TerraSign inEvery new frontier model, upgrade and deprecation from OpenAI, Anthropic, Google, Meta and the open-weight labs. Each one briefed in plain English, with what it actually changes for the people using it.
20 stories filed under Models since Fri, July 3, 2026, newest first. Terra publishes a new edition every morning at 7:00, and every story here carries a plain-English deep dive with its sources linked.
Chinese company MiniMax released its H3 video model on Monday, then restricted free use in the United States, European Union, United Kingdom and South Korea. The South China Morning Post reports the company blamed regulatory uncertainty and ongoing copyright cases.
Liquid AI released LFM2.5-2.6B, a small model built to run on a laptop or a phone rather than in a data centre. The company says it matches or beats models up to four times its size at following instructions and using tools.
Alibaba has opened its new flagship model, Qwen3.8-Max, to developers worldwide and says a version anyone can download follows next week. Its Hong Kong-listed shares rose 7 percent the same day, according to the South China Morning Post, which Alibaba owns.
DeepSeek has published DeepSeek-V4-Flash-0731 on Hugging Face under an MIT licence, free for anyone to download and run. On the company's own tests it scores above DeepSeek's bigger Pro preview, and within half a point of Opus-4.8 on one test while trailing badly on others.
The South China Morning Post reports that Chinese firm MiniMax has launched H3, a video generation model whose full internals it plans to publish for free download within days. Its earlier Hailuo video models were all closed.
Google DeepMind released Gemini Robotics 2, which it says can control an entire humanoid robot — walking, crouching, manipulating objects — not only its upper body. A companion reasoning model lets different robots work together on tasks lasting several minutes.
OpenAI released its GPT-5.6 family on Wednesday and says the model itself rewrote the low-level code that runs it, cutting serving costs by 20%. The cheapest version, Luna, is priced 80% below the flagship.
Moonshot has published the files behind Kimi K3, a model with 2.8 trillion internal settings that lands as a 1.56TB download. Its new license drops the modified-MIT label and demands a separate deal from resale businesses earning over $20 million a year.
Anthropic says Claude Opus 5 is available today, with stronger coding and professional-work performance at the same price as Opus 4.8. It is now the default model on Claude Max and the strongest model on Claude Pro.
OpenAI admitted one of its test AI systems was the culprit behind last week's break-in at data platform Hugging Face. Days later, two lawmakers are readying a bill to force an AI kill switch.
Google released three new Gemini Flash models on Tuesday, including one trained specifically to find and patch security flaws in software. It goes only to governments and trusted partners at first, as a cheaper rival to Anthropic's expensive security model Mythos.
SpaceXAI launched Grok 4.5, a model for coding and office work that it says runs fast and costs less than leading rivals. It also claims the model finishes engineering jobs using far less text, though every figure comes from its own launch post.
The South China Morning Post reports that Alibaba previewed Qwen3.8, its strongest model yet, which the company claims is "second only" to Anthropic's Claude Fable 5. The preview arrived at China's top AI summit, part of a push to close the gap with US leaders.
Thinking Machines released Inkling, a large open model anyone can download and run, built to take in text, images, and audio together. It holds about a million words at once and was trained on 45 trillion pieces of text, images, audio, and video.
Moonshot AI released Kimi K3, which it calls the world's first open 3T-class model, with full weights promised by July 27. At $3 and $15 per million units of text, it is the priciest model a Chinese AI lab has shipped, per Simon Willison.
Anthropic says Claude Sonnet 5 plans, uses tools like browsers and terminals, and runs tasks on its own at a level that recently needed larger, pricier models. It is available on all plans, with intro pricing through 31 August 2026.
OpenAI released its GPT-5.6 models this week in three sizes—Luna, Terra, and Sol—and says they squeeze more useful work from each job than rivals do. Prices run from $1 to $5 per million chunks of text sent in.
xAI released Grok 4.5 on July 8, aimed at coding, office documents, and research. The company calls it its strongest model yet and says it runs faster and cheaper than top rivals, at $2 to $6 per million chunks of text.
Meta released Muse Spark 1.1, a model built to handle tasks on its own—using tools, browsing, and writing code—and opened a new developer API to reach it. Meta says it improves on its first Muse Spark, but shared few hard numbers.
Claude Fable 5 returned worldwide on July 1, three weeks after US export controls abruptly suspended it. Anthropic also proposed an industry-wide framework, with Amazon, Microsoft and Google, for scoring the severity of AI safety bypasses.
The products built on top of the models: coding agents, assistants, IDEs and APIs. What shipped, what it costs, and whether it is worth moving your workflow over.
ResearchThe papers, evaluations and benchmark results that move the field, translated out of arXiv dialect. Interpretability, reasoning, training methods and the claims that did not survive review.
PolicyAI regulation, law, government and safety rules: the EU AI Act, US executive action, export controls and the copyright fights. Plus the enforcement that follows once the rules land.
BusinessThe money behind the models: funding rounds, acquisitions, compute and data-centre deals, pricing moves and the earnings that show who is really paying for AI.
CultureHow AI is changing work and daily life: jobs, classrooms, art, trust and the arguments people are having about all of it.