New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
OpenAI just expanded its Trusted Access for Cyber program and introduced GPT-5.4-Cyber, a fine-tuned model built specifically for defensive cybersecurity work. The permissive variant lowers refusal boundaries on legitimate security tasks and adds binary reverse engineering to analyze compiled binaries for vulnerabilities. It allows verified defenders to leverage frontier capabilities without running into standard safety guardrails.
Meta just published an updated Advanced AI Scaling Framework alongside the safety report for its Muse Spark model, introducing new risk evaluations for loss of control, cybersecurity, and bioweapons. Instead of relying on hardcoded refusals, Meta is now training models to understand the reasoning behind safety rules, confirming Muse Spark lacks the autonomy to pose runaway risks.
Meta just unveiled Muse Spark, its first native multimodal reasoning model from Meta Superintelligence Labs, featuring a new Contemplating mode that orchestrates parallel agents to rival Gemini Deep Think and GPT Pro. It scored 58% on Humanity's Last Exam and integrates health reasoning trained with 1,000 physicians. It is now live on meta.ai, with private API previews underway.
World Labs just rolled out updates for its Marble spatial intelligence platform, introducing Marble 1.1 and Marble 1.1-Plus. Marble 1.1 fixes visual artifacts while improving lighting and contrast, while the new 1.1-Plus model is built for scale to generate larger, more complex interactive 3D environments.
Anthropic just launched Project Glasswing alongside Apple, Microsoft, AWS, and Nvidia to hunt down critical software flaws using Claude Mythos Preview, an unreleased frontier model. Mythos autonomously discovered thousands of zero-days across major operating systems, including decades-old bugs in OpenBSD and FFmpeg. Anthropic is offering $100M in computing credits to defenders while keeping the model private.
Meta is adopting a hybrid release strategy for its next-generation AI models under Alexandr Wang, planning to open-source smaller versions while locking down its largest frontier models. Brought in via a $15B Scale AI deal, Wang is designing the new lineup to bounce back from Llama 4's lag, bypassing OpenAI's enterprise turf to push consumer AI directly across WhatsApp and Instagram.
Alibaba just rolled out Qwen3.6-Plus, a native multimodal model built for real-world agentic workflows. It packs upgraded coding execution, sharper visual reasoning, and a 1M context window enabled by default via API, with open-source versions on the way.
Z AI just dropped GLM-5V-Turbo, a vision-coding model that turns rough sketches, screenshots, and screen recordings into runnable code with matching animations. It packs a 200K context window, beats Claude Opus 4.6 on several design benchmarks, and integrates out of the box with coding agents like Claude Code.
Alibaba's Qwen team just dropped Qwen3.5-Omni, a native omnimodal model that handles text, image, audio, and video simultaneously. The highlight is "Audio-Visual Vibe Coding," letting you talk to your camera while the AI builds a working website or game in real time. It natively processes up to 10 hours of audio and outperforms Gemini-3.1 Pro in audio tasks.
Anthropic just started testing Claude Mythos 5.0 with select early-access users after internal leaks revealed the new flagship. Sitting above the Opus tier, it brings a step change in coding, reasoning, and cybersecurity. Why it matters: The lab is being unusually cautious with rollout, as the model's offensive cyber capabilities are advanced enough to outpace current defenses.
Google just launched Gemini 3.1 Flash Live, its highest-quality audio model built for faster, more natural voice interactions. It handles complex multi-step reasoning and interruptions with ease, powers Search Live across 200+ countries, and watermarks all audio outputs with SynthID.
OpenAI is reportedly pulling the plug on Sora and all related video products, freeing up compute for its upcoming flagship model 'Spud'. The video team will pivot to robotics world simulation, leaving Disney's $1B partnership on ice. Basically, Sam Altman is ditching resource-heavy side quests to double down on the core frontier race.
Google DeepMind just showed off Gemini 3.1 Flash-Lite generating entire websites in real time. Instead of serving pre-built pages, the demo browser builds each screen dynamically as you click and search. It is an impressive showcase of how fast and responsive lightweight models have become.
Mistral AI just rolled out Voxtral TTS, its first text-to-speech model built for real-time voice agents. The 4B-parameter model runs at a 70ms latency, supports expressive speech across nine languages, and can clone a voice from a 3-second sample. It is available via API, in Mistral Studio, and as open weights on Hugging Face.
Luma Labs just introduced Uni-1, an AI model that reasons and generates pixels simultaneously. Instead of treating reasoning and visual generation as separate steps, it combines both into a single unified process that Luma calls less artificial and more intelligent.
Sakana AI just launched Sakana Chat, powered by its new Namazu alpha series that adapts open-weight frontier LLMs for Japan through post-training. The models slash DeepSeek's 72% question refusal rate on sensitive topics to nearly zero while keeping benchmark parity in reasoning and coding. The web chat also packs real-time search capabilities.
Cursor just launched Composer 2, its new frontier-level coding model now available directly in the editor. Trained via continued pretraining and reinforcement learning on long-horizon tasks, it hits 61.3 on CursorBench while cutting pricing to $0.50/M input and $2.50/M output tokens. A faster variant is also available and set as the default option.
Xiaomi just officially launched MiMo-V2-Pro, its flagship agent model that previously dominated OpenRouter under the stealth codename "Hunter Alpha." It packs over 1T total parameters (42B active) and a 1M context window, claiming to outcode Claude 4.6 Sonnet on serious software tasks. Why it matters: Hardware giants are racing to build the central brains for autonomous AI agents.
MiniMax just launched M2.7, a reasoning-focused model that autonomously handled 30% to 50% of its own reinforcement learning workflow, including debugging and metric analysis. It hit a 66.6% medal rate on MLE Bench Lite and scored 56.22% on SWE-Pro. Why it matters: it marks an early shift toward recursive self-improving agent architectures, available now via API.
OpenAI just launched GPT-5.4 mini and nano, lightweight models built for high-volume workloads and tool-heavy coding. The mini variant runs over 2x faster than GPT-5 mini while nearly matching the flagship GPT-5.4 on SWE-Bench Pro, at just $0.75 per million input tokens. Nano drops the entry cost to $0.20, perfect for background subagents.
