How chatbots and other language models are built, trained and improved.
Language models are the systems behind chatbots, writing assistants and code helpers: trained on enormous amounts of text to predict what comes next, then shaped by further training to follow instructions and answer safely. This page collects what we have written about them — new models and what they can actually do, the training and alignment methods behind them, how they are priced and served, and the independent results that test the claims. Where a story turns on a technical idea, such as attention, context windows or fine-tuning, we link to the lesson that explains it properly rather than leaving the term hanging.
The French company Mistral has opened a public preview of Mistral Large 4, a natively multimodal model with 1 trillion parameters, of which 49 billion run for any one word, and a 1 million token context window. It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacentres, on data spanning more than 160 languages. Mistral says it will publish the open weights by the end of October 2026. Its most striking claim is in security: on a test that asks a model to reproduce a real software vulnerability and then patch it, Mistral Large 4 scores 82%, the highest of any model, while Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse the task.
References: Mistral: model documentation, parameters and pricing · Mistral on Hugging Face, where the weights are expected
From 9 October, free Gemini users will only get the Flash-Lite model, and AI Plus subscribers lose Pro.
OpenAI has added an Ultrafast mode to GPT-6 Astra, running on NVIDIA Blackwell GPUs. NVIDIA says it generates tokens up to 8x faster than the standard mode. It is available in the OpenAI API now and open to eligible ChatGPT Work and Codex users. Check OpenAI's pricing page before using it.
References: OpenAI: Ultrafast pricing · NVIDIA blog
Google's new frontier model can write answers about 15 times longer than before, goes to cyber defenders first, and is already rewriting Google's own code. Here is what it has done, and why it is different.
The whole internet is now filled with AI-written text. Humans are not putting any effort into writing text with AI, and almost a third of text is now written by AI. This has led to worthless training of AI, and it is not gaining any more knowledge now.