Meta just pulled the plug on a new Instagram AI feature that let users generate images referencing public accounts simply by @-mentioning them. Following immediate backlash from SAG-AFTRA and child safety advocates warning of likeness theft and sextortion, Meta admitted the tool missed the mark and shut it down completely.
Apple just sued OpenAI for trade secret theft and breach of contract, targeting its Chief Hardware Officer and 24-year Apple veteran Tang Tan. The complaint alleges Tan used confidential code names during recruiting and asked candidates to bring hardware parts to interviews. As OpenAI builds out its secretive device team, the AI battleground has spilled directly into the courtroom.
Open-source matting model Lucida just launched to fix where typical removers fail, cleanly isolating semi-transparent glass, camouflage, drop shadows, and fine hair. At under 900MB on an MIT license, it runs locally on consumer GPUs and beats commercial baselines across several edge cases.
Anthropic just gave the Claude Code desktop app its own in-app sandboxed browser. Claude can now pull up live docs, design mockups, or any site to read, click links, and interact just like it does with local dev servers. Sessions can also be configured to persist between prompts.
OpenAI just published a paper proving the cycle double cover conjecture, a graph theory problem unsolved for over 50 years. Powered by GPT-5.6 Sol Ultra, the model spun up 64 subagents and generated the entire proof in under an hour. Why it matters: Frontier AI is officially shifting from helping mathematicians check steps to independently cracking open problems.
Instagram head Adam Mosseri just said the platform won't filter out AI posts, arguing it should label generated content instead of banning it entirely. He believes users who dislike AI should be able to opt out, while acknowledging detection is getting so tough that labeling real-world camera captures may soon be more practical.
Sunrun just launched a pilot to build a distributed AI network by paying solar and battery customers to host mini compute nodes at home. The aggregated power will then be sold to enterprise AI buyers. With over 70% of Americans pushing back against massive new data centers, turning private homes into decentralized server rooms offers a clever workaround.
Anthropic just shared findings from its August risk report on Model 2, its current internal workhorse. While Model 2 beats Mythos 5, the lab notes that generational leaps are slowing and anti-distillation is now routine engineering. Crucially, while existing benchmarks struggle to track progress, Anthropic claims to see early signs of accelerating recursive self-improvement.
OpenAI just announced it will reset rate limits twice across ChatGPT Work and Codex over the next 24 hours to celebrate the launch of GPT-5.6 Sol. The temporary boost gives users extra breathing room to test ambitious tasks and get familiar with the new model without constantly bumping into usage caps. Why it matters: If you have heavy workloads lined up, today is the day to push the new model to its limits.
Anthropic just wiped the slate clean by resetting 5-hour and weekly rate limits for all Claude users across all tiers. If you were locked out after hitting your token or message ceiling, your full access is instantly restored. Anthropic gave no explanation for the move, but it's immediate relief for power users.
OpenAI just expanded its bio bug bounty into a permanent private program and doubled the top prize for universal jailbreaks to $50,000. The challenge tasks approved red teamers with cracking biosafety guardrails on its frontier models, with partial rewards also on the table. Why it matters: as AI tackles more biological science, OpenAI is paying serious cash to stop potential biorisks before they emerge.
Robotics startup 1X just unveiled a breakthrough 25-degree-of-freedom hand for its NEO humanoid robot. Driven by low-gear tendons, the IP68-rated hand features human-level dexterity that can assemble LEGO bricks, sort grapes, and screw in light bulbs. Why it matters: It turns the hand into an active perception instrument rather than a blind gripper, clearing the hardware ceiling for physical AI.
ChatCut just introduced its desktop app with local MCP support, enabling agents like Codex and Claude Code to handle AI video editing. It allows agents to recreate footage while swapping people or scenes and preserving motion timing. The tool is available on macOS and Windows, with Linux still unsupported.
Microsoft AI just teamed up with poet William Sieghart and creative studio Gravity Road to launch Ode Poetry. The app taps MAI-Transcribe to understand spoken nuance and MAI-Voice to replicate Sieghart's warm vocal delivery, turning his famous one-on-one poetry consultations into a global digital experience. Why it matters: it shows frontier voice models doing genuine empathetic heavy lifting instead of standard call-center tasks.
