Frontier Model Releases
Anthropic's new flagships. Anthropic released Claude Fable 5.2 and Mythos 5.2 with improvements in coding, reasoning, vision, and faster outputs; Fable price drops to $8 per million input tokens and supports zero data retention. The models top Artificial Analysis Intelligence Index v5, and some analysts call the capability 'Jagged AGI' rather than full AGI.
- Claude Fable 5.2 and Mythos 5.2 cement Anthropic's AI lead just before its IPO(opens in a new tab)
- Anthropic releases new Mythos and Fable models amid RSI concerns(opens in a new tab)
- Anthropic launches new Claude Fable and Mythos 5.2 models(opens in a new tab)
- Fable 5.2 and Jagged AGI(opens in a new tab)
Meta's small agent model. Meta released Muse Ember, its lightest and most cost-efficient model, designed for repetitive agent tasks like tool calls, coding subtasks, routing, and extraction at $0.10 per million input tokens. It handles text, images, audio, video, and up to 1M tokens, and Meta claims better cost per task than Claude Haiku 5.5 or GPT-6 Luna.
Gemini for robotics. Google released Gemini 4 Argon, a general-purpose physical-reasoning model for robot control, Real2Sim scene construction, multimodal labeling, and policy improvement, scoring 87% on ERQA without helper tools.
Seedream 5.5 image editing. ByteDance's Seedream 5.5 image generation model promises pixel-stable editing across up to 100 rounds without color drift, native RGBA transparency, and improved dense text rendering; it's rolling out via Jimeng, Doubao, and Volcengine Ark.
Local & Open Models
Audio embeddings in GGML. clap.cpp implements CLAP audio embedding inference in C++/GGML with CLI, HTTP server, and Python/Rust/Node bindings, enabling text-based search of audio files locally.
400M model for NPCs. A 400M encoder model fine-tuned for game NPC judgment calls answers yes/no and pick-one questions in about 0.1s on a 3060 Ti, with 42 adapters for common scenarios like threat, apology, and memory matching, available open source.
2B model plays Doom. A LoRA fine-tuned 2B decision model on MLX can play Doom's Defend the Center arena, averaging 23.2 kills over 20 episodes after training in 43 minutes on a Mac using game frames as input.
Comet 60B GGUF. Multiverse Computing's Comet 60B (A12B) compressed Nemotron 3.5 Super was converted to GGUF with MTP head support, runs on stock llama.cpp, text-only (no vision tower).
Tools & Frameworks
Mixed-vendor GPU inference. A fork of Project MAYA enables LLM inference across CUDA and ROCm GPUs simultaneously, with the CUDA backend handling attention/dense layers and ROCm handling complementary experts, achieving up to 48% faster prefill/decode on mixed setups; PR welcomed by maintainer.
Prompt optimization library. LangChain released Promptim, an experimental library that automates prompt improvement using a dataset, custom evaluators, and optional human feedback, based on evaluation-driven development.
llm-anthropic adds Fable. Simon Willison updated llm-anthropic to support Anthropic's new Fable 5.2 and Mythos 5.2 models; he notes a bug where max reasoning level fails to return images after spending 128k thinking tokens.
Claude Code secrets leak. Security researchers found Claude Code's settings.local.json files, which store approved shell commands including sensitive credentials, sometimes get committed to public GitHub repos, exposing thousands of secrets.
Baidu's automated research agent. Baidu's ERNIE Evolve automated research system improved inference code for ERNIE-5.5-Flash, beating public benchmarks on 3 of 4 metrics and achieving real production gains on Kunlunxin P800, though system not released.
Industry & Business
Anthropic IPO plans. Anthropic reportedly plans to set a valuation range this week and begin trading next week, with some investors pushing for a valuation below the $1.5T target, down from $2T floated months ago.
Nvidia spy scandal. Nvidia CEO Jensen Huang admitted an affair with Alice Huang, a Megaspeed exec suspected of spying for China; Nvidia opened an investigation and Huang pulled out of public events. Reports indicate Huang hired crisis PR firm Heller Inc. a year ago, a short seller flagged the relationship in May, and at least four securities law firms have opened investigations, with a case already filed; Nvidia stock slid 4%+. Congress is demanding answers.
- What could China get from Jensen Huang?(opens in a new tab)
- Everything we know so far about the Nvidia spy scandal targeting CEO Jensen Huang(opens in a new tab)
- Jensen Huang’s affair with alleged Chinese spy puts Nvidia’s China ambitions under scrutiny(opens in a new tab)
- The future of AI hinges on… Jensen Huang’s sex life?(opens in a new tab)
- Jensen Huang pulls out of ElevenLabs Summit amid Nvidia controversy(opens in a new tab)
- A short seller flagged Nvidia CEO’s ties to a suspected Chinese spy six months before the scandal(opens in a new tab)
- The race is on to sue Nvidia(opens in a new tab)
- Jensen Huang reportedly engaged a PR firm with a history of handling high-profile sex scandals a year ago(opens in a new tab)
- Here’s what Trump said he knew about Jensen Huang(opens in a new tab)
Anthropic buys turbines. Anthropic is spending $3.4B on natural gas turbines from OnePWR, a shift from renting compute, to secure 1.7 GW of power as grid queues lengthen.
Apple charges for AI. Apple introduced $49.99 Intelligence Usage Packs in Logic Pro to extend access to cloud-based AI features (OpenAI/Google Gemini) beyond usage limits, with unused credit expiring after 12 months.
ChatGPT Enterprise on AWS. OpenAI's ChatGPT Enterprise is now available via AWS Marketplace, allowing AWS customers to use existing purchasing and billing, with same models and features as direct.
Policy, Safety & Hardware
New AI regulator in Congress. Scott Wiener, who won Pelosi's seat, says METR's RSI report is very concerning and hopes Congress will act quickly on AI safety regulation.
Tencent warns on bioweapons. Tencent's AI chief says AI bioweapons are coming, argues neither fully open nor fully closed models solve the problem, and urges industry-wide caution, citing Hugging Face using open Chinese model GLM-5.2 when closed models refused.
Trump on AI safety. In a 60 Minutes interview, Trump dismissed AI risk as "not nearly as real" but claimed the DOJ threatened a top AI executive with jail, and said China's Xi is "all in to SI".
Google's new TPUs. Google unveiled TPU 8i Max, a dual-chip package doubling compute and HBM bandwidth for large contexts, and TPU 8tx, a confidential computing TPU that cuts latency up to 2x versus TPU 8t.
That's everything for this day – about a 5-minute read.