Grok’s 200k Token Threshold Can Double Your API Bill

👁 116 reads

Grok’s pricing includes a threshold that can double your bill in an instant - and we haven’t seen it covered in any English-language guide. After you cross 200,000 tokens in a request, all tokens in that request are charged at the higher rate. Not just those above the threshold. All of them.

What the Pricing Really Looks Like

The same thresholds and multipliers apply to Grok 4.6 and 4.5. In the older 4.3 and the 4.20 family, rates are lower ($1.25 and $2.50), but the mechanism is identical.

The Threshold Jump in Numbers

Let’s calculate just the input cost, without the model’s response:

Between the third and second rows there are two thousand tokens - roughly four pages of text. And a 100% price increase. This isn’t a table error - it’s how the provider describes billing in the documentation.

Why This Hurts More Than It Looks

Three reasons why this threshold is more dangerous than a simple rate increase:

1. Conversation history is appended to every request. You don’t need to send 200,000 tokens at once - it’s enough for the conversation to grow to that size. Then every subsequent message, even a single sentence, bills the entire context at double rate.
2. Reasoning tokens count too. The provider explicitly lists them among billable token types, alongside input, output, image, and cached tokens. Higher reasoning mode means more tokens, and with long context at double price.
3. Crossing the threshold is invisible. The model doesn’t warn you. You’ll only see it on your bill.

Four Things That Actually Lower Your Bill

1. Keep Requests Below the Threshold Deliberately

If your material has 230,000 tokens, splitting it into two parts of 115,000 each gives you 2 × $0.23 = $0.46 instead of $0.92 for one request. Half the price for the same content. Splitting makes sense whenever the task can be performed on parts.

2. Use Prompt Caching

Tokens served from cache cost $0.50 instead of $2.00 - they’re four times cheaper. This matters when you send the same system instruction or the same reference material with every request. The fixed part of your prompt should be at the beginning and remain unchanged, because that’s what gets cached.

3. Don’t Set Maximum Reasoning by Default

Reasoning in Grok 4.7 is configurable, and its tokens are billed like output - meaning $6, or $12 with long context. It’s the fastest way to get a bill you didn’t expect.

4. Check Which Endpoint You’re Using

This is an item almost no one writes about: requests directed to the US endpoint are billed with a 10% surcharge over global rates. If you don’t have a requirement for inference to happen in the US, this surcharge is pure waste.

When Long Context Pays Off Anyway

We don’t want to discourage large requests - there are tasks where splitting material ruins the result. This is an opinion backed by practice: tasks requiring comparison of the whole, for example finding contradictions between two distant fragments of a document, simply won’t work when performed on parts because the model never sees both fragments at once.

In such cases, the double rate is the price for a result you can’t get otherwise. It’s just worth knowing you’re paying it. More about what actually fits in the window and where the limit of usefulness lies in the article about the 500,000-token window.

Quick Test: How Much Does Your Typical Conversation Cost

1. Take your longest real request and count its characters. Divide by 4 - you’ll get an approximate token count for English text.
2. If the result is above 200,000, multiply rates by two.
3. Multiply by the number of requests per month.
4. Compare with the subscription price - $30 for SuperGrok. I’ve compiled plan details in a separate breakdown.

The four-characters-per-token converter is an approximation and will overestimate or underestimate depending on the text. It’s sufficient for budget decisions, but not for client billing.

Frequently Asked Questions

At how many tokens does Grok charge more?

From 200,000 tokens in a request. Above this threshold, all tokens in the request are billed at the long-context rate.

Is only the excess above the threshold charged at the higher rate?

No. The documentation explicitly states that the long-context rate covers all tokens in the request when the prompt reaches the threshold.

How much does the full 500,000-token window cost?

$2.00 for input alone - at the long-context rate of $4 per million.

Do cached tokens also get more expensive?

Yes, from $0.50 to $1.00 per million. The proportion doesn’t change - they’re still four times cheaper than regular input.

Can you avoid the double rate?

Yes, by splitting material into parts below the threshold. This won’t work for tasks requiring comparison of the entire material at once.

What’s this surcharge for the US endpoint?

Requests directed to the regional US endpoint are billed at a rate 10% higher than the global rate.

Kinetyra