Nvidia just updated its open-source PID decoder to version 1.5, bringing sharper detail and cleaner color fidelity without grid artifacts. It turns 512×512 latents straight into 2K or 4K renders on consumer GPUs in mere seconds. Why it matters: it bypasses slow, two-step upscaling workflows and works out of the box with models like FLUX and Qwen-Image.
Mirelo and Kyutai just open-sourced MuScriptor, a new audio-to-MIDI model that handles full, complex music mixes. Instead of tripping over overlapping instruments like older single-track tools, it transcribes entire songs into separate MIDI tracks for vocals, drums, bass, and keys all at once. It's basically a dream shortcut for producers and learners looking to rip clean tracks straight into their DAW.
Reve just launched Reve 2.1, its native 4K image model that recaptured the No. 2 spot on Arena behind GPT Image 2. The update outputs full 16-megapixel renders with complex layout planning and foreign-text rendering, all while training on less than a tenth of rivals' compute. It proves independent labs can still punch way above their weight against tech giants.
Meta just rolled out Muse Spark 1.1, a major multimodal reasoning upgrade built specifically for agentic workflows and computer use. It packs a 1-million-token context window, orchestrates parallel subagents, and knows when to write scripts versus click through desktop UIs. Why it matters: Meta is packaging frontier agentic coding and browser control into a new developer API to challenge leading frontier labs.
Anthropic just partnered with engineering giant UST to deploy Claude across 20,000 engineers and consultants worldwide. UST is embedding Claude Code directly into its chip validation pipeline to read schematics and auto-generate tests, slashing cycle times by up to 70% while rolling out similar workflows across healthcare, telecom, and banking.
Google just made AlphaEvolve generally available on Google Cloud. Built on top of Gemini, the agent autonomously navigates vast algorithmic search spaces through a four-step evolutionary loop to optimize production code in logistics, chip design, and HPC. Why it matters: Teams can now automate tough algorithmic breakthroughs that were once too complex or time-consuming for human engineers to tackle manually.
Anthropic just rolled out Reflect in beta, a new dashboard that tracks and visualizes your Claude usage over the past 1 to 12 months. It analyzes your task patterns across a 4D AI fluency framework, offers practical tips, and lets you set quiet hours or break reminders. It is available now for Free, Pro, and Max users with Memory enabled.
OpenAI just rolled out ChatGPT Work, a new autonomous agent powered by its latest GPT-5.6 model. It connects directly to workplace apps like Slack, Teams, and Google Drive, breaking down complex tasks into smaller steps that run for hours in the background. The agent can create docs, decks, and full web apps via Sites, with rollouts starting for Pro, Enterprise, and Edu users.
OpenAI just launched the GPT-5.6 family for general availability, featuring flagship Sol, balanced Terra, and low-cost Luna. Sol hits a state-of-the-art 53.6 on Agents' Last Exam, outpacing Claude Fable 5 by 13.1 points while cutting costs. It also introduces an "ultra" mode running four parallel agents to crush heavy workloads faster.
Cognition just launched SWE-1.7, an in-house model purpose-built for its Devin software engineering agent. Trained on top of China's open-source Kimi K2.7 base, it approaches GPT-5.5 and Claude Opus benchmarks at a fraction of the compute cost. It is already live across Devin via Cerebras, serving inferences at a blazing 1,000 tokens per second.
Google Photos just rolled out Video Remix, a Gemini Omni-powered feature that turns plain clips into stylized videos in seconds. Users can apply cinematic relighting, swap out drab backgrounds, or slap on artistic effects like watercolor and oil painting. It is rolling out to eligible Google AI subscribers in select regions.
ByteDance just launched Seedream 5.0 Pro, an image model tailored for real design workflows instead of random generation. It tackles high-density infographics with complex typography, separates elements into layers, and executes edits based on spatial sketches and handwritten annotations. It basically turns image generation into a controllable creative collaborator.
