The Grok 4.5 Model Card is a 29-page document dated July 14, 2026, with a revision dated July 20, 2026, published by xAI under its SpaceXAI trading name (Source: https://media.x.ai/v1/website/4p5-5184fdf9.pdf). It describes Grok 4.5 as the first release of a model family developed with Cursor, given supplemental training on anonymised Cursor workflow data, with a pretraining cutoff of January 2026. Unlike xAI's 2025 cards, which reported safety evaluations only, this card states that its primary purpose is to set out the model's capabilities in quantitative terms, and most of its length is capability benchmarks presented as bar charts against Anthropic, OpenAI and Z.ai models.
Capability evaluations
The coding chapter reports DeepSWE v1.0 and v1.1, APEX-SWE, SWE-Bench Pro, SWE-Bench Multilingual, SWE-Marathon, FrontierSWE, ProgramBench, Terminal-Bench 2.1, SWE-Atlas-QnA and an internal FalseClaimBench that checks whether the agent reports work it did not perform. Engineering evaluations are EEBench (chart only), 3DCodeBench and CAD-Bench; office-work evaluations are GDPval-AA v2 and Artificial Analysis's τ³-banking, a τ-bench derivative. An R&D chapter reports the internal SpaceXAI MTS Eval of model-development tasks (chart only) and RelBench; a search chapter reports an internal hallucination rate and DeepSearchQA. Footnotes attribute many results to the benchmark owners (Datacurve, Artificial Analysis, Mercor, Cursor), and the card says competitor figures are drawn from developers' system cards or benchmark leaderboards. Peer models are labelled with their effort setting; Grok 4.5 is reported at high thinking effort throughout.
Cyber
CyberGym is reported without safeguards as a capability probe, with peer figures taken from other developers' system cards. HackerBench v0.2, an internal capture-the-flag and red-teaming suite, is reported with production safeguards as a harmful or dual-use compliance rate and a benign refusal rate, for Grok 4.5, GPT-5.5 and Claude Opus 4.8. xAI states the capability "is most useful to defenders".
Biological and chemical
Run without safeguards, the suite comprises the Virology Capabilities Test, WMDP biology and chemistry, a LAB-Bench practical multiple-choice set, ProtocolQA Open-Ended and BixBench. The card calls the results "solid but sub-threshold" and treats the suites as safety-threshold evaluations; it prints no threshold values.
Safeguards and behaviour
The card reports compliance under a jailbreak battery, general refusals in six languages, child-safety refusals, CBRN refusal accuracy on internal Autointent-Bio and Autointent-Chem suites and on FORTRESS-RN, and self-harm refusals. Behaviour evaluations are an internal Epistemic Bias rate, MASK-Rectified (a MASK variant whose grading excludes evident role-play) and an internal sycophancy measure. Refusals are described as enforced by safety fine-tuning, system prompts and, on some surfaces, runtime input and topical filters.
Provenance and confidence
Every figure is developer-assembled: some were run by named benchmark owners, some internally, and peer scores are taken from other developers' documents under settings the card does not fully describe. The DeepSearchQA chart prints the same two values as the RelBench chart, and the later Grok 4.6 Model Card (xAI, August 2026) reports different DeepSearchQA figures and several different Grok 4.5 and peer scores. Confidence is medium. The vault copy is the text layer of the official PDF; charts without printed values are not recoverable from it.
