Mistral gives coding agents an open-weight upgrade

Welcome back. Following Meta's lead, Europe is going all in on open-weight models. Mistral just dropped its most powerful coding and security model yet, claiming the top spot over the competition. Meanwhile, Anthropic is giving devs more room to work with, with a significant increase to Claude’s API request size limit.

Also: Elon Musk shares an unreal Neuralink demo, comparing the real dollar value of LLM subscriptions, and 10 code slop indicators in your codebase.

Mistral gives coding agents an open-weight upgrade: The French AI lab just dropped Mistral Large 4 on Hugging Face, and it’s gotten a serious boost. Mistral says extra reinforcement learning pushed its Terminal-Bench 4.0 score from 28% to 41%. It now also matches China’s GLM-5.3 on Artificial Analysis’s Coding Agent Index while using 47% fewer tokens per task. That could mean big savings if you’re running agents at scale, although closed models like Anthropic’s Mythos 5.1 still lead. Grab the weights here.

Anthropic gives developers 4x more room per request: The AI lab just bumped the request size limit for Messages and Token Counting on the Claude API from 32MB to 128MB. That’s good news if your long computer-use sessions or hefty PDFs kept hitting the old cap with context to spare. Requests with over 20 images can also use 3000px images, up from 2000px. Cloud platforms still cap at 20–32MB, though. Check the docs.

Chinese labs reportedly switch distillation from Claude to OpenAI: The ChatGPT maker’s GPT-6 Astra has become the new favorite for training cheaper rivals on frontier model outputs. According to a report, Moonshot and DeepSeek both used Astra for models released this month, though neither has confirmed it. Earlier this month, Anthropic reported that new safeguards cut data scraped from Claude models in distillation attacks by 90% to 99%, which likely explains the shift to Astra. And if you’re building on frontier APIs, tighter usage checks could be coming your way.

Sensors, cameras, and microphones capture what’s happening in the real world, but turning raw signals into events your app can act on usually means a custom model per use case: collect, train, tune, repeat.

On Nov 5, see how Archetype AI’s Newton, a physical world model pretrained on time-series sensor and video data, helps developers skip that cycle. Learn:

Save your spot

An identity crisis. It’s 2026, and while some software engineers are depressed, it appears that others are having the time of their lives. Or at least that’s how Anthropic’s head of platform, Katelyn Lesse, described the state of engineering in a recent post. Her point is that AI won’t deliver on its potential until engineers solve a whole other set of problems.

From chats to swarms. Agents have come a long way from chatbots waiting for your next prompt. Now they run longer in the cloud, spin up subagents, and can even hand off jobs like grading, advising, and “dreaming.” Some teams are starting to point thousands of persistent agents at research and hardware design. This scale creates some big engineering challenges.

The new to-do list. Lesse points to four areas that need work in this situation:

Still plenty to build. If you’ve been wondering whether your engineering skills are becoming obsolete, Lesse has good news. The problems worth solving are changing, and at the end of the day it is your skills that could decide whether they get solved. Her advice is to go deep, learn new things, and form your own opinions. You can read the full essay here.

Factories, wells, and grids generate machine data every second. Tiger Data keeps it all in one Postgres database:

Start a free trial on Tiger Cloud now. No credit card required.

How to build and stress-test a SaaS with Claude Code: This tutorial walks through building a full B2B app from one PRD, deploying it, then handing the live product to TestSprite to try and break it in real browsers. You’ll learn how to test signup, logins, plan limits, checkout, and promo codes, feed failures back to Claude Code, and keep fixing until everything passes.

Treg (6.5k ⭐): Think OpenRouter, but for the tools your agents use. Give an agent one token and it can search and call thousands of APIs, CLIs, and shared skills, from SEO and enrichment to social, ads and web search, without signing up for every provider or exposing your keys. You can also plug in your team’s own tools and credentials and share them across agents.

How to orchestrate a one-person software factory: This guide shows how to run multiple coding agents without becoming their full-time project manager. You’ll learn how to split work into lanes, use orchestrators and owners, coordinate everything through GitHub, build reviews into the workflow, and only pull yourself in when a decision actually needs you.

An AI engineer shared a two-line fix for review loops. Codex Astra flags bugs that are technically possible but almost never happen, Opus fixes them, and Astra flags the next one. Forever.

Real bugs still get fixed. Low-scoring ones land as filed issues instead of churn. Tune the threshold of 5 to your project.

A founder just open-sourced one of the cleanest dashboard designs on the internet, and you can grab the full code on GitHub

If you enjoyed this newsletter, you should definitely check out our IT and cybersecurity newsletters. Subscribe using the links below:

We put together 20+ hands-on guides covering coding agents, agentic engineering, and leadership frameworks to help you level up your career. You can get them here.

Advertise with us: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team