Models, agents and infrastructure

DeepSeek and Google release agent-focused models, while OpenAI expands persistent ChatGPT context and Mistral outlines a European compute push.

Briefs

Frontier models

DeepSeek launches V4-Pro and open agent harness — DeepSeek launched V4-Pro with adjustable reasoning effort and Responses API support. It also released its MIT-licensed Cordis-based agent harness in preview, while reporting strong coding, terminal and vulnerability-discovery results in its release announcement, Terminal-Bench comparison, security benchmark comparison and harness release.
Gemini 3.7 Flash promises cheaper coding agents — Google says Gemini 3.7 Flash improves coding and agent work while launching at half Gemini 3.6 Flash’s introductory price. The model is available through the API and AI Studio, with vendor results detailed in the launch announcement, pricing and benchmark release and Google announcement.
OpenAI tests GPT-5.6 Sol Ultrafast mode — OpenAI is previewing an Ultrafast mode for GPT-5.6 Sol that it says can run up to 14 times faster than standard mode. The API preview begins with a select customer group, as outlined in the Ultrafast mode announcement.

Agent engineering

Rails benchmark tests coding agents on real tasks — Rails compared eight models across 21 practical Ruby tasks. Claude Opus 5 led its three-run evaluation, while cost, duration and steps varied widely in the benchmark announcement.
Grok coding-cost comparison lacks like-for-like evidence — Grok 4.6 reportedly completed a Rust library rewrite for about $55 in token costs. The coding demonstration compares it with Fable’s earlier implementation.

Platforms and privacy

ChatGPT adds computer activity history — OpenAI says ChatGPT’s desktop app can use activity across apps and websites to inform later chats. Computer History could reduce repeated explanations during ongoing work, as described in the OpenAI announcement.
X opens ranking weights and tests visibility labels — X has open-sourced code for filtering and ranking its For You timeline, including visibility labels and their effects. Its restricted Under the Hood pilot lets eligible users inspect labels applied to accounts and posts through the X Open Source release.

Infrastructure and robotics

Mistral plans a European compute business — Mistral plans regional inference endpoints, a Priority Tier and five-year European Compute Unit commitments. Its technical write-up sets targets of 200MW by 2027 and 1GW by 2030.
Tau pilots humanoid cleaning in San Francisco — Tau Robotics is piloting a paid San Francisco cleaning service at $30 per robot-hour. Invited households can book one or two robots for cleaning and decluttering, shown in its service site and timelapse announcement.

Interviews & Talks

Karpathy maps modern LLM capabilities and limits

Andrej Karpathy’s lecture gives a clear tour of how modern language-model products work in practice. He separates a model’s trained knowledge from its context window, then shows why useful systems often need external tools. Search can bring in current facts. A Python interpreter can handle calculations more reliably. Code tools, multimodal input, memory and custom assistants each extend what a chat interface can do.

The useful lesson is not to treat every answer as the product of a single all-knowing model. The surrounding tools shape the result, and each tool brings its own strengths and failure modes. That makes the lecture a practical guide for choosing when to search, calculate, code or provide more context.

Worth watching

00:04 — Karpathy explains the context window as the model’s working memory and distinguishes it from the model’s stored knowledge.
24:32 — He demonstrates how web search supplies current information that a model’s training data may not contain.
53:09 — He shows how a Python interpreter can improve numerical reliability while warning that AI-generated analysis still needs scrutiny.