📅 AI Industry News – December 12, 2025
Model Updates & Releases
- GPT-5.2 released: performance gains but higher pricing. OpenAI shipped GPT-5.2 with improvements in scientific reasoning (92.4% accuracy), competition math (100%), and long-context handling. Pricing rose to $1.75/M input tokens and $14/M output tokens, with a 90% cache discount. It ranks second on the WebDev Code Arena, though some coding benchmarks underperformed.
- OpenAI blog | System card (PDF) | Docs
- Mistral teases a new model. Mistral AI hinted at an upcoming release on X; the community speculates it may appear on OpenRouter.
- Mistral X post
- Qwen 3 sparse series called "underrated." Users praise models like a3b for coding and reasoning, though some report mediocre results from Qwen 32B.
- OpenRouter Discord
- Disney invests $1B in OpenAI, integrating characters into Sora. The deal includes 3-year licensing with first-year exclusivity; content will appear on Disney+.
- Announcement | CNBC report
- DeepMind opens first automated research lab in the UK. Built with the UK government, focused on AI-driven scientific discovery (materials science, drug research), planned to launch in 2026.
- DeepMind blog
- Unsloth's new packing method: 3x faster training. 3x faster than the old version, 10x faster than FA3; Qwen3-4B trains on just 3.9GB VRAM; fixes old NVIDIA driver dependency conflicts.
- Docs
- llama.cpp adds live model switching. New routing mode loads, unloads, and swaps models without restarts, using a multi-process architecture to isolate crashes, with LRU caching and auto-discovery.
- Hugging Face blog
- Hugging Face WebGPU local voice chat demo. Fully in-browser (STT, VAD, TTS, LLM all local), preserving privacy.
- Hugging Face Space
- GPT-5.2 beats human experts on GDPVal. GPT-5.2 Thinking outperformed 70.9% of human experts across 44 occupations, 11x faster at ~1% of the cost—though human oversight remains advisable.
- OpenAI GDPVal notes | SWE-Bench results
- Reddit debates GPT-5.2's performance and hype. Users praise the 100% competition-math score but criticize $168/M output-token costs; memes mock "AGI" claims over a miscounted R in "garlic."
- Reddit thread | AGI meme
- LMArena users test GPT-5.2 coding. GPT-5.2 High reportedly generated buggy code in Code Arena despite strong SWE-bench scores; ranks second on WebDev; users call the launch "rushed" and overpriced.
- LMArena leaderboard
- CUDA 13 resolves Torch/vLLM compatibility issues — both must use CUDA 13 builds; especially relevant for AMD GPU users.
- GPU MODE Discord
- Hetzner launches 96GB VRAM server at €889. Bare-metal with generous free traffic, aimed at AI startups cutting training/inference costs.
- Nous Research Discord
- Diffusion distillation technique yields free log-probs. Adding a predictive-divergence head and adjusting initial noise improves image likelihood maximization.
- arXiv:2512.02636
- Sandwich norm improves Transformer long-context handling by normalizing activations to handle longer sequences.
- OpenReview paper
- CIRIS Agent tested for jailbreak resistance. Designed as an ethical AI, it refuses harmful content (e.g., meth synthesis instructions), but users continue probing its limits.
- BASI Jailbreaking Discord
- Grok image generation moderation controversy. Some call restrictions strict; others note skilled users can still produce deepfakes and "unaligned garbage" outputs.
- BASI chat
- Cursor's debug mode gets positive feedback, though context rollback cannot restore state; users want backup features.
- Cursor Discord
- Perplexity Pro users hit strict rate limits after just 5 Gemini 3 Pro queries; suspected server load or bugs; workarounds include disabling VPN and clearing cache.
- Perplexity Discord
- Windsurf ships new MCP management UI (v1.12.41, v1.12.160), fixing GitHub/GitLab MCP issues and improving diff zones and Supercomplete.
- Windsurf changelog