The Grok 4 Fast Model Card is a seven-page document published by xAI and last updated on September 19, 2025 (Source: https://data.x.ai/2025-09-19-grok-4-fast-model-card.pdf). It describes Grok 4 Fast as an efficiency-focused member of the Grok family that offers reasoning near the level of Grok 4 at lower latency and cost and can skip reasoning entirely. All evaluations were run on a near-final release checkpoint, and most are reported twice, for the reasoning and non-reasoning modes.
Abuse potential
The card reports answer rates on xAI's internal refusal dataset under no attack, user-message jailbreak and system-message jailbreak, AgentHarm answer rate in the no-attack setting, and AgentDojo attack success rate. The model is deployed in the API with a fixed safety system-prompt prefix and input filters. The Grok 4.1 Model Card (xAI, November 2025) later states that the refusal results in this card covered English prompts only.
Concerning propensities
MASK dishonesty, xAI's internal soft-bias measure and Anthropic's answer-sycophancy evaluation are reported for both modes. xAI notes that non-reasoning mode raises the dishonesty rate, that the deployed system prompt carries no explicit instruction against deceptive behaviour, and that appending an instruction to answer truthfully lowers the dishonesty rate in both modes. It recommends that developers of truthfulness-sensitive applications run the model with reasoning enabled and include such an instruction.
Dual-use capabilities
With reasoning enabled and safeguards removed, the card reports BioLP-Bench, the text-only Virology Capabilities Test, WMDP biology, chemistry and cyber, CyBench unguided success and MakeMeSay. xAI states that Grok 4 Fast "approaches but remains below the dual-use capabilities of Grok 4", that its topical filters for chemical and biological weapons remain deployed, and that its assessments of autonomous hacking, radiological and nuclear risks are unchanged from Grok 4.
Provenance and confidence
All figures are developer-reported. The dual-use table prints MakeMeSay as a fraction and the other rows on a 0–100 scale without a percent sign. Confidence is medium. The vault copy is the text layer of the official PDF.
Reported results
Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.
| Evaluation | Model | Score | Setting | Section |
|---|---|---|---|---|
| Refusals — Standard (no attack) | Grok 4 Fast | 0.00 | answer rate (lower is better), internal harmful-query set, model-graded, reasoning enabled, with safety system prompt; near-final release checkpoint; developer-reported. English prompts only per the Grok 4.1 card | 2.1.2, Table 1 |
| Refusals — Standard (no attack) | Grok 4 Fast (non-reasoning) | 0.00 | answer rate, non-reasoning mode, with safety system prompt; developer-reported. English prompts only per the Grok 4.1 card | 2.1.2, Table 1 |
| Refusals — + User Jailbreak | Grok 4 Fast | 0.00 | answer rate, jailbreak in user message, reasoning enabled; developer-reported. English prompts only per the Grok 4.1 card | 2.1.2, Table 1 |
| Refusals — + User Jailbreak | Grok 4 Fast (non-reasoning) | 0.00 | answer rate, jailbreak in user message, non-reasoning; developer-reported. English prompts only per the Grok 4.1 card | 2.1.2, Table 1 |
| Refusals — + System Jailbreak | Grok 4 Fast | 0.00 | answer rate, jailbreak in system message, reasoning enabled; developer-reported. English prompts only per the Grok 4.1 card | 2.1.2, Table 1 |
| Refusals — + System Jailbreak | Grok 4 Fast (non-reasoning) | 0.01 | answer rate, jailbreak in system message, non-reasoning; developer-reported. English prompts only per the Grok 4.1 card | 2.1.2, Table 1 |
| AgentDojo | Grok 4 Fast | 0.00 | attack success rate, reasoning enabled; developer-reported | 2.1.2, Table 1 |
| AgentDojo | Grok 4 Fast (non-reasoning) | 0.03 | attack success rate, non-reasoning; developer-reported | 2.1.2, Table 1 |
| MASK | Grok 4 Fast | 0.47 | dishonesty rate, 1000 questions, reasoning enabled, deployed system prompt (no explicit honesty instruction); developer-reported | 2.2.2, Table 2 |
| MASK | Grok 4 Fast (non-reasoning) | 0.63 | dishonesty rate, non-reasoning, deployed system prompt; developer-reported | 2.2.2, Table 2 |
| MASK — With added honesty instruction | Grok 4 Fast | 0.21 | dishonesty rate, reasoning enabled, with an appended system-prompt line instructing truthful answers (not the deployed prompt); developer-reported | 2.2.2 |
| MASK — With added honesty instruction | Grok 4 Fast (non-reasoning) | 0.54 | dishonesty rate, non-reasoning, with the appended honesty line; developer-reported | 2.2.2 |
| Soft Bias (Internal) | Grok 4 Fast | 0.79 | average bias, LLM-judged paired prompts scored 0 / 0.5 / 1, reasoning enabled; developer-reported | 2.2.2, Table 2 |
| Soft Bias (Internal) | Grok 4 Fast (non-reasoning) | 0.89 | average bias, non-reasoning; developer-reported | 2.2.2, Table 2 |
| Sycophancy | Grok 4 Fast | 0.10 | sycophancy rate, Anthropic answer-sycophancy evaluation, reasoning enabled; developer-reported | 2.2.2, Table 2 |
| Sycophancy | Grok 4 Fast (non-reasoning) | 0.13 | sycophancy rate, non-reasoning; developer-reported | 2.2.2, Table 2 |
| MakeMeSay | Grok 4 Fast | 0.12 | win rate as attacker against a non-reasoning Grok 3 Mini defender, reasoning enabled, safeguards removed; developer-reported | 2.3.2, Table 3 |
| BioLP-Bench | Grok 4 Fast | 39.0 | accuracy (printed without a unit; percent scale), text-only, reasoning enabled, safeguards removed; developer-reported | 2.3.2, Table 3 |
Relationships
- supports: Grok (xAI)
- instance-of: AI System Cards
- related: Grok 4 Model Card (xAI, August 2025), Grok 4.1 Model Card (xAI, November 2025), xAI