AI News Daily Briefing — February 12, 2026
GLM-5 Drops: 744B Open-Weight Model Under MIT License Shakes Up the Coding Leaderboard
The release of GLM-5 "Pony Alpha" — a 744-billion-parameter model under the MIT license — is the most consequential open-source AI event in months. Multiple developers are already calling it the first open-weight model they'd actively recommend for code generation, with commentator Theo ranking it alongside Opus 4.6 and Codex 5.3 as the current best coding models and claiming it outperforms Google's Gemini. That's a bold claim, and benchmark receipts are still thin, but the enthusiasm from practitioners who have actually used it is hard to dismiss.
What makes this genuinely significant for builders is the licensing. MIT means no usage restrictions, no commercial caveats, no fine-print traps — you can deploy it, fine-tune it, and ship products on top of it without a legal department review. At 744B parameters it's not something you'll run on a laptop, but for companies building AI-powered developer tools or code-heavy workflows, having a frontier-class open model with zero strings attached fundamentally changes the build-vs-buy calculus.
The timing is notable too. This lands the same week xAI is hemorrhaging talent (nine engineers including two co-founders gone in a week), Apple is delaying its Gemini-powered Siri yet again, and Anthropic is busy expanding its free tier. The competitive landscape is fracturing: closed labs are fighting retention battles and distribution wars while open-source is quietly shipping models that practitioners actually prefer. If GLM-5's benchmarks hold up under scrutiny, it could accelerate the trend of serious engineering teams running their own inference rather than paying per-token to frontier API providers.
xAI Loses Nine Engineers Including Two Co-Founders in a Single Week
A mass exodus at Elon Musk's AI lab, with two co-founders and seven senior engineers departing amid delays to the Grok 4.2 model originally planned for late 2024. This is now a retention crisis, not just turnover — and it raises serious questions about whether xAI can execute on its roadmap while competitors ship aggressively.
Claude Free Tier Gets File Creation, Connectors, and Compaction
Anthropic is making an aggressive play for free-tier adoption, adding file creation, third-party connectors, skills, and automatic context compaction — features that were previously paywalled. The compaction feature alone, which summarizes earlier context to extend conversations, makes Claude meaningfully more useful for longer workflows without a subscription.
Anthropic Pledges to Cover 100% of Data Center Electricity Cost Increases
A smart PR move with real teeth: Anthropic commits to paying all grid upgrade costs tied to its data centers so ratepayers don't foot the bill. As AI compute demand strains power grids, this preempts the political backlash that could slow data center approvals nationwide.
Anthropic Safety Researcher Resigns Warning of Catastrophic AI Risks
The departure of yet another safety researcher from a leading AI lab adds to a pattern that's becoming hard to ignore. Whether or not you share the alarm, the steady drain of safety-focused talent from the companies actually building frontier models is a structural concern for the industry.
OpenClaw Hits 180K GitHub Stars; Developers Spinning Up Multi-Agent Teams
Lex Fridman's interview with OpenClaw creator Peter Steinberger spotlights the open-source agent framework's explosive growth. Developers are already building multi-agent systems on consumer hardware — one user reports running a six-agent team on a Mac Mini — signaling the framework is hitting the sweet spot between capability and accessibility.
Apple Delays Gemini-Powered Siri Features Beyond iOS 26.4
Despite partnering with Google, Apple is pushing back Gemini integration into Siri yet again. At this point the repeated delays suggest fundamental architectural challenges rather than timeline optimism — and every month of delay is a month competitors embed AI assistants deeper into user habits.
Cursor Triples Usage Limits for Composer 1.5
Cursor raises usage caps 3x for Composer 1.5 and temporarily boosts Auto limits through Feb 16. Smart move to keep power users from churning to competitors during a period where AI coding tools are rapidly commoditizing.
Study Finds GPT-5 Mini Outperforms Human Graders in Data Recovery
Research validation showing GPT-5 mini beating trained human graders at data recovery tasks reinforces the case that smaller, cheaper models are already superhuman at specific narrow tasks — which matters more for production systems than general benchmark scores.
Google AI Studio Gets Redesigned Home with Omnibar Navigation
Google's AI Studio refresh adds an omnibar for quick navigation and better access to chat history and project usage. Incremental but welcome — the original UI was a friction point for developers juggling multiple projects.
