Founded in 2023 and backed by substantial venture funding, Mistral AI has built a credible alternative to closed AI labs by releasing capable open-weight models alongside commercial offerings.
Mistral's Leanstral 1.5 focuses on generating formal proofs in Lean, aiming to make machine-verified mathematics more accessible to developers and researchers.
A benchmark of GLM-5.2 running on AMD's MI355X reports 2,626 tokens per second per node while claiming more than 2x lower cost than an equivalent Nvidia Blackwell setup — a signal that AMD is becoming a real inference option.
GitHub has added Kimi K2.7 Code to Copilot's roster of selectable models, giving developers another coding-focused engine to route their tasks through—no longer a preview or waitlist.
Cloudflare is forcing AI companies to distinguish training and agent crawlers from search indexers — or face automatic blocking across a large share of the web.
Wix-owned app-building platform Base44 is rolling out a proprietary AI model, betting that owning the model layer is the only durable competitive advantage in the crowded no-code AI space.
Google is rolling out Gemini's personalized image generation to free-tier users in the US, letting the model draw on your interests and connected Google app data to produce tailored visuals.
The platform behind the most widely cited AI model rankings reached a nine-figure revenue run rate by commercializing the same human-preference data that powers its free leaderboard.
A circulating report claims the Trump administration authorized over 100 companies and government agencies to use an Anthropic AI model, but key details remain unverified and the model name cited does not match any known Anthropic product.
With Anthropic's export controls still blocking access across much of Asia, a wave of regional startups is shipping frontier-class models that compete directly with Claude. Here's what builders need to know.
The gap between open-weights and proprietary LLMs has narrowed dramatically — but it hasn't closed. Here's where the differences still matter for builders making architecture decisions.
OpenAI has quietly limited who can access GPT-5.6 following a government request, while publicly stating that gatekeeping AI models this way should not become standard practice.
