Grok API pricing for every model, newest first

Each row shows xAI's own price. Input and output are per million tokens, and cached is what a token read back from the prompt cache costs, which is most of the prompt in a long chat or an agent loop.

The newest model isn't the lowest cost one xAI sells. For the example request Grok 4.7 costs $0.155 and Grok Build 0.1 costs $0.07. Over 10,000 calls that's $1,550 against $700.

Grok Build 0.1 is built for coding agents in an editor or a terminal, at $1 input and $2 output per million with a 256K window. Grok 4.20 Multi-Agent prices its tokens like any other row. But one request fans out to several agents that each read and write, so a single call can use many times the tokens of a normal one. Budget it per task, not per call.

Long prompts have their own rate. On Grok 4.7, a prompt past 200,000 tokens bills $4 input and $12 output per million, so check where your longest prompts land. The largest window belongs to Grok 4.20, at 2M tokens, against 500K on Grok 4.7.

Caching is where most of the savings are. On Grok 4.7 a cached token costs $0.50 per million, 75% off the normal input rate, and an agent that resends the same prefix every turn pays that rate for most of its input.

Start with Grok 4.7, the newest Grok, and test Grok Build 0.1 next to it. Run the same prompts through both and compare the cost per accepted answer, since a retry pays for the whole call again. For coding agents, give Grok Build 0.1 a run too, since that's the job it was trained for.

Keep the cached part of your prompt stable, because a prefix that changes every turn can't be read from cache. Reach for Grok 4.20 when your prompts don't fit in 500K tokens. Prices change, so check the table again before you commit to a large budget.