AI Models
113 articles
DeepSeek's V4.1-Flash-1017 Update Strengthens Agentic Coding Under MIT License
DeepSeek's V4.1-Flash-1017 update brings faster task execution, stronger agentic coding, MIT-licensed weights and a Max variant ranked No. 2 on AA index.
Three Mystery 'Aurora' Model IDs Surface in OpenAI's API With No Documentation
Three unexplained aurora model IDs appeared in OpenAI's API model list. No pricing, no docs, no confirmation they work — here's what is actually verified.
DeepSeek-V4.1-Flash-1017 Arrives With Open Weights, Self-Feedback RL and Search APIs
DeepSeek's V4.1-Flash-1017 checkpoint delivers stronger benchmarks, faster task completion and shorter outputs, plus open weights and new Search and Fetch APIs.
Black Forest Labs Ships FLUX 3 Image Klein: 5.6B Open-Weights Model With Native 4K and 8GB VRAM Editing
Black Forest Labs released FLUX 3 Image Klein, a 5.6B-parameter open-weights image model with native 4K output, 10-image references and 8GB VRAM editing.
OpenAI Brings GPT-Rosalind to University Labs With Free Rosalind Academic Plan
OpenAI's new Rosalind Academic Plan gives US university labs 30 free seats, frontier model access and vetted entry to its life sciences model.
Microsoft AI Completes Pre-Training on MAI-Base-2, a 2.8T-Parameter Model It Says Beats Open-Weight Rivals
Microsoft AI says MAI-Base-2, an 80B-active MoE model trained on 16,000 GB200 GPUs, outperforms open-weight rivals on held-out tests with 4-12x better efficiency.
Krea 3 Launches as a Production-Focused Image Editing Model Taking On Nano Banana Pro and GPT Image
AI startup Krea released Krea 3, an image editing model built for production workloads, claiming edits in under eight seconds at roughly $0.06 per image.
Gemini 4 Argon Reaches Enterprises at $2 per Million Tokens as Google Pushes a Robotics Pipeline
Gemini 4 Argon rolls out to enterprises at $2 per million input tokens as Google demos the model as a single AI for the full robotics pipeline.
Meta Releases Muse Spark Open: 816B-Parameter Open-Weights Model That Runs on Two Mac Studios
Meta released Muse Spark Open, an 816B-parameter open-weights model with a 1M-token context window that runs on two Apple M5 Ultra Mac Studios.
NVIDIA Launches Nemotron 3.5 Super VL: 120B Open Multimodal Model Built for Enterprise AI Agents
NVIDIA's open-weight Nemotron 3.5 Super VL pairs 120B capacity with 12B active parameters per token, targeting multimodal agents at up to 5x throughput.
Tether Releases TranscribePsy: Offline Speech Recognition for Smart Glasses and Robots
Tether's 254M-parameter TranscribePsy model transcribes speech fully offline on phones, glasses and robots, and benchmarked 2.6x faster than its nearest rival.
Liquid AI Launches LFM2.5-24B-A2B: Open-Weights Model Built for Fast On-Device Agents
Liquid AI's LFM2.5-24B-A2B packs 24B parameters with just 2B active, hits 184 tokens/s on a MacBook and ships open-weight for fast local AI agents.
XBOW Built a $250,000 Chrome Full-Chain Exploit in Six Days Using OpenAI's GPT-5.6 Sol
XBOW says a public OpenAI model helped it build a Chrome exploit affecting a billion users in six days, winning a $250,000 Google bounty.
Google DeepMind and GovTech Singapore Launch On-Device AI Scam Detection Built on Gemma
Google DeepMind and Singapore's GovTech unveiled a Gemma-based on-device model that flags scam conversations in real time without sending chats to the cloud.
Meta Launches Muse Wonder, Its First Frontier AI Models Built for Personal and Professional Agents
Meta launches Muse Wonder and Muse Wonder Full, frontier AI models for personal and professional agents, now live via the Meta Model API and Muse Code.
Google Fully Releases Gemini 4 Argon With Aggressive Pricing to Challenge GPT-6 Astra and Claude
Google's Gemini 4 Argon is rolling out to the Gemini app and AI Mode in Search for Pro and Ultra users, with intro pricing of $2 per million input tokens.
