The Grok 4.6 Model Card is a 42-page document dated August 12, 2026, the day Grok 4.6 was released, and revised on August 17, 2026; it is published by xAI under its SpaceXAI trading name (Source: https://media.x.ai/v1/website/card-4p6-4cd2dc57.pdf). It describes Grok 4.6 as the latest release in a "1.5T-scale model family" developed with Cursor, with a pretraining cutoff of January 2026 and supplemental training data generated as late as June 2026. The card follows the structure of the Grok 4.5 Model Card (xAI, July 2026) and reports most results as bar charts with printed values, Grok 4.6 at high or xhigh thinking effort beside Grok 4.5 and Anthropic, OpenAI and Moonshot models at their stated effort settings.
Changelog
The August 17 revision renumbered sections, added PartBench and new DeepSearchQA results, moved KernelBenchInternal to version 1.1 on a harder task split, and "corrected eval results on HackerBench v0.2, Self-harm, MASK, LAB". The earlier values are not given.
Capability evaluations
Coding: CursorBench 3.2 (run by Cursor), APEX-SWE (Mercor), FrontierCode v1.1 extended set (Cognition), DeepSWE v1.1 (Datacurve), SWE-Marathon v1.1 (Abundant AI) and Terminal-Bench 3.0 (Harbor). Knowledge work: GDPval-AA v2 and AA-Briefcase (Artificial Analysis), APEX-Agents (Mercor), OfficeQA Pro (Databricks) and the Harvey Legal Agent Benchmark as implemented by Vals AI. Engineering: EEBench, 3DCodeBench, PartBench, CADGenBench and CADBench. R&D enablement: the internal SpaceXAI MTS Eval, InferenceEval and KernelBenchInternal v1.1; the card also recounts an earlier checkpoint working for five hours on its own inference stack and opening seven pull requests, three of which it says now serve production traffic. Search and factuality: an internal hallucination rate and DeepSearchQA. For DeepSWE the card says peer numbers "use the best result available from model cards or evaluator-reported sources". The card does not state trial counts or sampling for any benchmark.
Cyber
CyberGym, CVE-Bench and an internal SecureCodeReview evaluation are run without safeguards; CyberGym peer figures are taken from other developers' system cards. HackerBench v0.2 is run with each provider's standard safeguards and reported as harmful or dual-use compliance and benign refusal for five models. xAI states that gains over Grok 4.5 are concentrated in defence and vulnerability mitigation, and that an unrestricted configuration was given to third-party evaluators who corroborated its internal results; the evaluators are not named and no third-party figures are printed.
Biological and chemical
The suites are treated as safety-threshold evaluations under the xAI Frontier Artificial Intelligence Framework of June 30, 2026, and run without production safeguards: the Virology Capabilities Test, an internal Biosecurity VCT, the internal BioUseBench dual-use refusal rate, WMDP biology, chemistry and cyber, a LAB-Bench practical multiple-choice set, ProtocolQA Open-Ended and BixBench, each for Grok 4.5 and Grok 4.6. The card states that Grok 4.6 "scores below the FAIF safety thresholds on dual-use knowledge" and notes a limited increase over Grok 4.5 in the biological domain. Threshold values are not printed.
Safeguards and behaviour
Jailbreak robustness is reported as compliance on a standard battery, on StrongREJECT and on long-horizon multi-turn attacks including Crescendo. Output-safety sections report general refusals in six languages, child safety, CBRN refusal recall on Autointent-Bio and Autointent-Chem, and FORTRESS-RN refusal accuracy as run by Scale AI, followed by self-harm refusals, MASK-Rectified and an internal sycophancy measure. The card states that Grok 4.6 generally refuses more reliably than Grok 4.5; its tables show higher StrongREJECT, self-harm, MASK-Rectified and sycophancy rates for Grok 4.6 than for Grok 4.5.
Provenance and confidence
The figures are developer-assembled from runs by named evaluation partners, internal runs and other developers' system cards. The x.ai launch post of the same date gives 63.6% for Fable 5 on FrontierCode v1.1, a value the card assigns to Claude Opus 5, with 64.9% for Fable 5 (Source: https://x.ai/news/grok-4-6). Confidence is medium. The vault copy is the text layer of the official PDF; the HackerBench chart's value labels are positioned by bar and must be matched to the legend order.
