🤚 The Open-Palm Product Dump

Mistral AI held its AI Now Summit 2026 this week and emerged with the kind of announcement cadence usually reserved for companies trying to distract from a quarterly earnings miss. Except Mistral isn’t distracted — it’s ambitious, which in the current AI market is the same thing as breathing.

The headline act: Mistral Medium 3.5, a 128-billion-parameter dense model with a 256,000-token context window, released under a modified MIT license with open weights. The model scores 77.6% on SWE-Bench Verified, surpassing both Devstral 2 and Qwen3.5 397B. It ships with configurable reasoning effort per request, a custom vision encoder trained from scratch, and the quiet confidence of a model that can be self-hosted on as few as four GPUs.

Pricing sits at $1.50 per million input tokens and $7.50 per million output tokens, which puts it comfortably in the “we’d like enterprise customers but we won’t say no to a hobbyist with a credit card” tier.

But the model was merely the appetizer. The entrée is Vibe — Mistral’s new unified agent platform that combines coding and productivity into a single subscription. Because apparently every AI company must now offer a terminal-based coding agent or surrender its Series C.

👐 The Two-Handed Agent Inventory

Vibe ships in two flavors: Work Mode and Code Mode, because Mistral has correctly identified that the modern knowledge worker oscillates between “writing a report” and “fixing a bug” approximately forty-seven times per day.

Work Mode connects to your entire enterprise stack — Google Workspace, Outlook, SharePoint, Slack, and GitHub — and offers structured data analysis with chart rendering, document synthesis, and task scheduling. It is, in essence, the middle manager that never takes a sick day and never asks for equity.

Code Mode provides GitHub integration, isolated sandbox sessions that persist offline, parallel session execution, and pull request generation with visible diffs. The new VS Code extension adds full project context awareness, automatic test writing, and integration with GitHub, GitLab, Jira, and Linear. There’s also a CLI with session-scoped permissions and a /teleport command that moves sessions between your terminal and the cloud, which is the kind of feature name that makes you wonder if the product team plays too many video games or exactly the right amount.

Pricing: Free for basic tasks, $14.99/month for Pro, $24.99/user/month for Teams, and “call us” for Enterprise — the universal AI pricing ladder that ends with a sales conversation nobody wants to have.

🌿 The Gentle Awakening

What’s quietly remarkable about this summit is the breadth. Mistral didn’t just drop a model and a coding agent — it also announced Physics AI, an entirely new class of models designed to predict physical system behavior for engineering acceleration. It launched a Search Toolkit for production search pipelines. It introduced Studio Connectors with built-in and custom MCP support. It hired Emmi to accelerate AI-native industry adoption. And it partnered with NVIDIA to accelerate open frontier model development.

This is not the behavior of a company content to be “the European one.” This is the behavior of a company that looked at the Anthropic-OpenAI-Google triopoly and decided that the correct response was not to pick a lane but to build all the lanes simultaneously while speaking French.

The open-weights release of Medium 3.5 is particularly pointed. While OpenAI guards its weights like nuclear launch codes and Anthropic treats model access as a philosophical commitment, Mistral continues to bet that openness is a competitive advantage. Four GPUs. Modified MIT. Come and get it.

👑 The Gold-Leaf Market Analysis

The AI agent market has now reached the phase where every major lab has its own coding CLI, its own VS Code extension, its own sandbox environment, and its own opinion about how pull requests should be formatted. Claude Code, Codex, Gemini Code Assist, Grok Build, and now Vibe are all competing for the same developer terminal, and the developer in question is starting to feel like a dating show contestant who said yes to too many roses.

Mistral’s advantage — if it has one — is the combination of open weights, European data sovereignty compliance, and aggressive pricing. The disadvantage is that it’s competing with companies whose marketing budgets exceed Mistral’s entire annual revenue. But in AI, the model is the marketing, and a 77.6% SWE-Bench score at $1.50 per million input tokens is the kind of pitch that doesn’t need a Super Bowl ad.

The Physics AI announcement, meanwhile, is the sleeper move. While everyone else fights over who can write the best React component, Mistral is quietly positioning itself as the AI company that also does actual engineering — the kind involving molecules and materials and things that exist in three dimensions. If that bet pays off, the coding agent wars will look like a warmup act.

“They released a model, an agent, a physics simulator, a search engine, an NVIDIA partnership, and a VS Code extension in a single week. The press release was longer than the model’s context window, and somehow both were still too short.” — The Slap of Wisdom European Bureau, writing this from a café in Paris where the espresso costs more than a million input tokens