Rocky Mountain Madman LLC

The standard way to keep an AI agent in line is to have a second AI read over its shoulder. Goodfire, the interpretability startup, has just shipped an alternative: monitors that watch what is happening inside a model's calculations while it works, instead of auditing the words it produces afterward.

Buried under this week's frontier-model headlines, Google DeepMind shipped something many working developers will get more practical use out of than any trillion-parameter announcement: EmbeddingGemma 2, a 740-million-parameter open embedding model released under Apache 2.0. It maps text, code, images, video, and audio into a single shared vector space, carries an 8K-token context window, and ships with modular text, vision, and audio encoders designed for on-device deployment.

In a week full of model launches, the most unusual announcement involved no model at all. OpenAI, Anthropic, and Google DeepMind have confirmed they spent several weeks coordinating on AI safety, jointly unveiling cyber-focused AI models, safeguards, and access programs developed collaboratively across the three labs. One report describes a formally announced joint standards body publishing common testing protocols for frontier models.

OpenAI's "28 Days of Continuous Releases" stunt hit Day 3 with its biggest consumer move yet: the GPT-6 model series is now rolling out globally to every ChatGPT user, free and paid. Plus and Team subscribers get GPT-6 Sol, while free and Go users will see GPT-5.6 gradually replaced by GPT-6 Luna. Alongside the rollout, OpenAI says Codex and Work weekly active users have passed 40 million, with usage quotas reset across the board.

Mistral has finally shipped a frontier-class model, and it is a statement piece. Large 4, immediately nicknamed "Le Chonk" by an internet that could not resist, is a natively multimodal mixture-of-experts model with 1.05 trillion total parameters and 49 billion active parameters, wrapped around a 1-million-token context window. It was trained from scratch on 3,800 Grace Blackwell GPUs in Mistral's own European datacenters — no rented American cloud, no borrowed infrastructure — and it is available right now as a public preview through Mistral's API.

Anthropic released Claude Haiku 5.5 on October 7 with a price cut that rearranges the math of production AI work: about 90 percent cheaper than Haiku 4.5 on the prompts that matter. Short-context requests up to 100,000 tokens now cost 10 cents per million input tokens and 50 cents per million output tokens, with cache reads at a penny. Above the 100k line the discount drops to 50 percent.

Reflection AI has unveiled Beam, a 501-billion-parameter text-only mixture-of-experts model with just 23 billion active parameters, trained on 23.8 trillion tokens with a 1-million-token context window. The company's headline claim: Beam matches GLM-5.2 on reasoning while using three to four times less inference compute. The weights are expected later in October under Apache 2.0, once red-teaming wraps up. In a market obsessed with parameter counts, Beam is a bet that the number on the box matters less than the number actually doing work.