xAI API

What xAI actually meters

xAI meters tokens, tool calls, media and stored files, and keeps the four apart. Tokens carry one twist worth planning around: the cache discount shrinks on the newest models. Grok 4.5 serves cached input at 15% of base and Grok 4.3 at 16%, but Grok 4.6 and 4.7 at 25%. Files bill again on top, at $0.025 per GiB per day.

Server-side tools bill separately from the tokens they consume. Web search, code execution and attachment search each cost $5 per 1,000 calls, collections search $2.50. X Search charges by item rather than by call: $5 per 1,000 posts returned and $10 per 1,000 profiles. Image understanding and remote MCP tools carry no invocation fee at all.

Three modifiers sit on top. Priority Processing doubles every token type, and xAI charges it only when the response confirms "service_tier": "priority". The US regional endpoint adds 10%, putting Grok 4.7 at $2.20, $0.55 and $6.60. Batch cuts 20%, on Grok 4.3 and the three Grok 4.20 builds alone, so batched Grok 4.3 lands at $1.00 and $2.00. Caching applies before every multiplier.

What the same Grok models cost elsewhere

The same Grok models cost the same elsewhere, nearly to the cent. Microsoft's retail price API lists Grok 4.6 in East US at $0.002 per 1,000 input tokens, $0.0005 cached and $0.006 output, which is $2.00, $0.50 and $6.00 per million: xAI's own figures, long-context tier included. Azure's Data Zone deployment adds exactly 10%, the same premium xAI charges for US-pinned inference. Azure also sells provisioned throughput at $1.00 per unit-hour globally, which xAI doesn't, and stops at Grok 4.6.

Cursor publishes the same $2, $0.50 and $6 for Grok 4.7. One number differs: Cursor puts long context above 256k input tokens and xAI puts it at 200k. We're not picking a winner, because each vendor describes its own serving boundary and both are first-party to the surface they sell. The effect is real, though. A 220k-token prompt bills at double on api.x.ai and at standard rates in Cursor.

What happens when you hit
the limit

Nothing throttles, because nothing is included. You load the account with prepaid credits before the first call, and spending stops when the balance does. Enterprise accounts invoice instead.

Rate limits bite first, and xAI ties them to cumulative spend since 1 January 2026: Tier 0 at $0, then $50, $250, $1,000 and $5,000. Tiers unlock automatically and never downgrade. Limits run per model on requests per second and tokens per minute, and the per-second figure is RPM divided by 48, so a minute's budget can't go out in one burst. Grok 4.7 starts at 150 RPS and 50M tokens per minute and reaches 500 RPS and 100M at Tier 4. Exceeding either returns HTTP 429.