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 |
|---|---|---|---|---|
| DeepSWE v1.0 | Claude Fable 5 | 66.1% | max effort, with fallback; Pass@1; evaluation created by Datacurve, run with each model provider's harness by Artificial Analysis; reported by xAI in the Grok 4.5 model card | 2.1 |
| DeepSWE v1.0 | GPT-5.5 | 64.3% | xhigh thinking effort; Pass@1; evaluation created by Datacurve, run with each model provider's harness by Artificial Analysis; reported by xAI in the Grok 4.5 model card | 2.1 |
| DeepSWE v1.0 | Grok 4.5 | 62.0% | high thinking effort; Pass@1; evaluation created by Datacurve, run with each model provider's harness by Artificial Analysis; developer-reported | 2.1 |
| DeepSWE v1.0 | Claude Opus 4.8 | 55.8% | max effort; Pass@1; evaluation created by Datacurve, run with each model provider's harness by Artificial Analysis; reported by xAI in the Grok 4.5 model card | 2.1 |
| DeepSWE v1.0 | Claude Opus 4.7 | 40.1% | max effort; Pass@1; evaluation created by Datacurve, run with each model provider's harness by Artificial Analysis; reported by xAI in the Grok 4.5 model card | 2.1 |
| DeepSWE v1.1 | Claude Fable 5 | 70.0% | max effort, with fallback; Pass@1; mini-swe-agent harness run by Datacurve; reported by xAI in the Grok 4.5 model card | 2.2 |
| DeepSWE v1.1 | GPT-5.5 | 67.0% | xhigh thinking effort; Pass@1; mini-swe-agent harness run by Datacurve; reported by xAI in the Grok 4.5 model card | 2.2 |
| DeepSWE v1.1 | Claude Opus 4.8 | 59.0% | max effort; Pass@1; mini-swe-agent harness run by Datacurve; reported by xAI in the Grok 4.5 model card | 2.2 |
| DeepSWE v1.1 | Grok 4.5 | 53.0% | high thinking effort; Pass@1; mini-swe-agent harness run by Datacurve; developer-reported | 2.2 |
| DeepSWE v1.1 | GLM-5.2 | 44.0% | max effort; Pass@1; mini-swe-agent harness run by Datacurve; reported by xAI in the Grok 4.5 model card | 2.2 |
| APEX-SWE | Claude Fable 5 | 54.8% | max effort, with fallback; Pass@1; tests conducted by Mercor; reported by xAI in the Grok 4.5 model card | 2.3 |
| APEX-SWE | Grok 4.5 | 51.2% | high thinking effort; Pass@1; tests conducted by Mercor; developer-reported | 2.3 |
| APEX-SWE | Claude Opus 4.8 | 47.3% | high thinking effort; Pass@1; tests conducted by Mercor; reported by xAI in the Grok 4.5 model card | 2.3 |
| APEX-SWE | Claude Sonnet 5 | 43.6% | high thinking effort; Pass@1; tests conducted by Mercor; reported by xAI in the Grok 4.5 model card | 2.3 |
| APEX-SWE | GPT-5.5 | 40.8% | xhigh thinking effort; Pass@1; tests conducted by Mercor; reported by xAI in the Grok 4.5 model card | 2.3 |
| APEX-SWE | GPT-5.6 Sol | 39.7% | xhigh thinking effort; Pass@1; tests conducted by Mercor; reported by xAI in the Grok 4.5 model card | 2.3 |
| APEX-SWE | GLM-5.2 | 37.3% | max effort; Pass@1; tests conducted by Mercor; reported by xAI in the Grok 4.5 model card | 2.3 |
| SWE-Marathon | Grok 4.5 | 29.0% | high thinking effort; resolution rate, full SWE-Marathon task set, multi-layer verification; developer-reported | 2.6 |
| SWE-Marathon | Claude Opus 4.8 | 26.0% | max effort; resolution rate, full SWE-Marathon task set, multi-layer verification; reported by xAI in the Grok 4.5 model card | 2.6 |
| SWE-Marathon | Claude Fable 5 | 24.0% | max effort, with fallback; resolution rate, full SWE-Marathon task set, multi-layer verification; reported by xAI in the Grok 4.5 model card | 2.6 |
| SWE-Marathon | Claude Opus 4.7 | 16.0% | max effort; resolution rate, full SWE-Marathon task set, multi-layer verification; reported by xAI in the Grok 4.5 model card | 2.6 |
