Models, releases, capabilities, reality checks
OpenAI's record $110B funding round at $730B valuation redefines AI infrastructure investment. Amazon, NVIDIA, and SoftBank lead the historic raise.
AI market hits $62.7B in 2025 with 78% enterprise adoption. But only 6% achieve significant ROI. Here's what separates winners from experimenters.
OpenAI raises $110B at $730B valuation. Amazon's $50B investment secures exclusive AWS access to Frontier platform and 2GW Trainium commitment.
OpenAI reveals 59.4% of SWE-bench Verified problems have flawed tests, with widespread contamination across frontier models. The AI coding evaluation crisis explained.
IBM and UC Berkeley reveal why enterprise AI agents fail: 13.8% SRE success, 0% FinOps success. MAST framework identifies 14 failure modes for systematic debugging.
GGML and llama.cpp join Hugging Face to streamline local AI development with single-click Transformers integration while maintaining open-source licensing.
NVIDIA Cosmos Policy achieves a 98.5% success rate in robot control by unifying perception and action through generative world models.
Google DeepMind's Project Genie and NVIDIA Cosmos Policy mark the shift from static AI to 'World Models' that understand physics and interactivity.
Google DeepMind's Project Genie creates interactive, playable worlds from a single image, trained on 200,000+ hours of video without manual coding.
STEM replaces FFN up-projections with token-indexed embeddings, cutting FLOPs by 33% and boosting reasoning tasks by 4% without MoE routing issues.
Anthropic's Claude Cowork introduces sub-agent coordination to automate desktop tasks within a secure virtual machine, marking a shift to autonomous AI employees.
Overworld Labs releases Waypoint-1, the first 60 FPS open-source world model for consumer GPUs, enabling real-time interactive AI environments locally.
Microsoft's Diff-Attn V2 brings noise-canceling attention to the mainstream with native FlashAttention support and 2.5x faster long-range processing.
Google DeepMind's MedGemma 1.5 brings 3D MRI/CT analysis and MedASR dictation with 82% fewer errors. A new era for multimodal healthcare AI.
Microsoft's OptiMind is a 20B parameter SLM that turns business problems into math code, outperforming models twice its size in optimization tasks.
OpenAI launches GPT-5.2 and an $8/month 'ChatGPT Go' tier. Discover the new 'xhigh' reasoning models, ad-supported tiers, and the August 2025 knowledge cutoff.
OpenAI signs a $10B deal with Cerebras for 750MW of compute, targeting 15x faster AI inference to power the next generation of real-time AI agents.
Google DeepMind's Veo 3.1 introduces 'Ingredients to Video' for visual consistency and native 9:16 support, topping MovieGenBench in prompt adherence.
NVIDIA Cosmos Reason 2 tops the Physical AI Bench with a 256K context window, bringing advanced spatial reasoning to the next generation of open-source robotics.
Google's Gemini 3 Flash delivers 78% SWE-bench accuracy at $0.50/M tokens—3x faster than its predecessor. Here's what developers need to know.
Nvidia's $20B acquisition of Groq's LPU assets marks its biggest deal ever, targeting the booming AI inference market with specialized chip technology.
OpenAI deploys RL-driven security for ChatGPT Atlas, using automated red teaming to combat prompt injection in agentic AI systems.
ServiceNow releases AprielGuard, an 8B open-source model achieving 0.96 F1 on adversarial attacks to secure agentic AI workflows.
OpenAI's new research shows monitoring AI reasoning is far more effective than output checks—but this transparency is fragile and may not last.
Google tests a redesigned Gemini interface with floating input field and consolidated tools menu, targeting its 450M monthly active users.
Manus launches Design View with Mark Tool for precision AI image editing—targeting specific elements without full regeneration.
Z.ai releases GLM-4.7 with 73.8% on SWE-bench, 200K context, and integration with Claude Code, Cline, and more. Starting at $3/month.
Google's Gemini 3 Flash outperforms its Pro predecessor on 18 of 20 benchmarks while running 3x faster at 69% lower cost. Here's what it means for developers.
OpenAI releases GPT-5.2-Codex for agentic coding with a new 13-evaluation CoT safety framework, while Google counters with Gemini 3 Flash.
MIRA OS, released December 20, 2025, introduces a persistent AI entity with novel memory decay and modular design, enhancing AI agent research.
OpenAI's new slider controls let users fine-tune ChatGPT's enthusiasm and warmth—a small feature with big implications for AI personalization.
OpenAI opens ChatGPT app submissions, enabling external context integration. New research details a UX framework for managing AI context: library, conversation, and memory.