Grok 4.1 Fast models are being shut down on the Gemini Enterprise Agent Platform on 20 August 2026
The xai/grok-4.1-fast-reasoning and xai/grok-4.1-fast-non-reasoning model IDs stop working on 20 August 2026. Here's how to migrate.
If your application calls xai/grok-4.1-fast-reasoning or xai/grok-4.1-fast-non-reasoning through the Gemini Enterprise Agent Platform, you have a hard deadline of 20 August 2026 to migrate. After that date, every API request to those model IDs will return a 400 error. There is no grace period.
This is not a soft nudge. It is a hard shutdown.
What is being turned off
Google announced the deprecation of the Grok 4.1 Fast model family on 17 July 2026. The two affected model IDs are:
xai/grok-4.1-fast-reasoningxai/grok-4.1-fast-non-reasoning
Both are served through the Gemini Enterprise Agent Platform’s Model as a Service (MaaS) endpoint:
https://aiplatform.googleapis.com/v1/projects/<your-project>/locations/global/endpoints/openapi/chat/completionsFrom 20 August 2026, any request using either of those model IDs against that endpoint will fail with a 400 error. Your application will not degrade gracefully. It will break.
A bit of background
The Grok 4.1 Fast family was designed for speed and cost efficiency. The non-reasoning variant was xAI’s most economical non-thinking model, well suited to high-volume, lower-complexity tasks such as summarisation and categorisation. The reasoning variant offered structured thinking at a similarly low price point.
It is worth knowing that xAI itself had already retired earlier Grok models, including Grok 3 and the original Grok 4 Fast releases, back in May 2026. The Gemini Enterprise Agent Platform’s 20 August cutoff is Google Cloud’s own enforcement on top of that, applied separately to its MaaS layer. If you were calling xAI’s API directly, your deadline was earlier. If you are calling through Google Cloud, 20 August 2026 is your line.
The Gemini Enterprise Agent Platform, which replaced Vertex AI following Google’s April 2026 rebrand, exposes over 200 models through its Model Garden, including xAI’s Grok family, Anthropic’s Claude, Meta’s Llama, Mistral, and Google’s own Gemini and Gemma models. Model retirements like this one are a predictable part of operating on a multi-vendor platform as the model ecosystem moves quickly.
What you should migrate to
Google’s release notes point to three migration paths:
Grok 4.2, available in Model Garden, suitable as a direct drop-in if you need continuity with xAI’s model family.
Grok 4.3, xAI’s current flagship, also available in Model Garden, with stronger general capability than the 4.1 Fast variants.
Grok 4.20 (Reasoning), xAI’s flagship reasoning model, noted for a low hallucination rate and strong performance on document understanding and long-horizon agentic tasks.
If your workload does not specifically require a Grok model, the broader Model Garden gives you alternatives across Gemini, Claude, Llama, and others.
The good news is that the migration itself is straightforward. Grok on the Gemini Enterprise Agent Platform uses an OpenAI-compatible interface, which means you can use the standard OpenAI Python library and simply swap the model ID string. For most teams, this is a one-line change in configuration, not a structural refactor.
What this means for you
If you have already migrated away from Grok 4.1 Fast, you can stop reading. If you have not, the path is clear:
- Identify every place in your codebase or configuration where
xai/grok-4.1-fast-reasoningorxai/grok-4.1-fast-non-reasoningappears as a model ID. - Replace it with your chosen successor, such as
xai/grok-4.2,xai/grok-4.3, orxai/grok-4.20-reasoning. - Test your application against the new model ID before 20 August 2026.
- Deploy.
One practical note: Grok models on the platform are subject to global quotas measured in queries per minute (QPM) and tokens per minute (TPM), with TPM covering both input and output tokens. If you are migrating a high-volume workload, check your quota headroom on the new model ID before you switch in production.
Because the Gemini Enterprise Agent Platform supports multi-model agent architectures, a single agent can call different models for different sub-tasks while sharing a unified identity, policy, and audit trail. That means you can migrate individual task steps to a new model without rebuilding your entire agent architecture.
The broader pattern worth noting
Model deprecations like this one will keep happening. The xAI model family in particular has moved quickly, with several generations retired within a single year. Building applications directly against specific versioned model IDs, without any abstraction layer, means every retirement becomes a production incident. If your team is running workloads on a multi-model platform, it is worth investing in a thin configuration layer that maps logical model names to actual model IDs, so future migrations require a config change rather than a code deployment.
The Gemini Enterprise Agent Platform’s release notes page is the authoritative source for upcoming deprecations. Keeping an eye on it regularly is a much better strategy than discovering a 400 error in production.