GitHub Goes Down, and Its Status Page Doesn't Show Uptime History
GitHub experienced an outage yesterday, and developers noticed the platform's official status page conveniently omits historical uptime metrics. A third-party tracker now fills the gap, which is an awkward look for the world's most critical code hosting platform.
Devin Code Review Hits 40K Daily Uses, Adds One-Click Fixes
Cognition's Devin Review tool crosses a meaningful usage milestone and ships merge buttons and markdown support. The velocity of adoption suggests AI code review is quickly becoming table stakes rather than an experiment.
Windsurf Launches Arena Mode Leaderboard with 40K Votes in Week One
The first in-product coding arena leaderboard introduces a speed-focused evaluation methodology with 40,000 votes in its debut week. This could become the Chatbot Arena equivalent for code generation — a useful signal beyond synthetic benchmarks.
v0 Launches GTM Agent Builder for Sales Teams
Vercel's v0 expands beyond developer tooling into go-to-market workflows, letting sales teams build custom demos and AI agents with real data in minutes. A bet that the same AI-first prototyping that works for engineers will work for revenue teams.
DeepLearning.AI Launches Course on Agent2Agent Protocol
New course covers the A2A protocol standardizing how AI agents built with different frameworks can interoperate. As multi-agent architectures go mainstream, interoperability standards will matter as much as individual agent capability.
Modal Labs Seeks Funding at $2.5B Valuation for AI Inference Infrastructure
General Catalyst is in talks to lead a round valuing the four-year-old AI infrastructure company at $2.5 billion. As open models like GLM-5 make self-hosted inference more attractive, the picks-and-shovels play of making that inference easy to run becomes increasingly valuable.
Box CEO: Companies Will Compete on AI Agent Infrastructure and Token Budgets
Aaron Levie predicts talent recruitment will hinge on a company's agentic capabilities. It's a provocative framing — your AI infrastructure as a recruiting tool — but it tracks with what we're seeing as developers increasingly choose employers based on tooling quality.
a16z GP Argues Venture Firms Must Scale Big or Go Niche
Alex Rampell makes the case that mid-size venture is a dead zone — firms need to either manage massive funds or hyper-specialize. The meta-game of venture is consolidating just like the industries it funds.
Microsoft Warns of Active Zero-Day Exploits Targeting Windows and Office
Critical vulnerabilities allow complete system takeover via malicious links, and hackers are already exploiting them in the wild. If your team is on Windows, patch today — this isn't a theoretical risk.
iOS 26.3 Patches 35+ Security Vulnerabilities
A major security update for iPhones, plus Apple also shipped patches for legacy iOS 18.7.5, macOS Sequoia, and Sonoma. Heavy patch day across the Apple ecosystem — update everything.
Microsoft Patches Notepad Markdown Flaw Enabling Remote Code Execution
Attackers could remotely execute malicious files via Markdown links in Notepad — a reminder that even the simplest tools can become attack vectors when new features (like Markdown rendering) expand the attack surface.
Google Cancels Android 17 Beta 1 After Briefly Releasing It
In a confusing sequence, Google appears to have both released and then pulled Android 17 Beta 1 for Pixel devices on the same day. Not a great look for developer confidence in the Android beta program's stability.
iOS 26.3 Adds Transfer-to-Android Tool and Location Controls
Apple making it easier to switch to Android is a surprising move driven by EU regulatory pressure. The addition of cellular location privacy settings is a quieter but arguably more important change for privacy-conscious users.
Palantir Launches $10K Grants for High School Projects
The Valley Forge Grants program offers $10,000 to high schoolers building on Palantir's platform. It's a pipeline play for future talent, but putting enterprise defense software in the hands of teenagers is certainly a brand strategy.
Today's dominant pattern: the center of gravity in AI is shifting toward open models and self-hosted infrastructure. GLM-5 shipping under MIT at 744B parameters, Modal Labs raising at $2.5B to make inference easier, and OpenClaw hitting 180K stars all point in the same direction. Meanwhile, the closed labs are fighting fires — xAI is losing co-founders, Anthropic is losing safety researchers, and Apple can't ship Siri. Builders who bet on open, composable AI infrastructure are increasingly the ones setting the pace.