Anthropic launched Claude Fable 5.2 and Mythos 5.2, achieving record benchmark scores including 41% completion of real freelance jobs, with tokens generated up to 30% faster and API costs cut by 20% versus the 5.1 generation.
Nous Research released Arete-27B, a supervised fine-tune of Alibaba's Qwen3.8-27B, reducing the model's pro-Chinese propaganda rate from 22% to under 1% while preserving general capabilities, demonstrating that state bias in frontier models is a shallow layer.
Baidu released ERNIE 5.5 Flash Base, a 280-billion-parameter sparse Mixture-of-Experts language model pre-trained on 33.5 trillion tokens, shipped without fine-tuning as an open-source foundation model for developers.
OpenAI released GPT-6 Cyber, a model built to find and exploit security flaws, to vetted customers, rating it 'Critical' under its safety framework. Researchers used the model to identify nearly 1,000 working Linux kernel exploits in a single run.
Ricursive Intelligence released Ripple-65B Preview, a 65-billion parameter dense open-weight model pretrained from scratch for agentic coding. The model scores 80.5 on Terminal-Bench 2.1 and is available under Apache 2.0 license.
Liquid AI released LFM3, a family of six open-weight, 2-billion-parameter multimodal models covering text, vision, and audio capabilities. The models are designed to run on enterprise-controlled devices with company-reported competitive performance.
OpenAI launched GPT-6.1 Terra, described as a fast, efficient model combining detailed visual reasoning and precise computer use capabilities. The model is available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex.
Hugging Face released Carbon-X, a 14.8 billion parameter open DNA language model that combines DNA sequence analysis with natural language processing, allowing users to query raw DNA in chat format and receive English responses.
Chinese AI company MiniMax launched M3.1 Flash, a 428B parameter reasoning model that scores 42 on the Artificial Analysis Intelligence Index—13 points higher than its predecessor—at less than half the cost. The model is priced at $0.15 per million input tokens and achieves 73.7% on coding benchmarks, matching DeepSeek V4.1 Flash performance.
Google unveiled Gemini 4 Argon, its first frontier model in seven months, with initial access restricted to trusted cybersecurity partners. The model matches GPT-6 Astra in benchmarks but faces internal employee skepticism over performance.
Researchers from Google Cloud AI Research, Washington University in St. Louis, and UNC Chapel Hill have released EnvHarness, a programmable layer that converts static agent benchmarks into adaptive training environments for AI agents.
OpenAI presented benchmarks at Hot Chips conference showing its 700W Jalapeño ASIC, co-developed with Broadcom, delivers up to 1.9x throughput per kilowatt and 3.6x lower latency compared to Nvidia's 1,400W flagship GPU. The chip is designed for fast AI inference at scale.
Hark, which raised $700 million in Series A funding in May, has unveiled Handoff, an AI agent that can navigate and complete tasks on websites like Target, Walmart, OpenTable and LinkedIn without requiring APIs. The company claims its agent outperforms previous-generation frontier models including GPT 5.4, GPT 5.5, Opus 4.8, and Gemini 2.5 Pro.
A multi-domain evaluation suite for coding agents spanning four work domains — Code, Web, Office, and Security — with contamination-resistant task construction that reverse-engineers tasks from real commits and pull requests, released openly with full task directories, evaluation harness, and reference solutions.
Chinese AI lab Moonshot AI launched Kimi K2.7 Code, a one-trillion parameter open-weights coding model priced up to 12x cheaper than GPT-5.5 and Claude Opus 4.8, though it trails those models in performance benchmarks and lacks independent verification.