Grok 4.7 Pricing vs 4.6: What Changes Your Task Cost?

Grok 4.7 retains Grok 4.6’s standard token rates, but that does not make a finished task cost the same. Longer reasoning, repeated context, retries and additional tool calls can change the bill even when the price per million tokens stays fixed.

The useful question is therefore two-part: what rate applies to this request, and how much usage was needed to finish the task? Prices below refer to the public xAI schedule checked on September 22, 2026, not an Ofox quote or a coding-tool subscription allowance.

Standard token prices

The official pricing table sets the long-context threshold at 200,000 prompt tokens or more. Once reached, the long-context rates apply to all tokens in that request, not only the portion above the threshold. These are pricing conditions, not two different context capacities.

The model has a 500,000-token context window. A large window is useful for fitting material into a request, but it does not make every request within that window cost the lower rate. Check the model card before estimating a long-context workload.

A small, reproducible cost example

Suppose a lower-tier request reports 40,000 uncached input tokens, 60,000 cached input tokens and 5,000 billed output tokens. The estimate is:

40,000 / 1,000,000 × $2.00 = $0.08
60,000 / 1,000,000 × $0.50 = $0.03
 5,000 / 1,000,000 × $6.00 = $0.03
Token subtotal                     = $0.14

This is a hypothetical calculation, not an observed invoice. It excludes separately billed tools and other service charges. If those 60,000 cached tokens were instead billed as ordinary input, the token subtotal would be $0.23. Use actual usage categories; do not assume that repeated text always produces a cache hit.

When a provider reports total input and cached input as a subset, subtract the cached portion before computing uncached input. Otherwise you will count the same tokens twice.

Why an agent task can cost more than expected

A task often spans several model requests. Each retry, review pass or tool response can create another turn. Record every request until the acceptance criteria pass, including failed attempts. Comparing only the last successful response hides part of the cost.

Reasoning settings also matter. Record the setting rather than comparing an unspecified 4.6 run with an unspecified 4.7 run. Artificial Analysis reports increased output-token consumption in its Grok 4.7 evaluation. That is evidence about its benchmark and settings, not a universal multiplier for your application.

A useful task log has these columns:

Fast and regional endpoints change the estimate

Fast has a separate tariff and is confined to Cursor and Grok Build. It is not a direct public API variant. At this check, the overview describes twice-standard rates, but the detailed long-context table does not follow that multiplier. Confirm the current Fast tariff for long prompts rather than multiplying the standard estimate blindly. The US regional endpoint has a separate 10% token premium. These are documented in the official overview; do not stack charges for combinations the provider does not offer.

For gateway use, check the gateway’s own tariff and promotion dates. A temporary discount at one provider does not change xAI’s list price, and neither price proves that another provider supports the same model.

Should you move from 4.6?

Keep one representative task and its acceptance tests stable, then compare complete runs. If the new model fixes a previously unsuccessful task, higher token consumption may be worthwhile. If both versions pass, compare total cost, latency and required correction rather than preferring the newer label automatically.

Use the access guide to choose the right product and the API setup guide to preserve the response information your integration needs. Our 4.6 versus 4.5 article covers the earlier version transition.