Agentic Coding Goes Commodity: xAI Undercuts, Mistral Rebrands, Copilot Expands

xAI just dropped agentic-coding pricing to $1/$2 per million tokens. Mistral retired Le Chat for Vibe. Copilot Studio added Mistral Medium 3.5. The agent layer is consolidating into the toolchain.

The agentic-coding market hit a pricing inflection this week. xAI shipped a $1-input / $2-output model purpose-built for agent loops, Mistral retired its consumer-chat brand and relaunched as a build-and-run agent platform, and Microsoft pulled Mistral's mid-tier model into Copilot Studio's governed agent runtime. None of these are isolated launches — they're the same trend showing up at three layers of the stack on the same week.

The pricing floor just moved

xAI's grok-build-0.1 announcement on May 29, 2026 puts the model into public beta at $1 per million input tokens and $2 per million output tokens — the same model that drives the Grok Build CLI. It ships day-one through OpenRouter, Vercel AI Gateway, Cursor, Hermes Agent, OpenClaw, Kilo Code, and OpenCode, which tells you xAI cares more about embedding into existing agent harnesses than about pulling users to its own surface. A day earlier, Grok Build 0.2.7 added shared terminals across subagents and a /usage command — small features that only matter once you're running multi-step agent loops in production, which is the workflow xAI is now pricing for.

Mistral picks a side

Le Chat is gone. The product is now Mistral Vibe, positioned as an AI chat and coding agent where you build, test, and run agents and apps, with custom-model training bolted in. The naming shift matters less than the surface-area shift: Mistral is no longer competing with ChatGPT on consumer chat, it's competing with Cursor and Claude Code on the build loop. The enterprise wedge landed the same week — Microsoft added Mistral Medium 3.5 to Copilot Studio on May 28, 2026 with in-region data control, giving European buyers a sovereignty-friendly model inside Microsoft's governed agent runtime.

What this means for builders

For working developers, the calculus on agentic coding just got simpler: at $1/$2 per million, you can run aggressive multi-step agent loops without flinching at the bill, and the model is already wired into the IDE and CLI tools you probably use. For buyers evaluating enterprise agent platforms, model choice inside governed runtimes (Copilot Studio, Bedrock, Vertex) is becoming the differentiator — not the runtime itself. And for anyone sizing infrastructure, ValorC3 cites Morgan Stanley estimates of $32.5–60 billion in new CPU demand from agentic workloads by 2030, with CPU-to-GPU ratios shifting from 1:8 toward 1:1 — agents are CPU-heavy in ways pure inference isn't.

The bottom line: If you're still paying premium-tier prices for agent loops, this is the week to re-benchmark — the floor is now $1 in, $2 out, and it ships inside the tools you already use.

Sources