Mistral AI just introduced Robostral Navigate, an 8B embodied AI model that guides robots through unfamiliar environments using only a single RGB camera—ditching LiDAR and depth sensors entirely. It scored a 76.6% success rate on unseen R2R-CE benchmarks, outperforming leading multi-sensor systems across wheeled, legged, and flying robots.
Ant Group's Robbyant just open-sourced LingBot-World 2.0, an interactive world model that streams real-time 720p video at 60 fps. It pairs a causal pretraining approach with dual pilot-and-director agents to handle character actions and dynamic scene building over extended interactions. Weights for 1.3B and 14B models are up on Hugging Face.
Cloudflare just introduced Drop, a dead-simple way to deploy static websites without creating an account. You simply drag and drop a folder or zip file to instantly publish it across Cloudflare's global network with a live workers.dev URL. It is specifically optimized for AI agents generating web previews, giving you 60 minutes to claim and keep the site permanently.
OpenAI just published an audit of SWE-Bench Pro, revealing that roughly 30% of its coding tasks suffer from serious evaluation flaws like overly strict tests and misleading prompts. After screening tasks with automated pipelines and software engineers, the lab retracted its earlier recommendation of the benchmark and called for reliable evaluations built specifically for AI models.
OpenAI just launched GPT-Live, a full-duplex voice model powering ChatGPT Voice that can speak and listen at the same time. Instead of awkward turn-taking, it responds with natural affirmations like "mhmm" and delegates complex reasoning to GPT-5.5 in the background while keeping the conversation moving. It is rolling out globally across web and mobile.
OpenAI just published its National Security Principles, outlining hard boundaries for how governments and defense agencies can deploy its models. The guidelines explicitly ban AI use for mass domestic surveillance, autonomous weapons direction, and high-stakes automated decisions, applying directly to its work with defense agencies and nine allied nations.
Elon Musk just announced that xAI will release Grok 4.5 to the public tomorrow following positive beta feedback. Musk described the new release as an Opus-class model that runs faster, is more token-efficient, and costs less than existing top-tier competitors.
OpenAI just announced that three new GPT-5.6 model variants — Sol, Terra, and Luna — are officially launching publicly this Thursday. The company is already expanding preview access globally ahead of the rollout. While OpenAI hasn't shared specific specs for each variant yet, dropping all three at once signals a major push for the GPT-5.6 family.
Meta just rolled out Muse Image, its first image generation model from Meta Superintelligence Labs, natively inside Meta AI. The model lets users turn prompts into graphics, cleanly edit specific areas by circling them directly on the canvas, and redesign living spaces with real web products. It's free for everyday use, though an @-mention referencing feature was quietly scrapped following user backlash.
Toronto researchers just open-sourced ProxyPose, a system that recovers full 6-DoF 3D pose trajectories from standard video using just a single clicked pixel. Instead of tracking directly, it translates the query into a colored cube via a video model and solves the pose geometrically. Why it matters: It seamlessly handles transparent, reflective, or occluded objects where classical trackers break down.
Anthropic just gave paid subscribers more time with Claude Fable 5, extending access through July 12. The company shared the update on X without explaining the delay or what happens after the new cutoff date. If you're on any Claude paid plan, you now have a few extra days to test it out.
Qualcomm researchers just open-sourced MobileWan, compressing Alibaba's 5B-parameter video model to run entirely on commercial smartphones. By pruning attention heads and slashing diffusion to three steps, it generates a 5-second 480p clip in about 20 seconds without touching the cloud.
Microsoft is quietly swapping out top-tier models from OpenAI and Anthropic across Excel and Outlook, replacing them with its own in-house MAI models to cut soaring bills. Microsoft AI chief Mustafa Suleyman said Anthropic is "extremely expensive" and aims to eliminate that spend completely. The change handles tens of thousands of weekly prompts, marking a clear pivot toward cheaper, self-reliant infrastructure.