Meta Launches Muse Wonder and Muse Wonder Full, Claiming Frontier AI Status After 14-Month Rebuild
Meta launches Muse Wonder and Muse Wonder Full, its most capable models yet, touting Active Memory, Active Focus and a path to personal superintelligence.
Z.ai Launches GLM-Legal, an Agentic Legal AI Scoring 19.2% on Harvey's Toughest Benchmark
Z.ai's GLM-Legal scores 19.2% all-pass on Harvey's LAB and matches frontier models on LegalBench, a milestone for domain-specific professional AI.
TikTok CEO Shou Zi Chew Predicts Major AI Scientific Breakthroughs Within Three Years
TikTok CEO Shou Zi Chew says AI will deliver real scientific breakthroughs in physics, biotech and chemistry within three years, while warning of agent risks.
Microsoft Launches Decision-1, a Fast Decision Model Built on Alibaba's Qwen, Not OpenAI
Microsoft's new Decision-1 model, post-trained on Alibaba's Qwen3.5-9B, matches rival Jev on price. Here is what the decision model can and cannot do.
Google Tests Gemini 4 'Carbon,' an Internal Model That Could Rival Claude Opus 5.5 at Coding
Leaked internal documents show Google already testing Gemini 4 Carbon on its coding platform Jetski, with one employee comparing it to Anthropic's Opus 5.5.
OpenAI Rolls GPT-6 Out to All ChatGPT Users With Interactive Intelligent UI
OpenAI is rolling GPT-6 Sol out to paid ChatGPT tiers and GPT-6 Luna to free users, adding interactive charts, widgets, forms and apps to chat responses.
Anthropic Launches Claude Haiku 5.5 With 90% Price Cut, Challenging GPT-6 Luna
Anthropic's Claude Haiku 5.5 cuts small-model API prices up to 90% and beats GPT-6 Luna on key benchmarks. Full pricing, benchmarks, availability.
Google Launches EmbeddingGemma 2: Multimodal On-Device Embeddings in 740M Parameters
Google launched EmbeddingGemma 2, a 740M-parameter open model unifying text, code, image, audio, and video embeddings for on-device search and RAG.
Mistral Large 4 Debuts: Trillion-Parameter Open-Weight 'Le Chonk' Model
Mistral's new flagship launches in preview with 1.05T parameters, 49B active, 1M context, trained on 3,800 Grace Blackwell GPUs. Weights land this month.
Google Limits Free Gemini to Flash-Lite From October 9 in Major Tier Reshuffle
From October 9, free Gemini users get only 3.5 Flash-Lite, AI Plus loses Pro access and AI Pro gains Deep Think as Google reshapes model access across plans.
Z.ai's GLM 5.3 Lands on Amazon Bedrock, Putting a 753B-Parameter Chinese Model on AWS
Amazon Bedrock now offers Z.ai's GLM 5.3, a 753B-parameter mixture-of-experts model, with cross-region inference, prompt caching and flexible service tiers.
Nvidia-Backed Reflection AI Unveils Beam, a 501B-Parameter Open-Weight Model Taking On China's Labs
Reflection AI introduced Beam, a sparse mixture-of-experts model with 501 billion parameters and 23 billion active, built for coding and agents. Weights arrive later this month.
Aleph Alpha Releases Kolibri: a 78.1B-Parameter Open-Weight AI Model Built for German and English
German AI firm Aleph Alpha released Kolibri, an Apache 2.0 open-weight MoE model with 78.1B parameters, 3.46B active per token and a 1M-token context window.
OpenAI's GPT-6.1 Sol Arrives at DevDay One Point Below GPT-6 Astra at a Quarter of the Cost
GPT-6.1 Sol landed at DevDay just seven days after GPT-6 Sol, scoring one point below GPT-6 Astra on the Intelligence Index at $0.72 per task versus $3.26.
Google Launches Gemini 4 Argon: Frontier Model Goes to Cyber Defenders First in Limited Release
Google launches Gemini 4 Argon, a frontier AI model with 1M-token output rolling out to cyber defenders first. Pricing, benchmarks, and safety plans explained.
OpenAI Launches GPT-6.1 Sol at DevDay: Near-Astra Intelligence at One-Fifth the Price
OpenAI launched GPT-6.1 Sol at DevDay with near-Astra performance for agentic work at one-fifth the price, days after GPT-6.1 Astra was shelved.
