xAI's Grok 4 Fast cuts reasoning costs by 98%

AI developers have long been stuck between two extremes: blazing-fast small models that hallucinate on hard problems, or snail-paced reasoning giants that drain your API budget in minutes. xAI is now attempting to kill that compromise.

XAI introduced Grok 4 Fast, a cost-efficient reasoning model designed to deliver top-tier performance with dramatic token efficiency. The company claims it matches the flagship Grok 4 across key reasoning benchmarks while consuming 40% fewer thinking tokens on average. Factoring in cheaper unit pricing, xAI says it slashes the cost of reaching frontier benchmark results by 98%, a claim verified as a top price-to-intelligence ratio by Artificial Analysis.

Unified architecture meets real-time search

Until now, switching between instant answers and deep reasoning typically required juggling completely different model checkpoints. Grok 4 Fast introduces a unified architecture where a single set of model weights handles both rapid responses and long chain-of-thought processing, controlled entirely via system prompts.

That flexibility translates to massive gains in autonomous research. Trained end-to-end with tool-use reinforcement learning, Grok 4 Fast can seamlessly scour the web and X, navigating links and parsing images and videos in real time. On LMArena's Search Arena, the model (under code name menlo) grabbed the #1 spot with an Elo score of 1163, beating OpenAI's o3-search by 17 points. In the Text Arena (code name tahoe), it landed at #8, matching the standard Grok 4 while outperforming every other model in its weight class.

Millions of tokens on a shoestring budget

Beyond raw intelligence, the technical specs aim directly at heavy enterprise workflows. Grok 4 Fast comes equipped with a massive 2M token context window.

API pricing starts at $0.20 per million input tokens and $0.50 per million output tokens for contexts under 128k, with cached inputs dropping to $0.05. For jobs exceeding 128k tokens, pricing bumps to $0.40 input and $1.00 output. Meanwhile, consumer users get the upgrade immediately: xAI opened Grok 4 Fast to all tiers on grok.com, including free accounts, across Fast and Auto modes.

Why it matters

For end users, getting unrestricted access to frontier reasoning and real-time social search at zero cost is an easy win. The lag between asking a complex current-events question and receiving a synthesized, sourced breakdown is shrinking fast.

For builders, this launch fundamentally changes the math on running autonomous agents and large-document reasoning. A 2M context window paired with a 98% cost reduction removes the financial friction of feeding full codebases or compliance manuals into reasoning loops.

We think xAI's real breakthrough here is architectural pragmatism. By unifying fast and slow modes into one weight set, it avoids the operational mess of routing between distinct networks. The risk, of course, is that giving an agile reasoning model free rein over noisy web and social data can introduce factual drift if tool triggers misfire. Still, by making high-density reasoning dirt-cheap, xAI just put serious pricing pressure on OpenAI and Anthropic.