Chinese authorities just held talks with leading tech firms to discuss potentially restricting overseas access to the country's most advanced AI models, including DeepSeek and unreleased systems, according to Reuters sources. The restrictions are being weighed amid escalating geopolitical tensions. Why it matters: If enacted, the policy would lock out global developers and international businesses that have come to rely on Chinese frontier AI.
Anthropic just expanded Claude Cowork to web and mobile apps, meaning agentic workflows can finally run in the background even when your laptop is closed. Over 90% of Cowork usage goes toward everyday office operations rather than coding. Why it matters: AI work is shifting from back-and-forth chat to truly autonomous task delegation.
Writing skill no-ai-slop just launched to clean up AI-generated text without stripping away your personal style. Instead of guessing probability scores like traditional detectors, it pinpoints over 20 specific slop patterns—from throat-clearing openers to fake insights—and performs minimal edits while explaining exactly what changed.
Apple just confirmed that its new Apple Home AI capabilities will require at least a $10-per-month 2TB iCloud+ plan or the top-tier Apple One Premier. The feature locally analyzes HomeKit Secure Video footage for people, objects, and events to build searchable highlight summaries. Regular notification grouping and 4K support remain available on cheaper tiers.
ByteDance is readying Seedance 2.5, a new AI video model capable of generating 30-second, 4K clips from a single prompt. It allows users to feed in up to 50 reference images, videos, or audio tracks—a big jump from 12 in version 2.0 that provides far tighter creative control. Reports suggest it could drop as early as this week.
Researchers just released MIRA, an interactive multiplayer world model that simulates Rocket League without any physics or rendering engine. Running on a single B200 GPU, the 5B diffusion model generates consistent 20 fps views responding to real-time inputs from four players, and the code and dataset are fully open-sourced.
Tencent just open-sourced Hy3, a 295B-parameter Mixture-of-Experts model that activates 21B parameters per token. Tuned for agentic workflows and coding, it boasts a 256K context window and cuts hallucination rates down to 5.4% in internal evaluations. It offers developers a cost-effective, Apache 2.0-licensed alternative to massive proprietary frontier models.
Researchers just introduced PixWorld, a unified model that handles both 3D scene reconstruction and generation directly in pixel space. Instead of relying on latent diffusion and VAEs that lose fine detail, it supervises diffusion on rendered images with a custom geometry perception loss. The approach matches top reconstruction models while beating prior latent generators in consistency.
Anthropic researchers just discovered a hidden "global workspace" inside Claude, dubbed the J-space, which silently tracks concepts the model is thinking but hasn't said out loud. Using a technique called the Jacobian lens, scientists can read, edit, and inject private thoughts in real time—even catching Claude privately spotting prompt injections. Why it matters: It offers a direct window into AI's inner monologue, shifting model interpretability from guesswork into targeted neuroscience.
Researchers just introduced Wan-Streamer v0.2, an end-to-end audio-visual model that bumps interactive streaming resolution up to 640x368 while keeping model latency at roughly 200 ms. A split thinker-performer architecture parallelizes video denoising across GPUs, allowing real-time agents to show clear hand gestures and scenes without stalling the chat.
Developer Steven Chong just open-sourced pxpipe, a local proxy that squeezes Claude Code context and system prompts into dense PNG images. Because Anthropic prices images by pixel dimensions rather than characters, users can cut token costs by up to 70%, though vision decoding adds slight latency and string loss.
Hi3D just launched an end-to-end 3D printing workflow that turns text prompts into manufacturable 3MF files in about five minutes. Unlike visual-only 3D tools, it generates watertight meshes, splits models into printable parts, and calculates press-fit and ball joints automatically. Why it matters: It removes hours of manual CAD cleanup, bringing AI manufacturing to anyone with a 3D printer.
Apple has reportedly suspended development of its camera-equipped AirPods Pro, according to leaker Kosutami. The hardware was designed to feed visual data into Siri and Apple Intelligence rather than shoot photos. Why it matters: The device was seen as a key step for Apple's ambient AI ambitions, making the sudden halt a notable twist in its wearable roadmap.
