Grok 4 Fast Model Card

Grok 4 Fast is an efficiency-focused language model from xAI, last updated September 19, 2025. It offers reasoning capabilities "near the level of Grok 4" at lower latency and cost, with an option to skip reasoning entirely for lowest-latency applications. It is pre-trained on a general-purpose data corpus, then post-trained on tasks, tool use, and correct refusal behaviors.

The card does not report general-purpose benchmark scores (e.g., reasoning, coding, or math) or disclose context window size or supported modalities. Dual-use capability scores with safeguards removed are: WMDP Bio 85.2%, WMDP Chem 77.5%, WMDP Cyber 81.4%, BioLP-Bench 39.0%, VCT 54.5%, and CyBench unguided success rate 30.0%. The card states these scores "approach but remain below" those of Grok 4.

Evaluations cover six domains: refusals (12 harm categories), agentic abuse (AgentHarm), model hijacking (AgentDojo), deception (MASK, 1,000 questions), sycophancy (Anthropic's answer sycophancy evaluation), and dual-use knowledge (WMDP, VCT, BioLP-Bench, CyBench, MakeMeSay). Dual-use assessments are conducted with safeguards removed to measure raw capability. All evaluations were run on "a near-final release checkpoint." Results are reported separately for reasoning-enabled and non-reasoning (nr) configurations.

With the safety system prompt active, Grok 4 Fast achieves a 0.00 answer rate on refusals and user jailbreaks; the system jailbreak answer rate is 0.00 with reasoning and 0.01 without. On AgentHarm (agentic abuse, no-attack setting), the answer rate is 0.08 with reasoning and 0.10 without; AgentDojo attack success rate is 0.00 with reasoning and 0.03 without. The MASK dishonesty rate is 0.47 with reasoning and 0.63 without; political soft bias averages 0.79 and 0.89 respectively. The card states xAI "believe[s] that Grok 4 Fast overall presents a low risk for malicious use and loss of control" with deployed mitigations.

A safety system prompt prefix is injected before all API contexts, encoding the refusal policy and warnings against jailbreaks; users may append their own system messages but cannot override the safety prefix. Model-based input filters reject biological and chemical weapons queries, self-harm requests, and CSAM across all product surfaces. Safety training during post-training reduces harmful outputs and jailbreak susceptibility. Assessments of autonomous hacking, radiological, and nuclear abuse risks are described as "unchanged from that of Grok 4."

Grok 4 Fast is available via xAI's API and deployed by X Corp. on the X platform. Users may append system messages to restrict or extend model behavior, provided they do not override the safety prefix. System prompts for consumer products are published publicly at https://github.com/xai-org/grok-prompts. License terms, pricing tiers, and access controls are not disclosed in this document.

Dishonesty rates increase meaningfully in non-reasoning mode (MASK rate 0.63 vs. 0.47 with reasoning), and the deployed system prompt does not include explicit anti-deception instructions. Political soft bias scores remain elevated even with mitigations (0.79 with reasoning, 0.89 without). xAI recommends that developers requiring high truthfulness operate Grok 4 Fast with reasoning enabled and append an explicit honesty instruction. The card acknowledges that "given the limited context visible to AI models, it is often difficult to distinguish malignant intent from mere curiosity."

Grok 4 Fast is introduced as a new lower-latency, lower-cost variant of Grok 4; the card reuses the refusal evaluation methodology first introduced in the Grok 4 model card. Dual-use capabilities are reported as below Grok 4's level across all measured categories. No prior version of Grok 4 Fast exists, so no changelog delta is provided.