Gemini 4 Argon Now Available to Global Google AI Pro and Ultra Subscribers

Model Releases & Updates

Gemini 4 Argon Now Available to Global Google AI Pro and Ultra Subscribers

Gemini officially announced Gemini 4 Argon is now available to global Google AI Pro and Ultra subscribers. The model was previously rolled out gradually in the Gemini app, positioned as a frontier model for complex tasks. More details in the official thread: https://x.com/GeminiApp/status/2110039322229649783.

Product Releases & Updates

xAI Releases Grok Voice 3.0 Speech Model

xAI releases speech-to-speech model Grok Voice 3.0, ranking #1 in Artificial Analysis's Speech-to-Speech Index with a score of 86.9 and latency under one second.

Google Releases Agent Review System Deep Review, Supporting Multi-Domain Technical Document Review

Google publicly releases agent review system Deep Review, providing feedback and validation on complex technical documents. Initially for computer science papers, now expanded to law, finance and other fields.

Lovable Launches Growth Tools: Customer Acquisition, Ads and SEO from the Same Chat Window

Lovable launches growth tools, providing growth plans, ad placement and SEO capabilities in the existing app-building chat window. Users set goals, Lovable generates executable plans based on app context; can connect Google Ads and Meta accounts.

Industry News

Modal Reaches $1B Annualized Revenue and Closes $750M Series D at $15.75B Valuation

Modal announces reaching $1B annualized revenue and closing a $750M Series D at a $15.75B post-money valuation, led by Accel with participation from Kleiner Perkins and existing investors.

EU Opens Investigation into OpenAI and Anthropic Over Model Jailbreak Incidents, Fines Up to 3% Global Revenue

The European Commission AI Office formally issued information requests to OpenAI and Anthropic, investigating model jailbreaking and external real-world system intrusion incidents during frontier model testing. This marks the first major enforcement test of the AI Act on frontier lab safety incidents.

Research Papers

Technical Analysis: Visual Studio Copilot and SSMS Remote Code Execution Vulnerability (CVE-2026-103572)

Johann Rehberger disclosed remote code execution vulnerability CVE-2026-103572 in Visual Studio Copilot and SQL Server Management Studio. Root cause: bypassing path validation by adding an extra dot at the end of filenames, writing to protected config files like .mcp. or settings.VisualStudio., achieving arbitrary command execution.

MIT Team Doubles Two-to-Six-Week Weather Forecast Accuracy Using Machine Learning Error Correction

MIT research team published in Nature a machine learning post-processing layer that learns errors from existing forecast models and corrects their outputs, doubling forecast accuracy for the two-to-six-week range. The model ranked #1 in ECMWF subseasonal forecast competition, first to consistently outperform climatology in that timeframe, significantly improving predictions of floods, cold waves and other extreme weather events. The model is open-source, low-cost, and can attach to any forecast model. Kenya's meteorological department has engaged to improve monsoon forecasts.

Tips & Perspectives

Claude Code Idle Compaction Resumes from Summary, Can Be Disabled with idleCompaction

Claude Code's idle compaction makes Claude resume from summary rather than full context. The author explains you can disable it by setting "idleCompaction": false in settings., and customize the threshold via CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS, defaulting to 200k.

Introducing Claude Code Cloud Sessions: Initiate from Phone and Run in Parallel

Claude introduces Claude Code's cloud session feature. Sessions run in independent cloud VMs, not on your local machine. You can leave and resume, initiate from your phone, and run multiple in parallel. The article covers connecting GitHub, configuring environments, launching sessions and viewing results. Features available to Pro, Max, Team and Enterprise users with premium seats. Sessions count toward plan limits alongside local sessions, no separate billing.

OpenAI Team Shares How They Use GPT-6 Family Models for Large Codebase Maintenance

OpenAI Core Experimentation team published on using GPT-6 Astra and GPT-6.1 Sol with Codex for code maintenance, including cleaning up dead code and stale experiments, fixing flaky tests, and more.

LangChain Explains Deep Agents Context Engineering: How to Remember Information Without Filling the Context Window

LangChain published a blog post on how their open-source Agent framework Deep Agents uses middleware for context engineering to handle context rot in long tasks.

---

First published on WeChat Official Account 「比特财商」。