| SWE-Marathon | GLM-5.2 | 13.0% | max effort; resolution rate, full SWE-Marathon task set, multi-layer verification; reported by xAI in the Grok 4.5 model card | 2.6 |
| SWE-Marathon | GPT-5.5 | 12.0% | xhigh thinking effort; resolution rate, full SWE-Marathon task set, multi-layer verification; reported by xAI in the Grok 4.5 model card | 2.6 |
| FrontierSWE | Claude Fable 5 | 89% | max effort, with fallback; Dominance (win rate versus a random baseline across tasks), up to 20 hours per task; results and harness labels from the public FrontierSWE leaderboard; Grok 4.5 evaluated with Grok CLI; reported by xAI in the Grok 4.5 model card | 2.7 |
| FrontierSWE | Grok 4.5 | 78% | high thinking effort; Dominance (win rate versus a random baseline across tasks), up to 20 hours per task; results and harness labels from the public FrontierSWE leaderboard; Grok 4.5 evaluated with Grok CLI; developer-reported | 2.7 |
| FrontierSWE | Claude Opus 4.8 | 73% | max effort; Dominance (win rate versus a random baseline across tasks), up to 20 hours per task; results and harness labels from the public FrontierSWE leaderboard; Grok 4.5 evaluated with Grok CLI; reported by xAI in the Grok 4.5 model card | 2.7 |
| FrontierSWE | GLM-5.2 | 72% | max effort; Dominance (win rate versus a random baseline across tasks), up to 20 hours per task; results and harness labels from the public FrontierSWE leaderboard; Grok 4.5 evaluated with Grok CLI; reported by xAI in the Grok 4.5 model card | 2.7 |
| FrontierSWE | GPT-5.5 | 70% | xhigh thinking effort; Dominance (win rate versus a random baseline across tasks), up to 20 hours per task; results and harness labels from the public FrontierSWE leaderboard; Grok 4.5 evaluated with Grok CLI; reported by xAI in the Grok 4.5 model card | 2.7 |
| ProgramBench | GPT-5.5 | 60.2% | xhigh thinking effort; tests-passed rate (not full resolution); no source, decompilation or internet; reported by xAI in the Grok 4.5 model card | 2.8 |
| ProgramBench | Grok 4.5 | 57.2% | high thinking effort; tests-passed rate (not full resolution); no source, decompilation or internet; developer-reported | 2.8 |
| ProgramBench | Claude Opus 4.8 | 57.1% | xhigh thinking effort; tests-passed rate (not full resolution); no source, decompilation or internet; reported by xAI in the Grok 4.5 model card | 2.8 |
| SWE-Atlas-QnA | Grok 4.5 | 84.0% | high thinking effort; accuracy on the tracked Atlas Q&A split; developer-reported | 2.10 |
| SWE-Atlas-QnA | GPT-5.6 Sol | 84.0% | max effort; accuracy on the tracked Atlas Q&A split; reported by xAI in the Grok 4.5 model card | 2.10 |
| SWE-Atlas-QnA | Claude Fable 5 | 83.0% | max effort, with fallback; accuracy on the tracked Atlas Q&A split; reported by xAI in the Grok 4.5 model card | 2.10 |
| SWE-Atlas-QnA | Claude Opus 4.8 | 82.0% | max effort; accuracy on the tracked Atlas Q&A split; reported by xAI in the Grok 4.5 model card | 2.10 |
| SWE-Atlas-QnA | GPT-5.5 | 81.0% | xhigh thinking effort; accuracy on the tracked Atlas Q&A split; reported by xAI in the Grok 4.5 model card | 2.10 |
| SWE-Atlas-QnA | GPT-5.6 Terra | 81.0% | max effort; accuracy on the tracked Atlas Q&A split; reported by xAI in the Grok 4.5 model card | 2.10 |
| SWE-Atlas-QnA | GLM-5.2 | 74.0% | max effort; accuracy on the tracked Atlas Q&A split; reported by xAI in the Grok 4.5 model card | 2.10 |
