Alibaba's Qwen team just dropped Qwen3-Max-Thinking, a new proprietary reasoning model built around advanced test-time scaling and autonomous tool use. Paired with web search, it notched a 49.8 on Humanity's Last Exam, beating both Gemini 3 Pro and GPT-5.2.
OpenAI’s GPT-5.2 was just caught citing Grokipedia—xAI’s controversial AI-generated encyclopedia—on sensitive topics like Iran and the Holocaust, according to tests by The Guardian. The catch is that Grokipedia previously faced backlash for citing neo-Nazi forums and questionable web sources. OpenAI responded that its model searches broad public web sources with safety filters in place.
Anthropic just open-sourced Claude’s Constitution under a CC0 license, an extensive manifesto written directly to the AI model rather than human readers. Instead of rigid dos and don'ts, it establishes a strict hierarchy prioritizing safety and ethics over helpfulness, openly considers whether Claude might possess consciousness, and explicitly directs the model to disobey Anthropic if ordered to do something shady.
Character.ai just made public the internal engineering techniques that powered its early pretraining cluster, originally spearheaded by co-founder Noam Shazeer. The headliner is Squinch, a 6-bit gradient compression algorithm that matched bfloat16 fidelity despite operating on a fraction of modern network bandwidth. Why it matters: These bandwidth-saving tricks offer valuable architectural blueprints for anyone training sparse MoE models or fine-tuning across bandwidth-constrained clusters.
OpenAI just rolled out a major security update to ChatGPT Atlas's browser agent, powered by an RL-trained automated red team. The internal attacker proactively hunts for complex, long-horizon prompt injection exploits before bad actors find them in the wild, baking defensive updates directly into the model checkpoint. Why it matters: as agents get the keys to your inbox and accounts, staying ahead of prompt injections is the only way to keep them safe.
Google just launched T5Gemma 2, an encoder-decoder model family built on Gemma 3 that introduces multimodal vision and massive 128K token contexts. By tying embeddings and merging attention mechanisms, Google squeezed parameters down to compact sizes (~370M to ~7B) ideal for on-device deployment. It is available as pre-trained weights ready for downstream fine-tuning.
Google just released FunctionGemma, a lightweight model tuned from Gemma 3 270M specifically for on-device function calling. Small enough to run locally on phones and Jetson Nano boards, fine-tuning boosts its task accuracy from 58% to 85%. Why it matters: It turns edge devices into fast, fully private agents capable of triggering native actions without touching the cloud.
OpenAI just rolled out GPT-5.2-Codex, an agentic coding model tuned for complex software engineering and defensive cybersecurity. It brings native context compaction for massive refactors, better Windows support, and sharper security auditing chops. It is available now in Codex for paid ChatGPT users, with API access rolling out soon.
Google just expanded its flagship lineup with Gemini 3 Flash, delivering Pro-grade reasoning at triple the speed and a fraction of the cost of Gemini 2.5 Pro. It scored an impressive 78% on SWE-bench Verified—surpassing even Gemini 3 Pro—and is already rolling out globally as the default free model inside the Gemini app and Search.
Meta just unveiled SAM Audio, bringing its "segment anything" vision to sound. The multimodal model lets users isolate specific sounds from messy audio mixtures using text prompts, visual clicks on video objects (like a guitar in a band), or time ranges. It's powered by a diffusion transformer architecture and available to test immediately in the Segment Anything Playground.
Google just updated its Gemini 2.5 Flash Native Audio model to make real-time voice agents significantly better at multi-turn conversations and function calling. The tech also lands in the Google Translate app with a live speech-to-speech beta that streams translated audio straight into your headphones while keeping your natural tone and pitch.
OpenAI just launched the GPT-5.2 model family—spanning Instant, Thinking, and Pro tiers—built specifically for professional workflows. The Thinking variant cuts hallucinations by 30% relative to GPT-5.1 and matches or beats human professionals on 70.9% of knowledge tasks across 44 occupations. It is rolling out to paid ChatGPT subscribers and API developers today.
Runway just announced GWM-1, its first family of general world models built on Gen-4.5 that generates frame-by-frame simulations in real time. It reacts to camera movements, audio, and robot actions across three specialized versions for explorable worlds, interactive avatars, and robotics. Why it matters: robotics teams can now test policies and generate counterfactual data inside a video diffusion simulator before touching physical hardware.
OpenAI just pulled the emergency brake, with Sam Altman declaring a internal 'code red' to accelerate the release of GPT-5.2 to December 9th. The move is a direct counter to Google's Gemini 3, with internal benchmarks reportedly putting OpenAI back on top. While dates could still slip due to capacity, OpenAI is clearly feeling the heat.
Google just introduced Gemini 3 Pro, upgrading its frontier model with advanced document, spatial, screen, and video reasoning. It can pinpoint pixel-precise coordinates, analyze video at 10 FPS, and reverse-engineer messy documents straight into structured LaTeX or HTML. Why it matters: Multimodal AI is shifting from basic image recognition to driving real-world robotics and desktop agents.
Google just rolled out Gemini 3 Deep Think to Google AI Ultra subscribers in the Gemini app. Aimed at tough math, science, and logic tasks, it uses parallel reasoning to test multiple hypotheses at once, hitting 41.0% tool-free on Humanity’s Last Exam and 45.1% on ARC-AGI-2 with code execution.
OpenAI is reportedly fast-tracking a new frontier model codenamed "Garlic" for an early launch in the new year. Trained on a significantly smaller dataset than GPT-4.5 to keep costs down, it rivals Google's Gemini 3 and Anthropic's Opus 4.5 in coding and reasoning. With Google closing the user gap and Sam Altman declaring an internal code red, OpenAI is feeling the pressure.
DeepSeek just launched DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, two reasoning-first models built specifically for AI agents. The standard V3.2 is now live across its app, web interface, and API, while the higher-end Speciale reasoning model is currently limited to the API.
