xAI has opened grok-build-0.1 on its public API in beta, making the model that already powers the Grok Build CLI usable without a SuperGrok or X Premium+ subscription. The framing is explicit: this is a coding model trained from the ground up for agentic work, not single-shot completion. It is built to plan what a task needs, execute step by step, debug its own failures along the way, and integrate with developer tools through MCP, the same tool-calling protocol the rest of the agent stack now speaks. The numbers attached: a 256K context window, enough for most per-file or per-module work, output served at 100-plus tokens per second, and pricing of one dollar per million input tokens and two per million output.
Where it sits is more interesting than what it does. The price lands squarely in the cheaper-models lane every lab is crowding into this week, alongside Fable 5's halving of the previous Mythos-class rate and Google's move on subscription pricing. The MCP-native part matters more than the speed: a model that speaks the standard tool-calling protocol out of the box plugs into the same substrate as everything else, including the gateway layer that the LiteLLM exploitation story showed is now under active attack. One honest caveat: the announcement ships no independent agentic-coding benchmark, no SWE-bench Verified number, no harness disclosure, so the codes-like-an-agent positioning is xAI's claim about its own model rather than a measured result.
The real signal is access, not a new frontier. The capability was already shipping inside a consumer product; opening it as a raw API priced for builders is a distribution move, putting a tool-fluent coding model in front of anyone with a key instead of anyone with a subscription. For builders the consequence is practical and small: it is one more cheap, MCP-speaking coding model to A/B against whatever you run now, and the only benchmark that counts is your own harness, because that is the one number the vendor does not control. Drop it into your agent loop, give it a real multi-step task, and watch whether the from-the-ground-up agentic training actually shows up where you work.