AI Dev News Digest: December 26th, 2025
2h·Apple, Disney, Zillow, Affirm,…
AI News: Holiday week recap – what mattered
People think the holidays are all vacation, but the tech world kept moving.
- NVIDIA announced a $20 billion deal with Groq on Christmas Eve to license inference chips that claim 10× better power efficiency than GPUs. (SiliconANGLE)
- GitHub shipped big updates: Copilot now has memory for your codebase, supports Agent Skills (custom workflows), and added Claude Opus 4.5 and GPT-5.2 to all paid tiers.
- Cursor started building its own AI model so it can run independently of OpenAI or Anthropic. (HackerNoon)
Key infrastructure moves
- NVIDIA–Groq: licensing deal looks designed to dodge antitrust scrutiny. Groq's LPU chips promise huge power savings for inference. (SiliconANGLE)
- Cursor builds in-house models: part of a trend where AI IDEs want full control over the stack. (HackerNoon)
GitHub & Copilot updates
| Feature | What it does |
|---|---|
| Copilot Memory (GitHub) | Stores repository-specific context so suggestions improve over time. |
| Agent Skills (GitHub) | Lets you package scripts and resources for repeatable tasks. |
| Claude Opus 4.5 (GitHub) | Anthropic's most capable model now available across all paid tiers. |
| GPT-5.2 (GitHub) | OpenAI's latest also landed in Copilot. |
| GitHub Actions lazy loading (GitHub) | Workflows with 300+ jobs load faster. |
| Dependabot dismissal requires review (GitHub) | Adds a safety step before closing alerts (Code Security only). |
New models and platforms
- Google Gemini 3 Flash – default model for the Gemini app, gives Pro-level reasoning at lower cost. (Google)
- T5Gemma 2 – open encoder-decoder models with 128K context windows, good for long-context tasks. (Google)
- Zhipu AI GLM-4.7 – an open-weight coding model that matches Claude Sonnet 4.5 on benchmarks; free to run locally. (Business Wire)
- Anthropic Agent Skills standard – portable across tools, centralised admin in enterprises. (AI Business)
Security & research
OpenAI said prompt injection may never be fully solved. They used reinforcement learning to find an exploit where a malicious email tricked the browser agent into sending a resignation letter. A new adversarially-trained model was released afterward. (OpenAI)
Tooling and integrations
- Vercel AI SDK 6 – adds first-class agent support, human-in-the-loop approvals, debugging tools, and full MCP with OAuth. (Vercel)
- Cursor buys Graphite – a code-review startup that lets teams work on multiple dependent PRs simultaneously. (TechCrunch)
- OpenAI ChatGPT App Directory – developers can submit third-party apps for discovery inside ChatGPT. (OpenAI)
- Replit in ChatGPT – describe an app, and Replit builds it in 2–4 minutes right in the chat. (Replit)
Startups & funding
- Dazzle AI (Marissa Mayer) – raised $8M to build a consumer AI assistant. (SiliconANGLE)
- Tuya "Hey Tuya" – a multi-agent home-assistant with 155M daily interactions. (PRNewswire)
Holiday perks
OpenAI and Anthropic doubled usage limits through New Year's for most users. OpenAI added a "Santa Codex" persona; Anthropic doubled limits for Claude Pro/Max subscribers (Dec 25–31). (Techloy)
Reading suggestions
- MIT Technology Review: the real cost of AI coding tools on junior developers. (MIT Tech Review)
- Google Blog: 2025 year-in-review, covering Gemini 3 and the shift to agents. (Google Blog)
- A blog post on "From vibe coding to VIABLE coding" – how senior engineers get more value from AI. (EveryDev.ai)
That's the rundown for this holiday week. Happy catching up!
Comments
Sign in to join the discussion.
No comments yet
Be the first to share your thoughts!