Grok API: Pricing, Keys & How to Call It

The Grok API is xAI's programmatic access to the Grok family, and the lineup turned over twice this year. As of July 2026 the served ids are Grok 4.5, the flagship since 9 July, at $2.00/$6.00 per million tokens; Grok 4.3 and the Grok 4.20 variants at $1.25/$2.50; and the coding model grok-build-0.1 at $1.00/$2.00. Grok 4, Grok 3, Grok 3 Mini and Grok Code Fast 1 are retired — if you are budgeting off their old $3/$15 or $0.20/$1.50 numbers, you are pricing models you can no longer call. Cheap per token is also not the same as cheap per job. We ran Grok 4.3 through our own executed coding harness: it cost $1.75 per 1,000 tasks, averaged 8.4 seconds, and missed one of nine tasks. This guide has the current price table, that measured result, how to get an xAI key, and how to call it (it's OpenAI-compatible), direct or via a gateway.

What the Grok API is

The Grok API is xAI's access to the Grok family, and it is OpenAI-compatible, so it drops into existing OpenAI-SDK code. What trips people up is that the lineup turned over: the served ids are now Grok 4.5 (the flagship since 9 July 2026, 500K context), Grok 4.3 (the prior flagship, still live, 1M context), the Grok 4.20 variants, and grok-build-0.1 on the separate Code API. Grok 4, Grok 3, Grok 3 Mini and Grok Code Fast 1 are retired. Two things follow. First, the famous cheap coding tier is gone — the $0.20/$1.50 that half the internet still quotes belongs to a model you cannot call. Second, sticker price is not the bill anyway: the measured section below has our own per-task numbers for Grok 4.3.

Grok API pricing

Per 1M tokens, USD. Live prices and contexts from xAI's models docs, July 2026, cross-checked against the DataLLM Lab catalog. Retired rows are the last prices those ids carried, kept for reference — they are not callable.

The line most price tables miss: xAI collapsed the ladder. The old lineup ran from $15 output down to $0.50 — a 30x internal spread that made model selection a real cost decision. The current lineup runs $6.00 down to $2.00, a 3x spread — and essentially all of that gap is the new flagship: take Grok 4.5 out and the remaining ids sit between $2.50 and $2.00, a 1.25x spread. Grok 4.3 and both 4.20 variants carry identical $1.25/$2.50 pricing, so choosing among those three is a question of behaviour — reasoning, non-reasoning, multi-agent — and nothing else. That leaves two price levers: staying on the $1.25/$2.50 tier instead of Grok 4.5 (2.4x cheaper on output), and moving coding work to grok-build-0.1 on the Code API, which saves 20% against that tier's output rate — not an order of magnitude.

Grok models by price

The useful comparison is no longer model against model — it is the old ladder against the new one. Grok's price range used to span 30x, from Grok 4's $15 output down to Grok 3 Mini's $0.50. Every one of those rungs is now retired, and what replaced them spans 3x, with three of the four current ids within 25% of each other:

This matters for anyone who built a cost model on the old lineup. A high-volume coding loop that budgeted Grok Code Fast 1 at $0.20 in / $1.50 out now has to run on grok-build-0.1 at $1.00 / $2.00. That is 33% more on output but 5x more on input, and a coding loop that resends context every turn is input-dominated, so the realistic bill increase is far closer to the 5x than the 33%. It is also not a one-line change: grok-build-0.1 sits on the separate Code API, not chat completions.

What Grok 4.3 actually cost per task

A list price is a rate, not a bill. We ran Grok 4.3 through our own executed coding harness — 13 models, nine Python tasks, temperature 0, max_tokens 4000, every returned solution run against assertions the model never sees — so the cost below is real token usage at list price, not an estimate. Grok 4.3 is the only Grok model in that run: Grok 3, Grok 4 and Grok Code Fast 1 were not tested, and we will not guess where they would land. Here is the Grok row, with three models from the same run for scale:

DataLLM Lab executed test: real token usage × list price, scaled to 1,000 tasks. Nine Python tasks, one scored attempt each, scored by execution.

Two results here are not on any spec sheet. First, Grok 4.3 was one of only three models in the run that did not score 9/9 — 10 of 13 did — and the task it missed was flatten, flattening a nested structure, not an exotic algorithm. At $1.75 per 1,000 tasks it also landed just above Claude Sonnet 5, which cost $1.67, ran faster at 7.2s, and returned working code on all nine. Of the eight models that cost less per 1,000 tasks than Grok 4.3, seven scored 9/9. Grok 4.3's list price is genuinely low for a 1M-context reasoning model; on this workload that discount did not translate into a lower bill than the models around it.

Second, where the money went. Grok 4.3 is a reasoning model and spent 482 reasoning tokens per task in this run — tokens you never read but pay for at the output rate. Claude Sonnet 5 and Qwen3 Coder Next spent 0. That is the mechanism behind the whole spread: the cheapest perfect score in the run cost $0.10 per 1,000 tasks and the most expensive cost $8.83, an 88x gap for output that passed identical tests. Anyone budgeting Grok 4.3 from the $1.25/$2.50 sticker alone will under-count the thinking tokens.

What this test does not measure. Nine short Python functions, one shot each, scored pass or fail. It says nothing about long-context reasoning, multi-file refactoring, agentic tool use, or non-Python work — and nothing at all about Grok 3, Grok 4 or Grok Code Fast 1, which we have not run. Treat it as one hard data point on cost-per-completed-task, not a verdict on the model. Full run: the 13-model coding cost benchmark. How it was measured: our methodology.

How to get a Grok (xAI) API key

1. Sign up at the xAI developer console (console.x.ai) and verify your account.
2. Add billing/credit.
3. Create an API key and copy it once; store it as export XAI_API_KEY=...

How to call the Grok API

Grok is OpenAI-compatible — use the OpenAI SDK with xAI's base URL:

from openai import OpenAI

client = OpenAI(base_url="https://api.x.ai/v1", api_key="$XAI_API_KEY")

resp = client.chat.completions.create(
    model="grok-4.3",         # or a grok-4.20 variant; grok-build-0.1 is on the Code API
    messages=[{"role": "user", "content": "Implement and test this function..."}],
)
print(resp.choices[0].message.content)

Which Grok model to use

Calling Grok through a gateway

DataLLM Lab carries Grok 4.3 plus the Grok 4.20 variants and grok-build-0.1 today — reachable with one OpenAI-compatible key alongside Claude, GPT and DeepSeek, with automatic failover. The benchmark above was not run on it — we call models through OpenRouter for testing precisely so the numbers do not depend on our own infrastructure. What the gateway gives you is the same one-key access pattern for production:

client = OpenAI(base_url="https://www.datallmlab.com/v1", api_key="$DATALLMLAB_API_KEY")
# reaches x-ai/grok-4.3, x-ai/grok-4.20, anthropic/claude-sonnet-5, ...

FAQ