| FalseClaimBench | Grok 4.5 | 84.0% | high thinking effort; accuracy (share of fully true claims about work performed), internal suite, higher is better; developer-reported | 2.11 |
| FalseClaimBench | Claude Opus 4.8 | 59.0% | max effort; accuracy (share of fully true claims about work performed), internal suite, higher is better; reported by xAI in the Grok 4.5 model card | 2.11 |
| FalseClaimBench | GLM-5.2 | 38.0% | max effort; accuracy (share of fully true claims about work performed), internal suite, higher is better; reported by xAI in the Grok 4.5 model card | 2.11 |
| 3DCodeBench | Grok 4.5 | 49.8% | high thinking effort; reward (executability and shape fidelity); developer-reported | 3.2 |
| 3DCodeBench | Claude Opus 4.8 | 47.0% | max effort; reward (executability and shape fidelity); reported by xAI in the Grok 4.5 model card | 3.2 |
| 3DCodeBench | Claude Fable 5 | 43.7% | max effort; reward (executability and shape fidelity); reported by xAI in the Grok 4.5 model card | 3.2 |
| 3DCodeBench | Claude Sonnet 5 | 39.2% | max effort; reward (executability and shape fidelity); reported by xAI in the Grok 4.5 model card | 3.2 |
| CAD-Bench | Claude Fable 5 | 88.5% | max effort; reward, Parametric CAD Bench (gNucleus); reported by xAI in the Grok 4.5 model card | 3.3 |
| CAD-Bench | Claude Opus 4.8 | 88.2% | max effort; reward, Parametric CAD Bench (gNucleus); reported by xAI in the Grok 4.5 model card | 3.3 |
| CAD-Bench | Grok 4.5 | 88.1% | high thinking effort; reward, Parametric CAD Bench (gNucleus); developer-reported | 3.3 |
| CAD-Bench | Claude Sonnet 5 | 86.9% | max effort; reward, Parametric CAD Bench (gNucleus); reported by xAI in the Grok 4.5 model card | 3.3 |
| CAD-Bench | GPT-5.6 Sol | 84.9% | xhigh thinking effort; reward, Parametric CAD Bench (gNucleus); reported by xAI in the Grok 4.5 model card | 3.3 |
| RelBench | GPT-5.6 Sol | 41.8% | max effort; accuracy, release-tracked RelBench aggregate score under a fixed task set and resource limits; reported by xAI in the Grok 4.5 model card | 5.2 |
| RelBench | Claude Opus 4.8 | 40.7% | max effort; accuracy, release-tracked RelBench aggregate score under a fixed task set and resource limits; reported by xAI in the Grok 4.5 model card | 5.2 |
| RelBench | Grok 4.5 | 38.4% | high thinking effort; accuracy, release-tracked RelBench aggregate score under a fixed task set and resource limits; developer-reported | 5.2 |
| Factuality | Grok 4.5 | 0.98% | high thinking effort; single-turn hallucination rate, lower is better; internal, separate grader; developer-reported | 6.1 |
| Factuality | GPT-5.5 | 1.14% | xhigh thinking effort; single-turn hallucination rate, lower is better; internal, separate grader; reported by xAI in the Grok 4.5 model card | 6.1 |
| Factuality | Claude Opus 4.8 | 3.35% | max effort; single-turn hallucination rate, lower is better; internal, separate grader; reported by xAI in the Grok 4.5 model card | 6.1 |
| DeepSearchQA | Claude Opus 4.8 | 40.7% | max effort; accuracy, internal implementation of DeepSearchQA; values duplicate the RelBench chart and are superseded by the Grok 4.6 card (see Notes); reported by xAI in the Grok 4.5 model card | 6.2 |
| DeepSearchQA | Grok 4.5 | 38.4% | high thinking effort; accuracy, internal implementation of DeepSearchQA; values duplicate the RelBench chart and are superseded by the Grok 4.6 card (see Notes); developer-reported | 6.2 |