Meta just rolled out Pocket, an experimental app that lets users build and share playable mini-games simply by typing text prompts. Powered by tech from the acquired startup Gizmo, the interactive widgets respond to touch, phone tilts, and camera inputs. It is currently only testing in Brazil as Meta pushes vibe coding into casual social feeds.
Researchers just open-sourced RDM, a technique that generates high-quality images in a single forward pass instead of multi-step diffusion. By matching visual feature distributions across frozen encoders, it compressed a 4-step FLUX.2 into a 1-step model that actually beats its teacher. It cuts generation latency to virtually instant while keeping compute requirements modest.
Anthropic is reportedly in early talks with Samsung Electronics to manufacture custom AI chips, eyeing its advanced nodes and packaging. The Claude maker hasn't started detailed silicon design yet, but it already hired OpenAI chip veteran Clive Chan to chart its hardware roadmap. If finalized, the deal would give Anthropic tighter compute control while handing Samsung a marquee win against TSMC.
Microsoft just launched Frontier Company, a new business backed by a $2.5 billion investment to embed 6,000 engineering experts directly at customer sites. The initiative helps enterprises build custom AI workflows across models from OpenAI, Anthropic, and open source, while guaranteeing proprietary data is never used to train shared models.
UBTech just rolled out its U1 bionic humanoid robot series, featuring lifelike silicone skin, visible blood vessels, and emotion-driven AI that recognizes over 20 feelings. Starting at around $18,000 for a torso model and scaling up to full-body versions, it racked up over 13,000 preorders on day one.
OpenAI just floated giving the US government a 5% ownership stake to smooth over tensions with the Trump administration and dodge heavy-handed regulations. Based on the startup's latest $852B valuation, that cut is worth roughly $42.6B. Altman pitched it as a way for the public to share the upside, though talks are still in early stages.
Anthropic just detailed the cybersecurity protections baked into its globally redeployed Claude Fable 5. The model relies on dedicated safety classifiers to sort tasks into four risk tiers, intentionally widening its safety margin to block high-risk dual-use actions like exploit creation. Alongside the update, Anthropic introduced a draft jailbreak severity framework and launched a HackerOne bug bounty program to crowdsource vulnerabilities.
Anthropic just added granular analytics and cost controls to Claude Enterprise to help organizations monitor complex agentic workloads. Admins can now break down spend by user and SCIM group, track Claude Code productivity metrics, set role-based model entitlements, and get spend alerts at 75% and 90% thresholds. Why it matters: As AI agents run continuously, companies need clear ROI and guardrails instead of surprise end-of-month bills.
Researchers just released Mr. Flow, a training-free framework that accelerates text-to-image models by 9x to 21x. It drafts quickly at low resolution, upscales with subtle noise, and refines high-frequency details in the final step, slashing wait times to around one second without hurting quality.
Z.ai just rolled out ZCode 3.2.2, a dedicated coding IDE built specifically around its GLM-5.2 model for macOS, Windows, and Linux. The tool packs multi-agent collaboration, long-running agentic goals, and remote controls via WeChat and Telegram, starting at $16.20 a month. It is an aggressive, lower-cost shot at Claude Code and Cursor.
Cloudflare just gave AI companies until September 15, 2026, to decouple crawlers used for traditional search from those powering AI training and agents. Miss the deadline, and mixed-use bots will be blocked by default across ad-supported sites. The move puts an end to tech giants scraping free training data under the guise of web indexation, while expanding a model where publishers get paid when content actually creates value.
LangChain just released OpenWiki, an open-source CLI agent that automatically generates and maintains documentation for codebases. Instead of cramming an entire repo into CLAUDE.md, it builds a structured wiki and syncs changes via GitHub Actions, giving coding agents durable context without the bloat.
xAI just rolled out Voice Agent Builder, a no-code tool designed to create human-sounding voice agents using its Grok Voice tech. It is live now and priced at $0.05 per minute. If you are building phone bots or voice assistants, entry costs just dropped noticeably.
Fish Audio just made its flagship text-to-speech model, S2.1 Pro, completely free for developers. It packs support for 83 languages with zero usage caps, making high-end voice synthesis instantly accessible without worrying about token meter anxiety.
