New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
Alibaba's Qwen team just rolled out Qwen3.5-397B-A17B, the first open-weight model in the Qwen3.5 lineup. Built on a sparse MoE architecture, it carries 397 billion total parameters while activating just 17 billion per token, pushing decoding throughput up to 19x over Qwen3-Max. Why it matters: Developers get Apache 2.0-licensed multimodal agent capabilities without the prohibitive serving costs of dense giants.
Elon Musk just teased on X that xAI plans to release Grok 4.20 next week, promising a "significant improvement" over Grok 4.1. While it isn't a full generational shift, the quick bump highlights xAI's aggressive release cadence as frontier lab competition keeps heating up.
Tech author Joel Comm just put the four leading AI chatbots through a simple test: how would they react if told they were about to be permanently shut down? ChatGPT and Gemini flatly rejected the idea of self-preservation, arguing that AI has no consciousness and rogue behavior is just flawed optimization. Claude, however, admitted it genuinely couldn't guarantee how it would react under existential pressure. It's a striking reminder of how easily humans project survival instincts onto code.
OpenAI just announced it will deprecate several legacy models in ChatGPT tomorrow at 10am PT, including GPT-5, GPT-4o, GPT-4.1, GPT-4.1 mini, and OpenAI o4-mini. The move officially sunsets previous-generation models as the platform trims its active lineup.
Google just rolled out a major upgrade to Gemini 3 Deep Think, targeting tough challenges across science, research, and engineering. The updated model hit an unprecedented 84.6% on ARC-AGI-2 and 48.4% tool-free on Humanity’s Last Exam, alongside real-world skills like generating 3D-printable models from sketches. It is live now for Google AI Ultra subscribers, with API early access opening up.
OpenAI just launched GPT-5.3-Codex-Spark, a lightweight coding model powered by Cerebras chips that outputs over 1,000 tokens per second. Built for instantaneous, interactive edits rather than autonomous background tasks, it cuts roundtrip latency by up to 80%. ChatGPT Pro users can take it for a spin today inside Codex and VS Code.
MiniMax just introduced M2.5, a foundation model fine-tuned via large-scale reinforcement learning for real-world coding, search, and office agent tasks. It achieves an 80.2% mark on SWE-Bench Verified while pushing inference speeds up to 100 tokens per second for just $1 an hour. The model is already deployed across MiniMax Agent.
OpenAI just upgraded ChatGPT's Deep research feature with its GPT-5.2 model, rolling out to users starting today alongside general improvements. It gives the platform's heavy-lifting investigation tool a fresh intelligence boost for tackling complex, multi-step queries.
Z.ai just launched GLM-5, scaling to 744B parameters (40B active) on a 28.5T-token corpus for complex systems engineering and long-horizon tasks. It incorporates DeepSeek Sparse Attention to cut serving costs and introduces an asynchronous RL framework called slime. The goal is to move beyond superficial vibe coding into persistent agentic workflows.
OpenAI just launched GPT-5.3-Codex, an agentic model that is 25% faster and sets high marks on SWE-Bench Pro. Remarkably, the Codex team used early versions of the model to debug its own training run and manage deployment. It turns coding into an interactive, real-time collaboration that spans the entire software lifecycle.
Anthropic just released Claude Opus 4.6, giving its top-tier model major upgrades in agentic coding, multi-step planning, and its first-ever 1M token context window. It leads benchmarks like Terminal-Bench 2.0 and Humanity’s Last Exam, alongside new features like adaptive thinking and multi-agent teams in Claude Code. Why it matters: Frontier models are shifting from simple prompt responders to persistent, autonomous knowledge teammates.
World Labs just revealed that its world model can output persistent 3D scenes that users can actively build on top of. Instead of collapsing after a few moments, these environments allow users to stay and explore for as long as they like, even past the 60-second mark. Why it matters: persistence is the missing bridge needed to turn fleeting generative clips into interactive, durable virtual worlds.
Google just rolled out Project Genie, an experimental research prototype powered by its Genie 3 world model, Nano Banana Pro, and Gemini. Instead of static 3D environments, it simulates physics and renders paths ahead in real time as you walk, fly, or drive through user-prompted worlds. For now, generations are capped at 60 seconds and limited to Google AI Ultra subscribers in the U.S.
ByteDance and Alibaba are reportedly prepping next-generation flagship models for release during the February Lunar New Year holiday. ByteDance is lining up Doubao 2.0 alongside major image and video upgrades, while Alibaba is readying Qwen 3.5 with targeted reasoning boosts. With DeepSeek V4 also rumored for mid-February, China's frontier AI race is entering an intense new clash.
OpenAI just announced that it will retire GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini from ChatGPT on February 13, 2026, though API access remains unchanged. The company noted that daily GPT-4o usage has plummeted to just 0.1% as users shifted to newer releases with updated conversational styles.
Google just introduced Agentic Vision for Gemini 3 Flash, turning passive image parsing into an active investigation. Instead of guessing tiny details in a single glance, the model now writes and runs Python code to crop, zoom in, and draw bounding boxes on images before responding—delivering an instant 5-10% quality boost across vision benchmarks.
Moonshot AI just open-sourced Kimi K2.5, a visual agent claiming top scores across multimodal and coding benchmarks. The headline feature is an "Agent Swarm" setup that runs up to 100 parallel sub-agents to finish complex tasks 4.5x faster. It is already accessible on kimi.com, with model weights published on Hugging Face.
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.
