What Grok 4.7 Doesn’t Know: May 2026 Cutoff & 5 More Gotchas

👁 117 reads

Grok 4.7 advertises real-time search, yet through the API it doesn’t know anything that happened after May 2026 by default. This isn’t a bug - it’s written plainly in the documentation - and it’s the first of six things you should know before you build something that relies on this model.

1. Knowledge Cutoff: May 2026

This is the most important number in this article. Everything that happened after that date doesn’t exist for the model until you provide it yourself - either in your prompt or through tools.

Why this is confusing: in the Grok app, it answers questions about today’s events, so you naturally assume the model simply knows. It doesn’t. The app performs searches and feeds it the results. Through the API nothing happens automatically.

The symptom you’ll recognize: the model responds with a confident tone and provides outdated information instead of saying it doesn’t know. This is behavior that’s most often mistaken for hallucination, but it’s often just missing data. I described a similar mechanism in structured output pitfalls.

What to Do About It

1. Enable the web search or X search tool if your queries concern current events.
2. Or provide fresh material in your prompt - then you pay for tokens, not tool calls.
3. Add explicitly to your system instruction: if the question concerns events after your knowledge cutoff date and you don’t have search results, say you don’t know. This doesn’t eliminate the problem but significantly reduces it.

2. Probability Analysis Fields Don’t Work - and Don’t Report Errors

In models 4.20 and newer, the logprobs and top_logprobs fields are not supported. The provider states they will be silently ignored if set.

This is the worst possible kind of incompatibility: your code works, the request goes through, the response comes back. Only the field you based your decision on is empty. If you have a layer that evaluates response confidence based on token probabilities, it will simply stop working on this model - without a single log entry.

3. Name Aliases Change the Model Under You

The provider uses three naming conventions:

The provider recommends aliases to most users so they automatically get updates. This is an opinion: in production, that’s bad advice. An alias means that one day your application will start running on a different model without your decision and without a deployment entry. If you have acceptance tests for response quality, an alias is exactly what will break them in the middle of the night.

4. Input Images Have Hard Limits

• Maximum file size: 20 MiB. A photo from a good camera without compression can exceed this limit.
• Only jpg and png. No webp, no heic, no pdf - and heic is the default photo format on iPhone.
• Number of images is not limited - that’s actually good news.
• Text and image order is flexible - description can go before or after the image.

The first two points are the most common cause of errors during initial photo analysis deployment. Converting to jpg on your application side solves both at once.

5. Not Every Model Accepts Batch Requests

Batch processing is a way to handle large volumes outside per-minute limits, with responses typically within 24 hours. Two things you need to know:

• Not every model accepts them - you need to check on the specific model’s page.
• Only some models have batch discounts. Twenty percent off for 4.3 and the 4.20 family. Grok 4.7 has no discount - batching here gives you a workaround for limits, not savings.
• Image and video generation works in batch mode but is billed at regular rates, without discount.

6. Reasoning Always Returns Encrypted

In the newer response interface, Grok 4.7 always returns encrypted reasoning content, even if you don’t request its inclusion. This is something to account for in two places: in the size of the response you’re processing, and in what you save to logs. If you log the entire response, you’re also logging that field.

Pre-Deployment Checklist

1. Do your queries concern matters after May 2026? If so - enable search or provide material.
2. Are you using logprobs? If so - this path won’t work, plan another.
3. Does your model name in configuration have a release date? If not - add it in production.
4. Are you converting images to jpg or png and checking size before sending?
5. Are you counting on batch discount? There isn’t one for 4.7.
6. Are your logs not saving encrypted reasoning?

If you’re just starting out, I described the entire path from account to first request in a separate guide.

Frequently Asked Questions

How far does Grok 4.7’s knowledge reach?

To May 2026. The provider states this date in the model documentation.

Does Grok know today’s news?

In the app yes, because the app searches for it. Through the API - only if you enable web or X search tools yourself.

Why doesn’t logprobs return data?

Because models 4.20 and newer don’t support it. The field is silently ignored without an error message.

What image formats does Grok accept?

Only jpg, jpeg and png, up to 20 MiB per file. Number of images in a request is not limited.

Which model name should I use in production?

The one with the release date. Aliases update automatically, which can change application behavior without deployment.

Is batch cheaper on Grok 4.7?

No. Twenty percent discount applies to models 4.3 and the 4.20 family. On 4.7, batch only provides a workaround for per-minute limits.

Kinetyra