# Grok Build 0.1 API Retirement: How to Replace the Model Without Breaking CI

Find stale model IDs, check what your xAI endpoint actually serves, and test a Grok 4.7 migration before changing every caller.

By Paras Madan | 2026-11-21


SpaceXSI [retired `grok-build-0.1` from the xAI API on November 19](https://docs.x.ai/developers/release-notes). If a coding assistant or review job still sends that model ID, a retry won't restore it. The retirement note does not name a drop-in replacement, and it does not say that the separate Grok Build app or CLI has shut down.

For an API integration, `grok-4.7` is a candidate to test. Move one read-only workflow before changing a model default shared by every job. The endpoint, request fields, response parser and usage assumptions all deserve a check.

## Find the ID at the point where requests leave

Look for `grok-build-0.1` in your application config, CI variables, queued jobs and gateway aliases. A model can disappear from the current code and still get picked by an older worker or a fallback. Record the effective provider, model ID and endpoint in staging logs without capturing API keys or request content.

Don't conclude the old ID still works just because its model page is indexed or a gateway lists it. The retirement concerns the direct xAI API. If you use a different gateway, confirm what it actually serves and bills instead of assuming the listing and the upstream route are equivalent.

## Check the model list with the same endpoint and key

The [xAI model API documentation](https://docs.x.ai/developers/rest-api-reference/inference/models) describes `GET /v1/models` as the list available to the authenticating key. Use the same approved endpoint and credentials as the worker to check whether the candidate appears. That is a better test of the route than reading a public model catalog.

If your app uses the US endpoint, read the [regional endpoint documentation](https://docs.x.ai/developers/advanced-api-usage/regions) before changing the base URL or model. It lists Grok 4.7 on the US route and notes that unsupported model requests there can return 404. A region-limited worker should fail clearly rather than silently try a different region or vendor.

## Verify the response your worker expects

Start with a small synthetic review request: give the worker a simple formatting-only change and ask for the same structured review output it already consumes. Run it on the same API surface as production. Check both a positive case that should create a finding and a negative case that should return none.

The [Grok 4.7 model page](https://docs.x.ai/developers/models/grok-4.7) documents its supported features. Separately, the release notes say that on the Responses API, Grok 4.7 returns encrypted reasoning content even when not requested in `include`. Test how your Responses client handles the item rather than assuming every result is the final text. If your workflow streams, verify that your parser assembles the final output correctly before applying the same change to batch callers.

Keep the model ID explicit for the first comparison. Grok 4.7 Fast in the Grok Build product is not a public xAI API option, so it can't substitute for a working API model ID.

## Update the cost estimate before moving long jobs

The official Grok 4.7 release notes list $2 input, $0.50 cached input and $6 output per million tokens below the 200k prompt tier, and $4, $1 and $12 respectively above it. Don't compare a short fixture at the lower tier with a production job that sends an entire repository into a much larger prompt.

Record input, cached input, output, latency and whether the review produced a usable result. A run that passes tests on small diffs but breaches the cost or latency budget on large ones isn't ready for a blanket migration. Change the smallest group of callers that shares an evidence-backed configuration, and keep a record of why you selected it.

I'm Paras. I write about AI tools and how engineering teams use them.

If you're changing an API or retiring a feature and want customers to know exactly what to do next, I can help with the writing.

[Book a 30-minute call](https://cal.com/paras-madan/30min).