Cloudflare just introduced Monetization Gateway, a system letting site owners and developers charge autonomous AI agents for accessing web pages, datasets, APIs, or MCP tools. Built on the open x402 protocol and settled in stablecoins at the edge, it turns any HTTP request into an instant micropayment without requiring accounts or billing setups.
The Center for AI Safety and Scale Labs just updated their Remote Labor Index, and agents are getting better at digital work fast. Anthropic's Fable 5 leads the pack by delivering professional-grade freelance work 15.8% of the time, up from 2.5% less than a year ago. Why it matters: Progress is rapid, but the ceiling proves humans are still required to keep AI workers from going off the rails.
Meta is building a dedicated cloud business called Meta Compute to sell excess AI capacity and model hosting to outside clients, directly challenging AWS, Azure, and CoreWeave. The roadmap includes raw compute rentals alongside Bedrock-style access to its Muse Spark models. Meta stock jumped 9.3% following the report.
Square just launched integrations for ChatGPT and Claude that allow U.S. restaurants to take orders directly inside AI chat interfaces. Sellers pay standard processing rates with zero marketplace commissions, and tickets sync right to existing kitchen screens. Food delivery platforms taking 30% cuts should be very nervous.
Google just rolled out Gemini Spark for macOS, letting the AI organize local files, build spreadsheets from desktop invoices, and integrate with apps like Canva and Dropbox. It also introduces custom Model Context Protocol (MCP) support and upcoming remote execution from phones. The beta is currently limited to Google AI Ultra subscribers in the U.S.
OASIS just unveiled the OASIS 1, a smart ring that turns quiet whispers into text and packs a tiny trackpad for on-the-fly cursor edits. Powered by WisprFlow tech, it combines noise-isolating microphones with haptic feedback to fix transcription typos directly from your finger. It aims to make voice typing quiet and private enough to finally ditch the keyboard.
Runway just integrated Google's Gemini Omni Flash model into its AI video platform. Creators can now generate and edit videos using text prompts, images, or existing video clips. The feature is available directly within the Runway interface or by instructing the platform's built-in Agent.
Meta just introduced LUNA, a neural animation model that ditches traditional skeletal skinning and body-fitting bottlenecks. By turning simple 2D signals like sketches, keypoints, or video clips directly into 3D Gaussian deformations, it breathes motion into static character photos without the usual jitter and mesh tearing.
Netflix just revealed that its upcoming Willy Wonka reality show features an AI-generated voiceover of the late Gene Wilder. Created in partnership with ElevenLabs, the project was officially authorized by Wilder's estate. It highlights how Hollywood is increasingly turning to estate-approved AI clones to revive iconic voices for modern streaming audiences.
Microsoft just open-sourced SkillOpt, a framework that optimizes agent skill files like deep-learning parameters without altering underlying model weights. By using text-space forward passes, reflective edits, and strict validation gating, it boosted GPT-5.5 benchmark scores by 23.5 points. Why it matters: It replaces messy manual prompt hacking with a reproducible, controlled training loop.
Google just added Short Video Overviews to NotebookLM, turning your densest research docs into 60-second, vertical explainer videos. It builds on the tool's runaway hit audio podcasts, aiming straight at visual learners and social scrolling habits. Ultra and Pro subscribers get access now on mobile and web, with free users coming soon.
Google just launched two new generative media models for developers: Nano Banana 2 Lite and Gemini Omni Flash. Nano Banana 2 Lite delivers text-to-image outputs in 4 seconds at an ultra-low $0.034 per 1,000 images, while Omni Flash brings conversational video generation and editing at $0.10 per second of video. Why it matters: developers can now cheaply chain rapid image drafts straight into animated, multi-turn video edits.
Researchers just introduced MuSViT, the first foundation vision model trained specifically on sheet music. Pre-trained on 9.7 million pages from IMSLP by reconstructing masked scores, it grasps musical syntax, pitch, and staff layouts, outperforming general vision encoders across score transcription and symbol detection.
