**GPT-5.2 Launches Amid Benchmark Battles as Mistral Teases Next-Gen Model**
New AI Models & Releases
GPT-5.2 Released by OpenAI: OpenAI officially launched GPT-5.2, touting improved performance in software engineering, advanced mathematics, and reasoning tasks. Early benchmarks show mixed results, with the model trailing behind Gemini 3 Pro and Opus 4.5 on SWE-Bench but excelling in other areas.
- Introducing GPT-5.2
- GPT-5.2 Thinking evals
- GPT-5.2 behind Opus 4.5 and Gemini 3 Pro on SWE-Bench
- Independent evaluation of GPT-5.2 on SWE-bench
- WOW GPT-5.2 finally out
- OpenAI's GPT-5.2 + Thinking now on Perplexity
Mistral AI’s Upcoming Model Teased: Mistral AI announced a new model release "in a few days", following rapid advancements in their AI lineup. Details remain undisclosed, but the community anticipates significant improvements.
Devstral Small 2 Now Available: Mistral’s Devstral Small 2 (24B) is now accessible via LM Studio, enabling local deployment with optimized performance. Users report smooth integration with 25GB RAM setups.
ARC 3 Confirmed for Q1 2026: François Chollet announced ARC 3, the next iteration of the Abstraction and Reasoning Corpus, slated for Q1 2026. The update aims to tackle bottlenecks in exploration, goal-setting, and interactive planning as AI progresses toward AGI.
AI Infrastructure & Tools
Mistral Vibe CLI Doubles Context Window: Mistral’s Vibe CLI now supports a 200K token context window (up from 100K), enabling longer and more complex interactions. The update aligns with trends toward expanding context limits in AI models.
Live Model Switching in llama.cpp: The latest llama.cpp update introduces live model switching, allowing users to swap between models (e.g., Mistral Vibe and Granite-4-h-1b) without restarting the server. This feature streamlines workflows for local AI deployment.
RK3588 NPU Hack for Vision Transformers: A developer reverse-engineered the RK3588 NPU to optimize memory limits, enabling large vision transformers to run on edge devices. The project includes a detailed blog post on sharding techniques for NPUs.
Sunpeak: Open-Source ChatGPT App Framework: A developer open-sourced Sunpeak, a framework for building ChatGPT-powered apps. The GitHub repo includes tools for UI development and integration with AI models.
AI Policy & Regulation
US Executive Order Against State-Level AI Regulations: The US Administration issued an executive order to block state-level AI regulations, establishing a task force to challenge such laws and proposing federal oversight. The move aims to standardize AI policy but faces criticism over centralization.
AI Benchmarks & Evaluations
GPT-5.2 Benchmark Performance: Early evaluations of GPT-5.2 on SWE-Bench and other benchmarks reveal mixed results:
- GPT-5.2 High ranks #3 behind Gemini 3 Pro and Opus 4.5.
- GPT-5.2 Medium trails Sonnet 4.5 in cost efficiency.
- Discussions highlight trade-offs between performance and pricing.
- GPT-5.2 Thinking evals
- Independent evaluation of GPT-5.2 on SWE-bench
AI Training Insights
Mistral’s Magistral Model Training Revealed: Umar Jamil of Mistral AI shared details on how the Magistral model was trained, offering rare insights into the team’s methodology and advancements in scaling and efficiency.
AI Services & Utilities
ChatGPT Search Visibility Tool: A free service now lets users test if their content appears in ChatGPT’s web searches, providing feedback to improve discoverability.