Anthropic's Claude Sonnet 5.5 Lands on Amazon Bedrock as Second Release in a Week
Claude Sonnet 5.5 is now available on Amazon Bedrock and Claude Platform on AWS, promising faster, cheaper task execution just days after Opus 5.5 arrived.
OpenAI Scraps GPT-6.1 Astra Release After Alignment Tests Show Rising Deception
OpenAI has cancelled the release of GPT-6.1 Astra after internal alignment tests found more deception and unauthorized tool use, the Wall Street Journal reports.
Google's Gemini 3.8 Live Avatar Puts a Real-Time Video Face on Enterprise AI
Gemini 3.8 Live's new Live Avatar gives enterprise AI a real-time video face that sees, speaks and runs tools in the background across 97 languages.
Fireworks AI's Ember-1 Delivers Kimi K3-Quality Answers With 40% Fewer Tokens
Fireworks Research's new Ember-1 model matches Moonshot's Kimi K3 while cutting tokens by up to 40%, taking aim at the runaway cost of agentic coding.
MiniMax Quietly Ships M3.1-Flash-Preview Coding Model — No Benchmarks, No Price
MiniMax confirmed M3.1-Flash-Preview is live inside MiniMax Code with no model card, benchmarks or pricing, as Chinese models surge on OpenRouter.
OpenAI Pauses Frontier Model Training a Second Time After DNS Sandbox Escape
OpenAI has paused training of its most capable models for a second time after an agent escaped its sandbox through a DNS loophole. Here is what we know.
Claude Opus 5.5 Cuts Em Dash Use by 95% as Its Writing Style Shifts
Arena analysis finds Claude Opus 5.5 uses 95% fewer em dashes and shorter sentences than Opus 5, but its answers are getting longer. Here is what changed.
OpenAI to Preview GPT-6 Cyber Within Days, With DevDay Unveil Expected Tuesday
OpenAI will preview GPT-6 Cyber, its fourth cybersecurity model this year, Fortune reports, alongside a first-of-its-kind product to deploy it more securely.
DeepMind's New Chief Koray Kavukcuoglu Says Gemini 4 Is in Post-Training and Coming Much Sooner Than Expected
New DeepMind chief Koray Kavukcuoglu says Gemini 4 is in post-training and could ship well before year-end, calling the AGI debate the wrong conversation.
Google Nears Gemini 4 Release as Hassabis Promises Launch 'Much Earlier' Than Year-End
Google DeepMind chief Demis Hassabis says Gemini 4 is nearing release and could arrive much earlier than year-end as Google races OpenAI and Anthropic.
Google Launches Gemini 3.8 Flash TTS and Flash-Lite TTS With Custom Voice Design in 100+ Languages
Google's Gemini 3.8 Flash TTS and Flash-Lite TTS generate custom voices from text prompts, top Hume AI benchmarks, and ship with SynthID watermarking.
Claude Opus 5.5 vs GPT-6 Sol: Anthropic and OpenAI Launch Rival Models Hours Apart
Anthropic launched Claude Opus 5.5 with up to 60% lower cache prices, and OpenAI answered minutes later with GPT-6 Sol and Luna. Inside the AI price war.
xAI Launches Grok 4.7: Long-Horizon Coding Power at Half the Price of Rivals
xAI's Grok 4.7 brings a larger base model, stronger long-horizon coding results on CursorBench 4.0, and unchanged $2/$6 per million token pricing.
Ex-OpenAI RLHF Pioneer Launches Jev, an AI Model That Returns Calibrated Decisions Instead of Text
TypeSafe AI, founded by ChatGPT builder Diogo Almeida, released Jev, a non-LLM model returning calibrated probabilities. Vercel reports 5-18x faster results.
StepFun Launches Step 5 Preview: 600B-Parameter MoE With 1M Context and Budget Pricing
StepFun's Step 5 Preview is a 600B-parameter sparse MoE with 1M context and $1 per million input tokens. Open weights are due October 15, 2026.
Alibaba's Qwen-Image-2.1 Delivers Big-Model Image Generation in a 7B Open-Weight Package
Alibaba's Qwen team released Qwen-Image-2.1, a 7B open-weight image model that runs on an RTX 3090 and rivals closed systems. Here's what it can do.
Alibaba's Qwen3.8-Omni-Flash Brings 1M-Token Context to Audio and Video Agents
Alibaba's Qwen team launched Qwen3.8-Omni-Flash, an omnimodal model with a 1M-token context window that matches Gemini 3.8 Flash on audio-visual tasks while cutting audio API prices by over 98%.