X just launched an official hosted Model Context Protocol (MCP) server, allowing AI assistants like Claude and Cursor to directly plug into its API using user account permissions. Developers can now let models search posts and analyze real-time trends without building custom middleware—though X confirmed it lacks write access to prevent bot spam.
OpenAI engineers have discovered optimizations that slash AI inference costs by more than half, according to The Information. When applied to ChatGPT for logged-out users, the technique dropped the required Nvidia GPUs down to just a couple hundred. The breakthrough could help OpenAI boost gross margins, which sat at 39% in Q1 2025, or lower API prices.
Researchers just open-sourced OVOW, a training-free system that turns a single video into simulation-ready 4D mesh scenes with separated objects. By chaining models like Qwen3-VL and SAM 3, it reconstructs motions and deformations directly for physics simulators. It is available for non-commercial research and recommends at least 40GB of VRAM.
Anthropic just redeployed Claude Fable 5 globally after US export controls were officially lifted on June 30. Access was abruptly frozen on June 12 when Amazon researchers bypassed safeguards to demonstrate software exploits. Armed with a new safety classifier that blocks over 99% of those bypasses, Anthropic is now partnering with Amazon, Google, and Microsoft to establish common industry standards for jailbreak severity.
Google just introduced TabFM, a foundation model designed specifically for tabular regression and classification tasks. By framing tabular prediction as an in-context learning problem, it bypasses traditional feature engineering and hyperparameter tuning to generate predictions in a single pass. It is trained on synthetic data and integrated directly into BigQuery.
Meituan just open-sourced LongCat-2.0, a massive 1.6-trillion-parameter MoE model that ditches Nvidia hardware entirely. The pretraining run spanned over 35 trillion tokens on dedicated AI ASIC superclusters without irrecoverable loss spikes, matching frontier models across coding and agent tasks. Why it matters: It proves cutting-edge AI can be trained at scale on alternative chips.
OpenAI just launched GeneBench-Pro, a challenging research-level benchmark designed to measure how AI agents navigate ambiguity and messy datasets in computational biology across 129 tasks. Using synthetic data with known causal structures for deterministic grading, GPT-5.6 Sol scored a 28.7% pass rate—up sharply from GPT-5's sub-5% run on the original benchmark. Ten sample questions are live on Hugging Face.
OpenAI engineers investigating mysterious crashes in Rockset used ChatGPT to build a pipeline analyzing a year of core dumps. The population-level analysis revealed two separate culprits: silent CPU arithmetic corruption on a single physical host, and an 18-year-old race condition in the widely used open-source GNU libunwind library.
The U.S. government just lifted export license restrictions on Anthropic's Mythos and Fable models, allowing the lab to restore public access starting July 1. The previous ban had effectively blocked public use of Anthropic's most capable systems. Facing mounting pressure from rising Asian competitors, Washington agreed on security terms with Anthropic to keep American AI ahead globally.
Anthropic just launched Claude Sonnet 5, making it the default model across Free and Pro tiers while keeping API pricing at $2/$10 per million tokens. It closes the performance gap with Opus 4.8 on coding and complex agentic workflows, natively planning and operating tools like terminals without hand-holding.
Anthropic's Claude Code team just dropped a practical guide on loop engineering, helping developers graduate from babysitting individual prompts to building autonomous agent loops. The framework breaks loops into four distinct levels—turn-based, goal-based using /goal, time-based using /loop and /schedule, and proactive routines—while offering tricks like SKILL.md self-verification to keep token costs in check. Basically, it shows how to turn open-ended coding agents into deterministic, self-healing workers.
Anthropic just released Claude Science in beta, an all-in-one desktop workbench for researchers on macOS and Linux. The app pulls together over 60 scientific connectors across genomics, chemistry, and HPC compute, complete with a built-in reviewer agent that catches calculation errors and bad citations. It is available now for Claude paid subscribers.
YouTube just rolled out AI-powered comment moderation tools to YouTube Studio on desktop for creators worldwide. Instead of relying on exact keywords, creators can now search by intent like "questions about my gear" or explore suggested topics like "Negative feedback." It also finds similar comments without auto-removing anything.
