News
Meta Releases Muse Spark Open, an 816B Open-Weight Model
Meta released Muse Spark Open on October 15, 2026. It's a huge 816B-parameter model for coding and agentic work. But it's not fully open source, and it needs serious hardware.
Category
Model releases
Verified coverage of new model launches, major upgrades and open-weight releases.
News
Nvidia in Talks to Buy Open-Source Model Startup Reflection AI, FT Reports
Reuters reported on October 10, citing the Financial Times: Nvidia is in talks with open-source AI startup Reflection AI, with options including additional investment or an outright acquisition. The company was valued at $25 billion in April; talks are early and could still fall apart.
News
xAI Adds Budget Tier to Grok Imagine Video 1.5: AI Video Generation from 2 Cents a Second
xAI added a Grok Imagine Video 1.5 Lite tier to its Imagine API docs — 2 cents per second at 480p, 3 cents at 720p. The 720p price is one-fifth of the full version and under one-third of Kling 3.0 Turbo.
News
OpenAI Quietly Lists gpt-rosalind-discovery on Pricing Page — Drug Discovery Model Surfaces Before Any Announcement
An never-announced model name appeared on OpenAI's API pricing page: gpt-rosalind-discovery. It sits next to gpt-rosalind-research under "life sciences models" at identical pricing. Nothing but a name and a price — no blog post, no docs, no system card. The use case can only be guessed from the name: most likely a new Rosalind-family member focused on drug discovery.
News
Claude Goes Beyond Chat: Anthropic Launches Motion Animations and Dashboards
Anthropic launched two beta features on October 8: Dashboards plugs data sources into Claude to generate live-updating dashboards from a sentence; Motion turns text and charts into editable animated explainers. Docs, Slides, and Design also graduated from beta, free tiers included.
News
Applications Hit 21x the Norm as Anthropic Pauses $1,000 API Credits for Startups
Anthropic's Claude startups program drew 21x the past five months' applications in 48 hours, with hundreds of thousands flooding in. The company paused two perks — a year of Claude Team and $1,000 in API credits. Already-claimed benefits are unaffected.
News
DeepSeek Processed 36.8 Trillion Tokens in a Week — More Than America's Big Four Combined
Per reports, VC firm Social Capital's analysis of OpenRouter data shows that in the last week of September, DeepSeek's models processed 36.8 trillion tokens on OpenRouter — more than OpenAI, Google, Anthropic, and xAI combined at 33.0 trillion.
News
OpenAI Rolls Out GPT-6 to Everyone: ChatGPT Answers Now Come With Interactive UI
OpenAI announced on October 7 that GPT-6 is officially entering ChatGPT: paid users get GPT-6 Sol first, free and Go users get GPT-6 Luna in batches from October 8. The simultaneously launched Intelligent UI puts charts, buttons, and interactive widgets directly in answers — output is no longer text-only.
News
Zhipu's GLM-5.3 Lands on AWS Bedrock with Usage-Based Revenue Sharing
AWS announced Zhipu's flagship open-weight model GLM-5.3 is fully available on Amazon Bedrock, with AWS reportedly sharing revenue with Zhipu based on usage. 753B-parameter MoE, 1M-token context, aimed at agentic coding and cybersecurity.
News
OpenAI Drops 722 Math Papers at Once: Unreleased Model's Solutions, Open on GitHub
On October 6, OpenAI published 722 math manuscripts solved by an unreleased frontier model to GitHub, open-sourced under Apache-2.0, some proofs with Lean formal verification. Its third math release in nine weeks drew a split reaction: applause from some, demands for evidence from others.
News
Anthropic Launches Claude Haiku 5.5: 75% Cheaper, Same Price as GPT-6 Luna
On October 7, Anthropic released Claude Haiku 5.5 — the third 5.5-family model in a month, aimed at high-frequency, low-cost tasks. Pricing is 75% cheaper than Haiku 4.5, matching OpenAI's GPT-6 Luna directly.
News
Bloomberg: US-China Top AI Model Benchmark Gap Narrows to 3%, a Record Low
Bloomberg Intelligence's October 5 report says that after DeepSeek released V4.1 Flash, the benchmark gap between the strongest US and Chinese AI models is just 3%, down from 15% at the start of the year. Scores are close, but US lineup depth and China's price-war reality haven't changed.
News
Google Launches Nano Banana 2.1: Image Model Goes 4K, Old Version Retires End of Month
On October 6, Google made image generation model Nano Banana 2.1 generally available, with up to 4K output and 14-reference-image blending. The previous Gemini 3.1 Flash Image shuts down October 29 — developers need to migrate soon.
News
EmbeddingGemma 2 Released: 740M Parameters, One Model to Search Text, Images, Audio, Video
Google DeepMind released EmbeddingGemma 2 on October 6 — a 740M-parameter open embedding model mapping text, code, images, audio, and video into one vector space. It runs offline on phones and computers; searching local files needs no internet.
News
Mistral Large 4 Released: European AI Tries to Catch Up with Chinese Open Models
Mistral AI released its flagship Large 4 on October 6, claiming it has "narrowed the gap with the strongest models." The company says it matches GLM 5.3 on long coding tasks, trained on just 4,000 chips, with open weights coming October 27.
News
Reflection AI's First Open-Weight Model Is Coming: Taking On DeepSeek and Qwen
Axios reported exclusively on October 4 that the Nvidia-backed AI startup Reflection AI will release its first open-weight foundation model this month, targeting DeepSeek and Qwen. Founded by two former DeepMind researchers, the company raised $2 billion last year at an $8 billion valuation — but the model name, parameter count, license and benchmarks are still undisclosed.
News
Aleph Alpha Open-Sources Kolibri: 78B Parameters, Only 3.46B Active per Token
German AI company Aleph Alpha released the open-weight model Kolibri on October 3 — 78.1B total parameters with only 3.46B active per token, Apache 2.0 license, English-German bilingual, context up to 1M tokens, built for sovereign AI local deployment.
News
PewDiePie Releases Local AI Model Ajax: OpenAI Banned His Account Twice Over "Distillation"
YouTuber PewDiePie released a homemade local AI model, Ajax (fine-tuned from Qwen3.5-9B), saying OpenAI banned his account twice during development — because he used OpenAI model outputs as training data for Ajax, i.e., distillation.
News
FLUX 3 Image Launches: Native 4K Output, Multi-Image Reference Editing
Black Forest Labs released the image model FLUX 3 Image on October 1. A single request accepts up to 10 reference images, supports natural-language local editing, with native output up to 4K. API is half-price at launch; the open-weight version is expected within weeks.
News
Claude Sonnet 5.5 Arrives: 30% Faster, Same Price, and Takes Aim at GPT-6 Sol
Anthropic has released the second model in the Claude 5.5 series, Sonnet 5.5. It responds over 30% faster than its predecessor, keeps API prices unchanged, and jumped from 10.3% to 70.6% on the Terminal-Bench 4.0 coding benchmark. Anthropic claims it 'clearly outperforms' OpenAI's GPT-6 Sol on long-horizon knowledge work.