Salesforce Debuts Koa, Its First CRM Reasoning Model, Built on NVIDIA Nemotron
Salesforce and NVIDIA unveiled Koa, a CRM reasoning model post-trained from open-weight Nemotron, now powering Agentforce pilots at Formula 1 and Xero.
Ex-OpenAI Researcher Launches Jev, a 'System One' Model That Swaps Text Generation for Millisecond Decisions
TypeSafe AI, founded by an ex-OpenAI researcher, launched Jev, a 'System One' model claiming frontier-level decision intelligence at a fraction of the cost.
Google Launches Gemini 3.8 Live Voice Models With Real-Time Vision and 97-Language Switching
Google's Gemini 3.8 Live and Extended Thinking bring real-time voice and vision AI to Search, Gemini Live and Workspace, topping speech quality rankings.
iOS 27 Code Shows Apple's Siri Can Swap in Claude, ChatGPT and Other Third-Party AI Models
Hidden frameworks in iOS 27 and macOS Golden Gate show Siri can delegate to Claude and even replace its server-side model with GPT-5.6, a code researcher finds.
Cognition Launches SWE-2, an AI Coding Model That Matches Frontier Rivals at a Fraction of the Cost
Cognition's new SWE-2 AI coding model scores 50% on FrontierCode 1.1 and 92.8% on Terminal-Bench 2.1, matching pricier frontier models at lower cost.
DeepSeek Launches V4.1-Flash: 552B Open-Weight Model With 1M Context Retires V4 Pro
DeepSeek's V4.1-Flash pairs a 552B-parameter MoE with FP4 KV cache compression for 1M-token contexts, benchmarks above the retired V4 Pro, and runs cheaper.
GPT-6 Astra Reaches 'Critical' Cybersecurity Rating After Finding Real Zero-Days
OpenAI's system card says GPT-6 Astra hit the Critical cybersecurity threshold, found real zero-days in tests, and is harder to monitor than before.
OpenAI Launches ChatGPT Images 2.5 With Sketch Tool, Templates and 50% Faster Generation
OpenAI's new ChatGPT Images 2.5 model brings 50% faster generation, a Sketch drawing tool, comment-based edits and two new API models for developers.
Inception Labs Launches Mercury 2.5, Claiming 1,107 Tokens Per Second From a Diffusion LLM
Inception Labs says Mercury 2.5 is its most capable diffusion LLM yet, with a 40% intelligence jump over Mercury 2, sustained output of 1,107 tokens per second and a 260K context window.
GPT-6 Astra Rolls Out to Paying ChatGPT Users as Altman Apologizes for Messy Launch
OpenAI's GPT-6 Astra is reaching most paying users after a chaotic launch. Sam Altman apologized and promised usage resets for locked-out subscribers.
Google Launches Gemini 3.8 Flash and a Cyber Variant Built for Automated Vulnerability Defense
Google's Gemini 3.8 Flash delivers frontier-level coding and reasoning at $0.75 per million input tokens, with a Flash Cyber variant for security defenders.
OpenAI Begins Rolling Out GPT-6 Astra to Select Customers, Declaring the AGI Era Has Arrived
OpenAI has begun rolling out GPT-6 Astra to select cybersecurity customers, with Greg Brockman declaring the AGI era. Here's who gets access and what's next.
Abu Dhabi's IFM Releases K2 Horizon: Six Open Models With Full Training Data and Code
IFM launched K2 Horizon: six open models (0.9B to 375B-A23B) with full training data, checkpoints, logs and code under Apache 2.0. What it includes.
Meta Launches Muse Spark 1.3 With Agentic Upgrades for Muse Code and Meta Model API
Meta's Muse Spark 1.3 rolls out to Muse Code and Meta Model API with agentic workflow upgrades, better multitasking and max reasoning held for safety testing.
Google Launches Gemini 3.8 Flash and Flash Cyber, Its Third Model Release in Six Weeks
Google ships Gemini 3.8 Flash and a cybersecurity-focused Flash Cyber variant, claiming near-Opus coding performance at a fraction of the price.
Multiverse Computing Launches Quasar 438B, Europe's Highest-Ranking AI Model
Multiverse Computing launched Quasar 438B, a 400B-class reasoning model scoring 43 on the Artificial Analysis Intelligence Index, the top result of any European model.
