xAI's new flagship exposes four reasoning effort tiers and quietly defaults to high. Any integration ported over without touching the request payload is now paying for deep reasoning on classification and routing calls that never needed it. Here is how to check, and what the new xhigh tier does to your benchmarks.
Grok 4.6 shipped this week at $2.00 per million input tokens and $6.00 per million output tokens with a 500,000 token context window, and the price is what everyone is discussing. The parameter worth your attention is in the reasoning documentation instead, because it is the one that decides what you actually pay.
xAI documents a reasoning effort parameter with four values: low, medium, high and xhigh. If your request does not specify one, the model runs at high. That default is a reasonable editorial choice by xAI, and it is the right setting for hard analytical work. It is the wrong setting for the large volume of unglamorous calls that make up most production agent traffic: intent classification, field extraction, routing decisions, short summarisation, tool argument formatting. Those calls do not benefit from extended reasoning, and under the default they are paying for it on every request.
The failure mode is quiet because nothing breaks. A team that repointed an existing integration at Grok 4.6 without editing the payload gets slightly better answers and a materially larger bill, and the two effects are difficult to separate after the fact unless somebody instrumented reasoning token usage before the switch. If you migrated in the last few days and your spend is running ahead of your projection, this is the first thing to check, before you go looking at prompt length.
The xhigh tier introduces a second, subtler trap. xAI documents xhigh as available on Grok 4.6 and later. On models that do not support it, including Grok 4.5, a request specifying xhigh is served at high rather than returning an error. Silent downgrade is good for portability and terrible for evaluation. If you are running a head to head between 4.5 and 4.6 with xhigh in both payloads, you are comparing high against xhigh and attributing the difference to the model generation. Confirm which effort level each side actually executed before you sign off on that comparison.
One more migration detail, and this one fails loudly rather than silently, which is preferable. xAI documents that presence penalty, frequency penalty and stop sequences cannot be used with its reasoning models. Requests carrying them return an error. Prompt configurations copied over from older non reasoning deployments will break outright rather than degrade, so it is worth grepping your codebase for those three parameters before you start the migration rather than during it.
Try it today
- Grep your codebase for Grok API calls and count how many set a reasoning effort value explicitly. In most codebases the answer is zero, which means every one of them is running at high.
- Classify your call sites by what the call actually does. Classification, extraction, routing, formatting and short summarisation go to low or medium. Multi step analysis, code generation and agentic tool loops stay at high.
- Set the effort value explicitly on every call site, including the ones you want at high. An explicit high is documentation; an implicit high is an accident waiting to be repeated by the next person who adds an endpoint.
- Run a one week A/B on your two highest volume call sites at low versus the current default, and score the outputs against your existing evaluation set. If quality holds, the saving compounds across every request forever.
- Before any 4.5 versus 4.6 benchmark, verify the effort level each side actually ran at rather than the one you requested, because xhigh is silently served as high on Grok 4.5.
- Grep for presencePenalty, frequencyPenalty and stop in any request path being migrated to a reasoning model, and remove them before you cut over.