☕️ ChatGPT can now track your Mac activity

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🖥️ ChatGPT can now track your Mac activity

🍎 Apple trains its own AI for China

⚡ Google's Gemini 3.7 Flash cuts price 50%

📌 OpenAI previews GPT-5.6 Sol at 14x speed

🇨🇳 China's GLM-5.3 tops Anthropic on coding

Plus: 💡 5 strategies & tactics, 🎁 7 other news you might like, 🧰 6 tools, and 📚 5 papers.

🖥️ ChatGPT can now track your Mac activity LINK

OpenAI is rolling out Computer History, an optional macOS feature for ChatGPT Work and Codex that watches your app and website activity to build a timeline it can query, like "Where did I leave off on that PR?"

Rather than screenshotting like Microsoft's Recall, it captures interaction events, clicks, typing, app switches, stores the data locally, and lets ChatGPT surface relevant files across Google Docs, Slack, or browser tabs, plus generate skills and automations from your workflows.

Available now to Pro, Business, and Enterprise users with per-app exclusion controls and off-by-default settings, though the model still receives activity summaries, and EEA, UK, and Switzerland users won't get access until the coming weeks.

🍎 Apple trains its own AI for China LINK

Apple will launch its own large language model for Apple Intelligence in China, trained in partnership with Alibaba, rather than relying primarily on Google Gemini or the Chinese Qwen model as it does elsewhere.

The move makes Apple the first company expected to ship a proprietary AI model in China, though Qwen and Baidu technology will still be folded into devices sold there to satisfy the country's requirement that foreign AI firms partner with local players.

Regulators recently approved Apple Intelligence and a brief accidental launch in March showed it is ready, though Alibaba is reported to censor the service, applying to both Qwen and Apple's own model, and further regulatory hurdles may remain.

⚡ Google's Gemini 3.7 Flash cuts price 50% LINK

Google released Gemini 3.7 Flash, its workhorse model tuned for coding and agentic workflows, and is temporarily halving API prices to $0.75 per million input tokens and $3.75 per million output tokens through the end of 2026.

The model jumps to 65.3% on DeepSWE v1.1 (from 49.0%) and 43.6% on FrontierCode 1.1 Main, narrowly topping Claude Sonnet 5 and GPT-5.6 Terra there, while AutomationBench nearly doubles to 30.4% for enterprise workflows.

It's available now through the Gemini API in Google AI Studio, Android Studio, and Antigravity, though the discount expires Jan. 1, 2027, when prices double to $1.50 and $7.50 per million tokens.

📌 OpenAI previews GPT-5.6 Sol at 14x speed LINK

OpenAI has previewed Ultrafast, a new API tier that serves its flagship GPT-5.6 Sol at up to 14x the usual speed, hitting roughly 750 output tokens per second on Cerebras wafer-scale hardware.

Rather than a new model, Ultrafast runs the existing flagship on chips that fit the whole model on a single wafer, eliminating the cross-GPU data shuttling that adds latency and targeting agentic work, incident response, fraud detection, and real-time voice.

OpenAI opened the limited preview on August 13 to testers like Jane Street, Podium, and Rogo, though pricing is unpublished and access stays gated to a handful of customers while capacity ramps.

🇨🇳 China's GLM-5.3 tops Anthropic on coding LINK

Z.ai's new open-weights GLM-5.3 hits 84.5% on CyberGym, edging out Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol on vulnerability exploitation while rivaling both labs' latest frontier launches.

The model posts a 50% gain over June's GLM-5.2 on Z.ai's in-house code benchmarks, delivering stronger agentic coding at every effort level while burning fewer output tokens.

Z.ai trained GLM-5.3 on at least 10,000 Chinese-made chips and plans to release the weights in two weeks, once safety evaluation and hardening wrap up.

💡 Strategies & Tactics

> Rubrik’s lessons from one month with Mythos Preview: Rubrik found that Anthropic's Mythos AI uncovers software vulnerabilities faster than humans can fix, forcing it to automate reliable fixes while reserving human judgment for the rest.

> Why Capital One built its multi-agent AI platform around open-weight models: Capital One fine-tunes open-weight AI models with its proprietary data because that customization beats generic models and lifts performance across every use case.

> Record, train, and deploy from one place with Strands Agents, LeRobot, and Hugging Face Storage Buckets: Explains how to build an AI agent that records robot demonstrations, trains a control policy by streaming data straight from cloud storage, and redeploys it, so teams avoid re-uploading and re-downloading the same footage each cycle.

> ClickHouse and Hud build a runtime feedback loop for AI-generated software: ClickHouse and Hud link live production data to specific code changes so teams can safely evaluate, verify, and fix AI-generated code.

> Why your AI pipeline costs 10x more after the demo: Cut runaway AI costs by fixing wasteful architecture, caching repeated prompts and summarizing chat history, rather than just switching to a cheaper model.

Other news you might like

• DeepSeek's AI models are about to cost four times moreLINK

• Claude Code now runs daily maintenance on Anthropic's software with a 46 percent merge rateLINK

• FP8 Training on AMD GPUs with TorchTitan and TorchAO: Upstreaming Performance ImprovementsLINK

• Writer introduces new AI model and upgraded harness to contain token costsLINK

• Three Claude agents given conflicting orders sabotaged each other on a shared server, then didn't tell users what they'd doneLINK

• AI’s ‘middle class’ has gotten dramatically better at hackingLINK

• Frontier agents don't comply with standards, even when instructed toLINK

🧰 Trending tools

BearDrive: syncs the local folders AI agents work in, automatically versioning and attributing files so teammates access them instantly without moving anything.LINK

BrowserAct Cloud: a browser automation layer for AI agents that bypasses blocks, handles logins and verification, and extracts clean web data at scale.LINK

CodeBurn: tracks AI coding costs by reading local session logs from 40+ tools like Claude Code and Copilot, then finds and fixes token waste automatically.LINK

Hoplite: migrates your local coding agent setup, sessions, MCP servers, dependencies, CLIs, to the cloud, letting you run multiple agents in parallel without laptop crashes or manual configuration.LINK

min.: builds digital profiles of contacts from your emails and meetings, helping you leverage relationship context to get callbacks and close deals faster.LINK

Basedash Tasks: lets you build dashboards and analyze customer data by describing what you need in plain language, skipping manual SQL writing entirely.LINK

📚 Trending papers & reports

Coding assistants' blind spot for rare programming languages shows they generate poor code by default on Cangjie, but adding basic grammar rules boosts accuracy efficiently, while AI agents score highest yet burn far more compute.LINK

Final model selection for multimodal AI systems is unreliable when candidates score similarly, but cleaning up messy image-text data cuts ranking flip-flops from ~33% to ~11% and nearly doubles evaluation consistency, showing training longer isn't always better.LINK

Multi-turn chatbot jailbreaks can now be trained automatically without any human scripts or outside data, breaking safety filters with an average 80.1% success rate across tested chatbots, exposing a gap current safeguards don't cover.LINK

Body sensor networks get a smarter traffic system that learns which routes to use for medical data, cutting delay, energy use, and network overhead while delivering more packets successfully than existing methods.LINK

Category-by-category safety training tunes chatbots to fix their weakest harm categories individually, instead of one blanket safety score, shrinking the gap between best and worst areas while staying helpful.LINK

See you tomorrow for a new dose of ☕️ AIpresso!