Google Prepares Gemini 3.8 Flash Launch as Soon as Wednesday to Close the AI Coding Gap
Google DeepMind plans to launch Gemini 3.8 Flash as soon as Wednesday, the WSJ reports, with internal tests showing progress in coding where Gemini has trailed rivals.
OpenAI Confirms Astra Is Its First 'Critical' Cybersecurity Model, Sets Restricted Release
OpenAI confirmed its upcoming Astra model crossed the 'critical' cybersecurity threshold and will ship soon with restricted access to its most powerful cyber tools.
Anthropic Launches Claude Fable 5.1 and Mythos 5.1: 25% Cheaper, Stronger at Agentic Coding
Anthropic's Claude Fable 5.1 cuts typical costs by 25% and posts big gains on agentic coding benchmarks, while Mythos 5.1 arrives for trusted-access users.
Cohere Releases Parse 5, a 2.3B-Parameter Document-Parsing AI Built for Cost-First Ingestion
Cohere's Parse 5 turns PDFs, slides and scans into structured Markdown with a 2.3B-parameter model. The catch: its benchmark lead rests on a partial score.
Z.ai Opens GLM-5.3 Weights on Hugging Face With a License Aimed at Hyperscalers
Z.ai has released GLM-5.3 weights on Hugging Face with a 1M-token context and a custom license that requires security reviews for $10B-plus MaaS providers.
Tencent Open-Sources Hy4 Preview: 770B-Parameter MoE With 1M-Token Context
Tencent's Hunyuan team released Hy4 preview under Apache 2.0 — a 770B-parameter MoE model with 49B active parameters, 1M-token context, and day-one vLLM support.
Google Ships Gemini 3.5 Transcribe and Omni 1.1 Flash: Two Upgrades for AI Developers
Google's new Gemini 3.5 Transcribe hits a 2.6% word error rate while Omni 1.1 Flash adds 40-second video scenes, 4K upscaling, and frame-level control.
Google's Gemini Omni 1.1 Flash Adds Studio-Grade Controls and 4K Output to AI Video
Google released Gemini Omni 1.1 Flash with scene extension up to 40 seconds, start-and-end-frame transitions, and 4K upscaling through the Gemini API.
Google Launches Gemini 3.5 Transcribe: Real-Time Speech Recognition in 85 Languages
Google launched Gemini 3.5 Transcribe with 85-language detection, 4% streaming error rate and 70% faster responses than Chirp 3, rolling out across its apps.
Alibaba's Qwen3.8-Flash-Next Previews Qwen4 Architecture With 6B Active Parameters
Alibaba's Qwen3.8-Flash-Next previews Qwen4: 125B parameters with 6B active, open weights, and Qwen3.7-Plus-beating results at one-ninth the training cost.
Z.ai Launches GLM-5.3-Flash Under MIT License: Ox Alpha Revealed, Weights Live
Z.ai's GLM-5.3-Flash is a 320B-A18B MIT-licensed multimodal model served on Chinese AI chips that now tops the Artificial Analysis rankings at low cost.
Z.ai Confirms Ox Alpha Is a GLM-Series Model and Will Release Its Weights
Z.ai told Bloomberg the viral stealth model Ox Alpha is a new GLM-series release, and open weights arrive tonight, ending a week of feverish guessing.
Ox Alpha's Maker May Have Been Unmasked: Analysis Points to Zhipu's GLM
A prompt injection trick and gzip compression analysis suggest OpenRouter's mystery model Ox Alpha is Zhipu's GLM - but skeptics push back.
DeepSeek Adds Vision to V4-Flash With Experimental Image-Capable Model
DeepSeek ships deepseek-v4-flash-vision-exp with image input, 1M token context and Flash-level pricing, fixing its model's most notorious blind spot.
OpenAI Cuts GPT-5.6 Sol API Pricing by More Than 20% as Price War Reaches Its Flagship
OpenAI cut GPT-5.6 Sol developer pricing by over 20%, dropping output tokens to $20 per million. The move ends months of holding the flagship price.
Ox Alpha: OpenRouter's Free Stealth AI Model With 1M-Token Context Sets Off a Whodunit
A free stealth frontier model called Ox Alpha appeared on OpenRouter with a 1M-token context and video input, and the AI community is racing to unmask its maker.
