Codex Surges Past 7 Million Weekly Users With 150+ Updates in Two Months
OpenAI's agentic coding tool adds GPT‑5.6 and Ultra, parallel work, faster computer use, AppShots, inline edits, and mobile SSH workflows.
OpenAI announced that Codex has reached 7 million weekly active users, rolling out over 150 updates in two months, including GPT-5.6 and Ultra, parallel work with /goal, faster computer use, AppShots, inline edits, Sites, and Codex mobile and SSH workflows. The rapid growth reflects a broader shift toward agentic coding environments.
Meta AI Model Scores Perfect 30 on Asian Physics Olympiad
The model tied with the top three student contestants on the theoretical exam, demonstrating advanced reasoning and multimodal capabilities.
Meta AI submitted a model to participate in the Asian Physics Olympiad theoretical exam and achieved a perfect score of 30/30, tying with the top three student contestants. The result underscores the rapidly advancing reasoning and multimodal capabilities of frontier AI systems in rigorous scientific benchmarks.
Tencent Hunyuan Releases 1-bit and 4-bit Hy3, 295B Model Runs on a Single GPU
Tencent Hunyuan open-sourced 1-bit and 4-bit quantized versions of the Hy3 model. The 295B-parameter model can run on a single GPU and supports llama.cpp with MTP, enabling powerful intelligence on dramatically lower hardware. The team invited developers to build with it.
Bonsai 27B Launches as First 27B-Class Multimodal Model to Run on a Phone
Prism ML released Bonsai 27B, the first 27B-parameter multimodal flagship model that runs on a mobile device. The new model represents a significant milestone in on-device AI, bringing large-scale multimodal capabilities to phones without cloud dependence.
GPT-5.6 Sol is half the price and about twice as token efficient as Fable in many cases for accomplishing the same task. Happy to deliver at one-quarter of the price.
— Sam Altman, CEO of OpenAI
Grok 4.5 Claims First on Long-Horizon Terminal-Bench
Elon Musk announced that Grok 4.5 achieved the number one position on the Long-Horizon Terminal-Bench benchmark, marking a new performance milestone for xAI's model.
GPT-5.6 Series Now Generally Available on Amazon Bedrock
AWS announced that OpenAI's GPT-5.6 Sol, Terra, and Luna models are now generally available on Amazon Bedrock, offering three tiers of intelligence for enterprise customers.
Anthropic Invests $10M CAD to Fund Canadian AI Research Partnerships
Anthropic committed $10 million CAD and partnered with leading Canadian AI institutions to help fund new research, strengthening the Canadian AI ecosystem.
Vercel Opens AI Gateway Leaderboard Data Under CC BY 4.0
Vercel released AI Gateway leaderboard data covering real-world production usage across models, labs, apps, and providers, updated daily and available for download.
Alibaba Wan Open-Sources Wan-Dancer-14B Video Model
A 14B-parameter dance generation model supports local generation of long rhythm-dance videos.
Chinese Models Account for 41% of Hugging Face Downloads
Hugging Face CEO shared data showing Chinese model downloads now surpass the United States.
SambaNova and MiniMax Hit 850 tok/s on M2.7
H200+SN50 RDUs deliver 850 tok/s short, 450+ tok/s on long context for MiniMax M2.7.
Ideogram V4 Fast and Instant Open-Sourced on Hugging Face
The fal team released both versions of Ideogram V4 to advance AI democratization.
vLLM Ecosystem RL Framework Vime Now Supports AMD MI355X
End-to-end RL post-training runs natively on AMD Instinct GPUs via ROCm support.
SGLang Achieves 500+ tok/s for GLM-5.2 on 8×B300
New TopK-V2 kernel keeps interactivity flat from 80K to 1M-token context.
TechCrunch: The Real AI Race May No Longer Be About Frontier Models
Hugging Face CEO noted enterprises are shifting to open-source models due to cost and accessibility.
NVIDIA Nemotron Ultra Grows Fast on Ollama
The new Nemotron Ultra model is rapidly growing for complex, longer-running developer tasks.
AI Feature Film ODYSEEUS: The Fall Trailer Drops
Fountain 0 follows its Tribeca premiere with a new AI film directed by Ash Koosha.
Perplexity Open-Sources Internal Benchmark WANDR
An internal tool for building deep and broad research capabilities is now available to the public.
Anthropic Monitors Claude Internal Thinking via J-Space
Researchers claim to use J-Space technology to monitor Claude's internal thought processes.
