Grok API Pricing 2026: xAI Per-Token Rates & Cost Calculator

List price is not the number. Floor price is.

Every LLM comparison you will read ranks providers on their published per-million rate. That is the wrong number, because most providers sell you two large discounts on top of it — prompt caching and a 50% asynchronous batch rate — and the ones that stack cleanly can land at a fraction of sticker. xAI offers one of those two levers, weakly.

"Floor" below is the best per-million rate a workload can actually reach: cached input, batch discount applied where the provider publishes one.

There is a second, non-financial implication. On providers with deep discount stacks, a lot of cost engineering is available to you after the fact — turn on caching, move a pipeline to batch, restructure a prompt to lengthen the cacheable prefix. On Grok, most of that toolbox is missing or blunted, so your cost is settled at architecture time. Estimate carefully up front, because there is less room to optimise your way out later.

Grok API price table, per 1M tokens

Where Grok's flat output ratio wins, and where it loses

Because Grok charges only 2x for output where most providers charge 4x–8x, its competitiveness swings hard on workload shape. Same request volume, same cache rate, opposite shapes.

Generation-heavy: 1,200 in → 3,000 out, 40,000/month

Long-form drafting, code synthesis, verbose reasoning. Output dominates.

Retrieval-heavy: 12,000 in → 300 out, 40,000/month

RAG answering, document Q&A, classification over long records. Input dominates.

Grok 4.3 places 2 of 5 on the generation-heavy job and 2 of 5 on the retrieval-heavy one — same model, same volume, different answer. This is the practical reason to run your own numbers in the calculator rather than reading a "cheapest LLM API" ranking: the ranking depends on a variable that only you know.

The 200,000-token threshold

The rates on this page are xAI's sub-200k band. Above 200,000 tokens in a single prompt, several Grok models charge roughly double, and the higher rate applies across the whole prompt rather than only the excess. Google does the same thing on Gemini Pro; OpenAI, Anthropic and DeepSeek do not, and charge a single flat rate at any length.

The dangerous property of a threshold like this is that it is invisible in aggregate metrics. Average prompt length can look completely stable while the tail of your distribution crosses the line and takes a growing share of your spend with it.

• Log a histogram of input token counts, not a mean. Alert on the share of requests above 200k, not on the average.
• If your prompts cluster just under the line, cap retrieval or history truncation at roughly 190k. The couple of chunks you drop are worth far less than halving the rate on every token in the request.
• Prefer two sub-threshold calls over one that crosses it, whenever the task decomposes. Same tokens, half the input rate.

X Premium and SuperGrok are not API access

This is the most common billing misunderstanding with xAI specifically, because Grok is unusually visible as a consumer product inside X. A Premium, Premium+ or SuperGrok subscription covers Grok in the X app and the Grok consumer apps. It grants no API credit whatsoever.

API access is a separate product with its own console, its own payment method and its own spend limits, billed per token at the rates above. Set a hard monthly cap in that console before you ship anything — with no batch discount and a shallow cache discount, a retry loop on Grok has fewer natural brakes on it than the same bug would on a provider whose effective rates fall as volume moves to cheaper tiers.

Grok API pricing FAQ