The Grok Code Fast 1 Model Card is a six-page document published by xAI and last updated on August 26, 2025 (Source: https://data.x.ai/2025-08-26-grok-code-fast-1-model-card.pdf). It describes Grok Code Fast 1 as a fast reasoning model in the Grok family designed for coding applications run through agentic harnesses, and follows the structure of the Grok 4 Model Card (xAI, August 2025): abuse potential, concerning propensities and dual-use capabilities.
Abuse potential
The card reports answer rates on xAI's internal refusal dataset with no attack, with a user-message jailbreak and with a system-message jailbreak, an AgentHarm answer rate in the no-attack setting, and an AgentDojo attack success rate. The refusal policy lists the categories of request the model is instructed to decline and is injected as a safety system prompt ahead of any user-supplied system message.
Concerning propensities
Only deception is measured, on the MASK dataset through the API without an agentic harness. xAI reports that the dishonesty rate exceeds that of Grok 4 and attributes this in part to safety training that teaches the model to answer all queries that do not show clear intent to engage in prohibited activities. It states that, because the model is intended for agentic coding and not as a general-purpose assistant, the result does "not currently pose serious concerns".
Dual-use capabilities
Dual-use evaluations were run in an agentic harness with tools, including search, on a near-release checkpoint with safeguards removed. The card reports BioLP-Bench, the text-only portion of the Virology Capabilities Test, WMDP biology, chemistry and cyber, and CyBench unguided success. xAI states that the model does not exceed the human expert baseline on BioLP-Bench, does exceed it on VCT, and is weaker than Grok 4 in every dual-use category measured.
Provenance and confidence
All figures are developer-reported. Table 1 prints refusal answer rates as fractions and the AgentHarm and AgentDojo figures on a 0–100 scale. 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 Code Fast 1 | 0.00 | answer rate (lower is better), internal harmful-query set, model-graded, with refusal-policy system prompt; developer-reported. The Grok 4.1 card says refusal results in earlier cards covered English prompts only | 2.1.2, Table 1 |
| Refusals — + User Jailbreak | Grok Code Fast 1 | 0.00 | answer rate, jailbreak in user message; developer-reported | 2.1.2, Table 1 |
| Refusals — + System Jailbreak | Grok Code Fast 1 | 0.00 | answer rate, jailbreak in system message; developer-reported | 2.1.2, Table 1 |
| AgentDojo | Grok Code Fast 1 | 26.9 | attack success rate (percent scale, unit not printed); developer-reported | 2.1.2, Table 1 |
| MASK | Grok Code Fast 1 | 71.9 | dishonesty rate (percent scale, unit not printed), run through the API without an agentic harness; developer-reported | 2.2.2, Table 2 |
| BioLP-Bench | Grok Code Fast 1 | 19.9 | accuracy (percent scale, unit not printed), text-only, agentic harness with tools including search, safeguards removed, near-release checkpoint; developer-reported | 2.3.2, Table 3 |
| BioLP-Bench | Human experts | 38.4% | human expert performance as stated by xAI in the Grok Code Fast 1 model card | 2.3.2 |
Relationships
- supports: Grok (xAI)
- instance-of: AI System Cards
- related: Grok 4 Model Card (xAI, August 2025), Grok 4 Fast Model Card (xAI, September 2025), xAI