anthropic won't charge for it, though enterprises absorb standard cloud costs and take on the operational burden of alert triage and incident response themselves.

💡 Strategies & Tactics

> How we make AI coding more cost efficient without sacrificing task quality: GitHub Copilot cut AI coding costs by measuring efficiency across whole tasks rather than per response, trimming unused formatting and needless retrieval turns without hurting quality.

> Co-Designing AI Models Using Speculative Decoding for Faster LLM Inference: Speed up language model inference by having a small draft model guess several tokens at once, then tuning guess length to your hardware.

> The Modern CUDA Toolbox in Practice: A Step-by-Step Optimization Walkthrough: Explains how to make GPU code safer and 300 times faster by using NVIDIA's debugging tools, optimized libraries, and parallel memory streams instead of hand-written kernels.

> Training a coding model to paint watercolours with TRL and OpenEnv: Explains how to train a language model to paint watercolors by rewarding it against a hand-rated image pool that encodes personal taste.

> How concerned should we be about Astra's recurrent architecture?: OpenAI's Astra keeps its hidden reasoning depth within twice that of GPT-4, so it barely reduces the ability to monitor the model's thinking for now.

Other news you might like

• Nvidia confirms it will buy Hugging Face for $12.9 billionLINK

• Coder launches Agent Relay with SpaceXAI, running Cursor’s cloud agents on customer infrastructureLINK

• Multiverse says its 438B model is fast enough for AI agents. The benchmarks tell a more complicated story.LINK

• PyTorch 2.14 Release BlogLINK

• OpenAI to Congress: we're building AI systems that can shut themselves downLINK

• HiddenLayer nabs $100M as enterprises rush to secure their AI deploymentsLINK

🧰 Trending tools

Agent Builder by Airtop: builds AI web agents from plain-English descriptions to log in, browse sites, extract data, and complete tasks without code.LINK

Readr: an ebook reader that answers questions about your book in the margin, gives spoiler-free recaps, and reads pages aloud on-device.LINK

Blume.codes: monitors your local coding agent sessions and converts repeated corrections into reusable rules and skills, working with Claude Code, Codex, and Cursor.LINK

Nex: a multimodal creativity platform with its own image synthesis model for creating, organizing, and sharing visual stories.LINK

Staats: cookieless, agent-native site analytics where your AI agent tracks deploys, flags spikes, and surfaces insights through chat instead of dashboards.LINK

ADR: secures enterprise AI agents through observability, security benchmarking, and threat detection, currently deployed in production at Uber.LINK

📚 Trending papers & reports

Vector search scaling shows that the cost of finding similar items in the databases powering AI search grows faster than vendors assume until datasets get very large, letting teams size infrastructure realistically.LINK

Screen-clicking software agents can now learn specialized professional workflows by watching an expert perform them once, turning recorded demonstrations into reusable skills, tested on a solar-power client's simulation tasks.LINK

Robot instruction-following gets far more precise when training pairs demonstrations with detailed how-to language like which arm to use, lifting real-world dual-arm task success to ~63 out of 100 versus ~50 for goal-only training.LINK

Image generator speedups pick which steps in an image-making model actually matter and skip the rest, delivering sharper results at the same compute budget with no retraining, cutting one quality-error score from 1.48 to 1.26.LINK

Auditory illusion tests reveal that today's audio AI systems mostly hear sounds literally rather than falling for the same tricks humans do, and none fully match how people actually perceive music, speech, and noise.LINK

See you tomorrow for a new dose of ☕️ AIpresso!