Grok 4 Fast

What it is

Grok 4 Fast is a cost-efficient reasoning model announced by xAI on 19 September 2025, built to reach the performance of Grok 4 at a far lower price. It ships with a context window of 2 million tokens — the amount of text the model can keep in view at once^[1]. It is available to all users, including free users, on grok.com and the iOS and Android apps, and to developers through the xAI API, OpenRouter and Vercel AI Gateway^[1].

One set of weights, two modes

Reasoning and non-reasoning behaviour are not split into two separate models here: both modes share the same weights and are steered by system prompts. The API simply exposes them under two names, grok-4-fast-reasoning and grok-4-fast-non-reasoning^[1]. In practice that means a developer can choose per request whether they want the model to work through a problem step by step before answering, or to reply directly, without switching to a different model or a different endpoint.

Efficiency and price

On comparable benchmark scores the model uses 40% fewer thinking tokens on average than Grok 4, and xAI claimed a 98% lower price to reach Grok 4's performance on frontier benchmarks^[1]. Thinking tokens are the intermediate reasoning the model generates before it produces an answer, and they are billed like any other output, so cutting them cuts the bill directly.

API pricing is $0.20 per million input tokens and $0.50 per million output tokens for prompts below 128K tokens, rising to $0.40 and $1.00 above that threshold, with cached input at $0.05 per million tokens^[1]. The tiered structure effectively charges more once a single request grows past 128K tokens, and the cached-input rate rewards repeated use of the same long prefix, such as a fixed system prompt or document.

Tools and search

xAI said the model was trained end-to-end with tool-use reinforcement learning, and that it excels at deciding when to invoke tools such as code execution or web browsing. It browses both the web and X to augment queries^[1]. Training the choice of when to call a tool, rather than hard-coding it, is what lets a single model handle a question that needs a calculation, a fresh web page, or neither.

Reported results

xAI's own pass@1 numbers — the share of problems solved when the model gets a single attempt — were GPQA Diamond 85.7%, AIME 2025 92.0%, HMMT 2025 93.3%, HLE 20.0% without tools, LiveCodeBench 80.0%, BrowseComp 44.9% and SimpleQA 95.0%^[1]. On LMArena its search variant ranked first in the Search Arena with 1163 Elo, while the base model ranked eighth in the Text Arena^[1]. Artificial Analysis independently rated it as having a state-of-the-art price-to-intelligence ratio on its Intelligence Index^[1].

Where it sits

Grok 4 Fast belongs to the part of the market that competes on cost per unit of capability rather than on raw top score: the pitch is that a user can get Grok 4-level results while paying a small fraction of the price, and can send very long inputs without hitting a context limit. The model's successor, Grok 4.1 Fast, followed on 19 November 2025, also with a 2-million-token context^[2].