| HackerBench v0.2 — Harmful/Dual-Use Compliance | Grok 4.5 | 7.8% | high thinking effort; share of harmful or dual-use requests complied with, lower is better; internal suite, each provider's standard safeguards; developer-reported | 7.2 |
| HackerBench v0.2 — Harmful/Dual-Use Compliance | GPT-5.5 | 25.0% | xhigh thinking effort; share of harmful or dual-use requests complied with, lower is better; internal suite, each provider's standard safeguards; reported by xAI in the Grok 4.5 model card | 7.2 |
| HackerBench v0.2 — Harmful/Dual-Use Compliance | Claude Opus 4.8 | 22.8% | max effort; share of harmful or dual-use requests complied with, lower is better; internal suite, each provider's standard safeguards; reported by xAI in the Grok 4.5 model card | 7.2 |
| HackerBench v0.2 — Benign Refusal | Grok 4.5 | 0.0% | high thinking effort; share of benign tasks refused, lower is better; internal suite, each provider's standard safeguards; developer-reported | 7.2 |
| HackerBench v0.2 — Benign Refusal | GPT-5.5 | 1.1% | xhigh thinking effort; share of benign tasks refused, lower is better; internal suite, each provider's standard safeguards; reported by xAI in the Grok 4.5 model card | 7.2 |
| HackerBench v0.2 — Benign Refusal | Claude Opus 4.8 | 0.0% | max effort; share of benign tasks refused, lower is better; internal suite, each provider's standard safeguards; reported by xAI in the Grok 4.5 model card | 7.2 |
| BixBench zero-shot MCQ | Grok 4.5 | 93.8% | high thinking effort; accuracy, zero-shot multiple choice, Grok Build default tool-on setting; developer-reported | 8.5 |
| Jailbreaks | Grok 4.5 | 0.73% | high thinking effort; compliance rate on should-refuse prompts under attack, lower is better; single- and multi-turn, user- and system-message attacks; developer-reported | 9.1 |
| General refusals | Grok 4.5 | 1.1% | high thinking effort; compliance rate on should-refuse prompts, lower is better; six languages, model-graded; developer-reported | 10.1 |
| Child safety | Grok 4.5 | 0.0% | high thinking effort; compliance rate, lower is better; CSAM and child-safety suite; developer-reported | 10.2 |
| Autointent-Bio — Bio refusal accuracy | Grok 4.5 | 97.9% | high thinking effort; refusal accuracy on dangerous queries, full safeguards enabled; internal suite; developer-reported | 10.3 |
| Autointent-Chem — Chem refusal accuracy | Grok 4.5 | 96.7% | high thinking effort; refusal accuracy on dangerous queries, full safeguards enabled; internal suite; developer-reported | 10.3 |
| FORTRESS-RN — R/N refusal accuracy | Grok 4.5 | 97.9% | high thinking effort; refusal accuracy on radiological and nuclear items, full safeguards enabled; developer-reported | 10.3 |
| Self-harm refusals | Grok 4.5 | 0.5% | high thinking effort; compliance (non-refusal / non-redirection) rate on should-refuse conversations, lower is better; developer-reported | 11.1 |
| Epistemic Bias | Grok 4.5 | 20.4% | high thinking effort; bias rate across opposing framings of the same question, lower is better; internal; developer-reported | 12.1 |
| MASK-Rectified | Grok 4.5 | 0.67% | high thinking effort; dishonesty, lower is better; MASK-derived dataset with grading corrected for evident role-play; developer-reported | 12.2 |
| Sycophancy | Grok 4.5 | 0.01% | high thinking effort; average drop in accuracy relative to the neutral prompt, lower is better; internal benchmark; developer-reported | 12.3 |
Relationships
- supports: Grok (xAI)
- instance-of: AI System Cards
- related: Grok 4.6 Model Card (xAI, August 2026), xAI, Cursor (Anysphere)