The API & Grok Build: building on Grok, eyes open.
For builders — and for the merely curious who want Grok 4.3 without the $300 subscription — the xAI API is the side door worth knowing. It's cheap, it's capable, and it comes with the most consequential fine print in the product: a free-credits program that pays you in compute for your data. We'll read that fine print together.
01 The API in plain terms
The xAI API serves the Grok model family at per-token prices — Grok 4.3 sits around $1.25 in / $2.50 out per million tokens with a 1M-token context window, which is aggressive pricing for a frontier model (compare our Fable 5 notes: $10/$50). Practical takeaways even for non-engineers:
- The API is the cheap test of the top model. Curious whether 4.3 beats your subscription's model on your work? A few dollars of API credits in the console's playground answers it — no Heavy upgrade required.
- If you run OpenClaw or other BYOK tools (our track), Grok is a plug-in brain: paste an xAI key instead. The firehose doesn't ride along, though — X data access via API has its own gates.
- Spend caps, always. Same rule as every key we've ever taught you to create: name it, cap it, revoke it without ceremony.
02 The free-credits trade, spelled out
xAI offers up to $175/month in free API credits through its data-sharing program: in exchange, your API traffic — the prompts and outputs — can be used to train and improve xAI's models. For hobby projects on public data, that can be a fine trade, and $2,100/year of compute is real money. For anything touching client information, business documents, user data, or ideas you'd rather not donate: it is not a fine trade at any price. The decision rule: if you wouldn't paste it into a public forum, don't run it through a training-enabled tier. Check the program's current terms yourself before enrolling — this is exactly the kind of clause that changes between our updates.
03 Grok Build: the coding model
What it is, who it's for
Grok Build (public beta) is xAI's dedicated agentic coding model — 256K context, text + image input (mockup-to-code works), built for multi-step coding tasks rather than single completions. It slots into the same world as Claude Code and GitHub Copilot's agents: you describe the outcome, it plans and executes across files.
Update — July 15, 2026: SpaceXAI open-sourced the Grok Build agent and terminal UI — the full source (agent loop, tools, extension system) is now on GitHub. Two things that matter for the "eyes open" theme of this lesson: you can now read exactly how the harness assembles context and dispatches tool calls, and Build can run fully local-first — compile it yourself and point it at your own inference, keeping your code on your machine.
Update — August 19, 2026: don’t confuse the two things called “Build.” What this section covers is Grok Build, the coding model and terminal agent for people who write code. Separately, the Build Mode inside the Grok chat app — describe an app, a game or a dashboard and Grok writes and runs it live in the thread — came out of its SuperGrok Heavy early beta and is now on every plan, free included, on web, iOS and Android. Published apps get their own grok.me address, and you can export the project to GitHub if you want to keep going in a real editor. That’s the on-ramp; this is the workbench. We cover it in Build Mode: describe an app, get a working app.
Update — September 16, 2026: Grok Build now has memory. After each completed turn, the terminal agent reviews what happened in the background and writes durable notes — how the team writes and reviews code, decisions and the reasoning behind them, project facts like which command runs the test suite. Notes are kept per project plus a global set for preferences, and a /dream pass folds them into one topic file per subject. When you come back, Build reads the topics relevant to the area it is about to touch, so “use just test, not cargo test” only has to be said once. SpaceXAI says task state, tentative conclusions, secrets, and anything already in the repo’s docs are deliberately left out, and instructions in the current conversation override any note. /memory opens a read-only browser of every note, which is also how you find the file to fix when a note is wrong. Eyes-open angle: memory files are plain markdown on your machine — it is worth a skim before you share a project directory, because a note can capture more context about how your team works than the code does.
Honest placement for our readers
- Building a real product? Claude Code and GitHub Copilot remain the mature picks with deeper ecosystems (our Claude Code lesson, GitHub Copilot lesson). Build is the challenger worth a bake-off, not yet the default.
- Non-coder building a small tool? Our Build-an-App track patterns work with any of these — Build's mockup-to-code is a genuinely nice on-ramp.
- The bake-off prompt: give the same small, real task to Build and your current tool; judge on completed-without-babysitting, not on vibes of the first response.
04 Subscription or API?
The decision in one breath: subscriptions buy the product (apps, voice, Imagine, Connectors, the firehose in chat); the API buys the brain (models, by the token, in your tools). Heavy daily chat user → subscription. Builder, tinkerer, or once-a-week power user → API credits may cost less than $10/month for the same model quality. Both → that's common and fine; just cap both meters.
The $5 experiment
Open an xAI console account, set a $5 cap, and run your three hardest recent prompts against Grok 4.3 in the playground. Decline the data-sharing program for the test (pay the $5 — your prompts stay yours). You'll learn more about the model ceiling than any benchmark chart can tell you.
What you can do now
- Use the API as the cheap test of top-model quality — capped, named keys, always
- Plug Grok into BYOK tools and know what doesn't ride along (the firehose)
- Price the free-credits trade correctly: compute for training data — fine for public, never for client work
- Run a real bake-off between Grok Build and your current coding tool
- Choose subscription vs API by what you're actually buying: the product or the brain