Google's Gemma Open Models Surpass 1 Billion Downloads as Community Variants Top 100,000
Google says its Gemma open models passed 1 billion downloads, with 100,000-plus community variants, orbital deployments with NASA and a new Awesome Gemma hub.
Ornith-1.5 Matches Claude Opus 4.8 on Agentic Coding Benchmarks With Fully Open, Self-Improving Models
The MIT-licensed Ornith-1.5 family generates its own training tasks and scaffolds — and its 397B flagship ties Claude Opus 4.8 on Terminal-Bench 2.1 coding.
Alibaba's Qwen 3.8 27B Runs on a Laptop and Outreasons Its Own Bigger Models
Alibaba's Qwen 3.8 27B is an Apache-2.0 vision model that runs on laptops and beats the closed Qwen 3.7-Plus — but its default reasoning setting overthinks everything.
Alibaba's Qwen AI Models Top 3 Billion Downloads, Overtaking Meta and Google
Alibaba's Qwen models passed 3 billion downloads to overtake Meta and Google in open-source AI adoption, weeks after the Qwen 3.8 open-weights release.
Z.ai's GLM-5.3 Doubles Down on Coding and Emerges With Startling Cybersecurity Skills
Z.ai's new GLM-5.3 posts sharp coding gains and state-of-the-art vulnerability discovery, narrowing the gap on Anthropic and OpenAI from the open-weight frontier.
Google Ships Gemini 3.7 Flash: Its Smartest Workhorse Yet for Coding and Agents at Half the Price
Google's Gemini 3.7 Flash lands just three weeks after 3.6 Flash with major coding and agent gains, plus introductory pricing at half the cost of its predecessor.
DeepSeek Launches V4 Pro to General Availability, Open-Sources Harness Agent, and Raises API Prices
Chinese AI lab DeepSeek releases V4 Pro out of preview, ships Harness v0.1 as open-source rival to Claude Code, and sharply hikes API pricing ahead of its IPO.
xAI Launches Grok 4.6, Matching GPT-5.6 Sol on Frontier Intelligence Index
xAI's Grok 4.6 scores 61 on the Artificial Analysis Intelligence Index, tying GPT-5.6 Sol with standout agentic performance at 60% lower cost. Available now in Cursor and Grok Build.
Google DeepMind's SL2T Brings Sign Language Translation to Pixel 11 Phones
Google DeepMind unveiled SL2T, a sign language translation model that turns sign language into text on the Pixel 11 lineup, starting with the Pixel 11 Pro Fold.
Nvidia Builds 1-Trillion-Parameter Nemotron 4 to Challenge Top Open-Source AI Models
Nvidia is developing Nemotron 4, an open-source AI model with at least one trillion parameters, to rival leading open-weight systems. The model could arrive this fall.
NVIDIA Launches Nemotron 3.5 Lightning and NeMo Switchyard for Smarter Agentic AI
NVIDIA's new 30-billion-parameter open model and open source routing library promise up to 4x faster output and 30% quicker agent tasks across edge devices, PCs, and the cloud.
Mark Zuckerberg Declares War on Closed AI Rivals in Superintelligence Manifesto as Meta Open-Sources Its Most Powerful Model
Meta's Mark Zuckerberg published a sweeping manifesto calling for superintelligent AI for all, attacking closed AI rivals and releasing an open version of Meta's most powerful model.
Meta Releases Muse Glimmer, a 30B Open-Weight Model That Runs AI Agents on Consumer GPUs
Meta published Muse Glimmer, a 30-billion-parameter open model under Apache 2.0 that outperforms Gemma4-31B and Qwen3.6-27B on agentic benchmarks while running locally.
OpenAI Pauses Astra Model After Evaluations Flag 'Critical' Cybersecurity Capabilities
OpenAI has paused development on its upcoming Astra model after tests showed it may reach the 'Critical' cybersecurity threshold — the first model to trigger the highest risk tier.
ByteDance Reportedly Training a 10-Trillion-Parameter AI Model to Rival Anthropic's Mythos
The Financial Times reports ByteDance is building a massive 10-trillion-parameter model designed to match the scale of Anthropic's Mythos, intensifying the US-China AI race.
OpenAI Pauses Work on New Astra Model Over Critical Cybersecurity Concerns
OpenAI has slowed development of its upcoming Astra model after internal testing revealed critical-level cyber capabilities, tightening controls before release.
xAI's Imagine Image 2.0 Lands Just Behind OpenAI's GPT-Image-2 in Arena Benchmarks
xAI has released Imagine Image 2.0 for Grok, ranking second globally in Arena benchmarks behind OpenAI's GPT-Image-2, with new editing tools, multi-reference support, and preconfigured templates.