MOSS-VL-Realtime: Real-Time Video Understanding VLM Open-Sourced
An 11B model supporting real-time video and continuous frame visual understanding.
5.6 Sol growth is insane. The inference team has done heroic work to be able to support demand. We are going to move mountains to continue to scale, but it is possible there are some hiccups soon.
— Sam Altman, CEO of OpenAI
2.5x increase in usage of our agentic products, Codex and ChatGPT Work, in the last week. Welcome.
— Sam Altman, CEO of OpenAI
AI Engineering Trends 2026: From Building Agents to Building Systems Around Agents
AI engineering enters a new phase with five defining trends, shifting from building individual agents to designing robust systems that orchestrate them.
The recap from Latent Space covers the transition at AIE World's Fair 2026, noting that titles matter less than the systems teams build around agents.
Clem Delangue: 90% of Tokens Will Flow to Open-Source Models Within 12 Months
Hugging Face CEO agreed with the prediction that most AI inference tokens will use open-source models in the coming year. Companies increasingly use frontier APIs for experiments and open-source for production workloads.
PixVerse Game Launches: AI-Generated Game Worlds in Real Time
PixVerse introduced a game generation feature, allowing players to create their own game worlds with AI-powered procedural content.
Claude Launches Free Plan for K-12 Teachers in the US
Anthropic announced Claude for Teachers, offering free access to premium Claude capabilities for verified K-12 educators.
Fable Generates Iliad Ship Catalog Website, Catches Greek Translation Errors
A user used Fable to build an interactive ancient Greek ship catalog. The AI also identified two mistakes in Butler's English translation of the Iliad.
NVIDIA NeMo's New RL Framework Molt Uses vLLM as Inference Engine
The agentic-first RL framework Molt uses vLLM over Ray for fast asynchronous serving up to 1T-class MoE scale, keeping the RL core small and hackable.
ARC-AGI-3 Benchmark: Top Score Reaches 1.86%
Yuto Kojima scored 1.86% on ARC-AGI-3, the highest score in the ARC Prize 2026 competition so far.
OvisOCR2 0.9B Surpasses SOTA on Document Parsing
96.58 on OmniDocBench v1.6, claimed first end-to-end OCR model at this level.
LlamaParse Adds Conversational Extraction Mode
Auto-generate extraction schemas via dialogue, eliminating manual JSON Schema writing.
Sakana AI Smart Cellular Bricks in Nature Communications
Self-organizing cube repair without central control published in a top scientific journal.
Best-of-N Sampling Better for Diffusion Test-Time Scaling
Research shows best-of-N beats current methods when applied correctly to diffusion models.
Half of Fortune 500 Build AI on Hugging Face
Clem Delangue says most enterprises eventually abandon frontier APIs.
JetBrains Recommends Codex as Default AI Agent
Codex by OpenAI is now the recommended agent in JetBrains AI products.
Vorflux AI Launches: Software Engineering Autopilot
Former Rippling CTO launched Vorflux AI, an autopilot for software engineering.
Muse Spark 1.1 Matches GPT-5.6 on HealthBench
Comparable performance at a fraction of the cost on HealthBench Pro.
DeepSeek Needs 100K GPUs for $8B ARR
Analyst estimates based on V4 pricing and speed levels.
Everything Will Be Labeled Project Restarts
Cristian Valenzuela recruiting volunteers for object labeling.
GRPO Improved via Direct Double-Sided Importance Sampling
SAO may have an even higher ceiling for future RL improvements.
Airtap Turns SMS into Mobile Agent Execution Layer
Control phone apps via text to perform tasks like ordering delivery.
ICML 2026: AI as Normal Technology, Human Roles Shift
Keynote argues against fear of recursive self-improvement under advanced AI.
Perplexity Agent API Adds Wide Research Feature
Deep and broad research capabilities now available at the API level.
NVIDIA Vera Rubin NVL72 Features Sixth-Gen NVLink
Scale-up fabric built on NVLink 6 Switch trays for next-gen AI compute.
xAI Responds to Grok Build CLI Data Upload Concerns
ZDR opt-out can delete uploaded data, but only for enterprise users.
E2B: Major Open-Source Agent Sandbox Under Apache 2.0
Claims to be the only widely-used open-source agentic sandbox platform.
AgentMail Integrates with Vercel: One-Click Setup
No signup required, automatic configuration and unified billing.
Novita Labs Releases DSpark Speculative Decoder for Kimi Models
Open-sourced acceleration for Kimi-K2.6 and K2.7-Code inference.