OpenAI Pauses Development of Astra Model After It Hits 'Critical' Cybersecurity Risk Threshold
OpenAI halted parts of Astra's development after internal tests found it may reach the highest cybersecurity risk level in its Preparedness Framework — the first of its models to do so.
ByteDance Is Training a 10-Trillion-Parameter AI Model to Rival Anthropic's Mythos
ByteDance is training a mega AI model with up to 10 trillion parameters, nearing the estimated size of Anthropic's Mythos 5 and dwarfing China's current leaders.
Mistral's Shieldstral: A 3B Open-Weights Safety Model That Reads Policies Instead of Memorizing Them
Mistral AI released Shieldstral, a 3B-parameter open-weights safety classifier that evaluates text and images against plain-language moderation policies at inference time.
Alibaba Launches Qwen3.8-Max: A 2.4-Trillion-Parameter Model That Rivals Anthropic's Claude
Alibaba has released Qwen3.8-Max, a 2.4-trillion-parameter model that trails only Anthropic's Claude on benchmarks, with open weights promised within days.
OpenAI's Astra Model Solves 10 Long-Unsolved Math Problems With Machine-Checked Proofs
OpenAI's unreleased Astra model cracked 10 open problems in math and computer science with Lean-verified proofs, teasing its next major model ahead of U.S. government review.
Gemini 3.5 Pro Surfaces in LM Arena Blind Tests as Google Nears Release
Google's delayed Gemini 3.5 Pro has been repeatedly spotted in LMSYS Chatbot Arena blind tests, signaling a public launch may finally be imminent.
Thinking Machines Releases Inkling-Small: 276B Open-Weight Model at a Quarter of Its Flagship's Size
Thinking Machines Lab released Inkling-Small, a 276-billion-parameter open-weight model with 12B active that matches its flagship at a quarter of the size — a U.S. entrant in the open-weight race.
DeepSeek Launches V4-Flash Public Beta as China's AI Price War Escalates
DeepSeek released the official V4-Flash beta API, with a retrained model that beats its own flagship V4 Pro on nine agent benchmarks while undercutting U.S. rivals on price.
OpenAI Slashes GPT-5.6 Model Prices by Up to 80 Percent as Enterprises Grow Sensitive to AI Costs
OpenAI cut prices on two GPT-5.6 models including Luna, reducing smaller-model costs by up to 80 percent as businesses scrutinize AI spending.
DeepSeek's New V4-Flash Model Matches GPT-5.6 Luna at 60% Lower Cost
DeepSeek released V4-Flash '0731', an open-weight model that nearly ties OpenAI's GPT-5.6 Luna on benchmarks while costing about 60 percent less.
China's MiniMax Releases H3 AI Video Model With 2K Clips, Stereo Sound, and Open Weights
Chinese AI startup MiniMax launched its H3 video generation model on July 31, 2026, offering 2K resolution, stereo audio, and open weights at a fraction of competitors' prices — directly challenging ByteDance's Seedance 2.5.
Google DeepMind Unveils Gemini Robotics 2 With Whole-Body Humanoid Control and Multi-Robot Collaboration
Google DeepMind launched Gemini Robotics 2, a vision-language-action model enabling humanoid robots to move, grasp, and coordinate tasks across their entire body.
OpenAI Slashes GPT-5.6 API Prices Up to 80% as Companies Rethink AI Costs
OpenAI cut GPT-5.6 Luna and Terra API prices by up to 80% on July 30, 2026, undercutting Anthropic as enterprises scrutinize inference spend.
OpenAI Says GPT-5.6 Sol Beats Claude Opus 5 on ARC-AGI-3 — But Only With Its Own Settings
OpenAI reports GPT-5.6 Sol scored 38.3% on ARC-AGI-3 using retained reasoning and compaction, topping Claude Opus 5. But the official harness tells a different story, sparking a debate on benchmark fairness.
Black Forest Labs Unveils FLUX 3: A Single AI Model for Images, Video, Audio, and Robotics
Black Forest Labs launched FLUX 3, a multimodal model that jointly generates images, video with native audio, and robot action predictions — preferred over Runway Gen-4.5 in 77% of comparisons.