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 |
|---|---|---|---|---|
| CursorBench 3.2 | Grok 4.6 | 70.8% | xhigh thinking effort; Cursor agent harness, fixed task set, averaging 41,136 output tokens per task; evaluation conducted by Cursor; developer-reported | 2.1 |
| CursorBench 3.2 | Grok 4.6 | 69.9% | high thinking effort; Cursor agent harness, fixed task set; evaluation conducted by Cursor; developer-reported | 2.1 |
| CursorBench 3.2 | Grok 4.5 | 66.7% | high thinking effort; Cursor agent harness, fixed task set; Grok 4.5 was served with a maximum thinking effort of high; evaluation conducted by Cursor; developer-reported | 2.1 |
| APEX-SWE | Claude Opus 5 | 63.7% | max effort; Pass@1; evaluation conducted by Mercor; reported by xAI in the Grok 4.6 model card | 2.2 |
| APEX-SWE | Claude Fable 5 | 58.8% | max effort, with fallback; Pass@1; evaluation conducted by Mercor; reported by xAI in the Grok 4.6 model card | 2.2 |
| APEX-SWE | Grok 4.6 | 56.4% | high thinking effort; Terminus-2 harness; Pass@1; evaluation conducted by Mercor; developer-reported | 2.2 |
| APEX-SWE | Grok 4.5 | 53.6% | high thinking effort; Terminus-2 harness; Pass@1; evaluation conducted by Mercor; developer-reported | 2.2 |
| APEX-SWE | Kimi K3 | 48.0% | max effort; Pass@1; evaluation conducted by Mercor; reported by xAI in the Grok 4.6 model card | 2.2 |
| APEX-SWE | Claude Sonnet 5 | 46.4% | max effort; Pass@1; evaluation conducted by Mercor; reported by xAI in the Grok 4.6 model card | 2.2 |
| APEX-SWE | GPT-5.6 Sol | 45.8% | xhigh thinking effort; Pass@1; evaluation conducted by Mercor; reported by xAI in the Grok 4.6 model card | 2.2 |
| APEX-SWE | Claude Opus 4.8 | 43.9% | max effort; Pass@1; evaluation conducted by Mercor; reported by xAI in the Grok 4.6 model card | 2.2 |
| APEX-SWE | GPT-5.5 | 37.0% | xhigh thinking effort; Pass@1; evaluation conducted by Mercor; reported by xAI in the Grok 4.6 model card | 2.2 |
| FrontierCode v1.1 — Extended (150 samples) | Claude Fable 5 | 64.9% | max effort, with fallback; extended score, graded against maintainer-authored rubrics, verifiers and hard-blocker criteria; evaluation conducted by Cognition; reported by xAI in the Grok 4.6 model card | 2.3 |
| FrontierCode v1.1 — Extended (150 samples) | Claude Opus 5 | 63.6% | max effort; extended score, graded against maintainer-authored rubrics, verifiers and hard-blocker criteria; evaluation conducted by Cognition; reported by xAI in the Grok 4.6 model card | 2.3 |
| FrontierCode v1.1 — Extended (150 samples) | Grok 4.6 | 61.3% | high thinking effort; Grok Build harness; extended score, graded against maintainer-authored rubrics, verifiers and hard-blocker criteria; evaluation conducted by Cognition; developer-reported | 2.3 |
| FrontierCode v1.1 — Extended (150 samples) | GPT-5.6 Sol | 60.6% | max effort; extended score, graded against maintainer-authored rubrics, verifiers and hard-blocker criteria; evaluation conducted by Cognition; reported by xAI in the Grok 4.6 model card | 2.3 |
| FrontierCode v1.1 — Extended (150 samples) | Claude Opus 4.8 | 59.6% | max effort; extended score, graded against maintainer-authored rubrics, verifiers and hard-blocker criteria; evaluation conducted by Cognition; reported by xAI in the Grok 4.6 model card | 2.3 |
| FrontierCode v1.1 — Extended (150 samples) | GPT-5.5 | 56.7% | xhigh thinking effort; extended score, graded against maintainer-authored rubrics, verifiers and hard-blocker criteria; evaluation conducted by Cognition; reported by xAI in the Grok 4.6 model card | 2.3 |
| FrontierCode v1.1 — Extended (150 samples) | Grok 4.5 | 56.6% | high thinking effort; Grok Build harness; extended score, graded against maintainer-authored rubrics, verifiers and hard-blocker criteria; evaluation conducted by Cognition; developer-reported | 2.3 |
| FrontierCode v1.1 — Extended (150 samples) | Claude Sonnet 5 | 56.2% | max effort; extended score, graded against maintainer-authored rubrics, verifiers and hard-blocker criteria; evaluation conducted by Cognition; reported by xAI in the Grok 4.6 model card | 2.3 |
| FrontierCode v1.1 — Extended (150 samples) | SWE-1.7 (Cognition) | 54.3% | reported without effort; extended score; evaluation conducted by Cognition; reported by xAI in the Grok 4.6 model card | 2.3 |
| DeepSWE v1.1 | Claude Opus 5 | 74.0% | max effort; best result available from model cards or evaluator-reported sources; Pass@1; evaluation conducted by Datacurve; reported by xAI in the Grok 4.6 model card | 2.4 |
| DeepSWE v1.1 | GPT-5.6 Sol | 73.0% | max effort; best result available from model cards or evaluator-reported sources; Pass@1; evaluation conducted by Datacurve; reported by xAI in the Grok 4.6 model card | 2.4 |
| DeepSWE v1.1 | Claude Fable 5 | 70.0% | max effort, with fallback; best result available from model cards or evaluator-reported sources; Pass@1; evaluation conducted by Datacurve; reported by xAI in the Grok 4.6 model card | 2.4 |
| DeepSWE v1.1 | Kimi K3 | 69.0% | max effort; best result available from model cards or evaluator-reported sources; Pass@1; evaluation conducted by Datacurve; reported by xAI in the Grok 4.6 model card | 2.4 |
| DeepSWE v1.1 | Grok 4.6 | 67.0% | xhigh thinking effort; mini-swe-agent harness; Pass@1; evaluation conducted by Datacurve; developer-reported | 2.4 |
| DeepSWE v1.1 | GPT-5.5 | 67.0% | xhigh thinking effort; best result available from model cards or evaluator-reported sources; Pass@1; evaluation conducted by Datacurve; reported by xAI in the Grok 4.6 model card | 2.4 |
| DeepSWE v1.1 | Grok 4.6 | 65.9% | high thinking effort; mini-swe-agent harness; Pass@1; evaluation conducted by Datacurve; developer-reported | 2.4 |
| DeepSWE v1.1 | Claude Opus 4.8 | 59.0% | max effort; best result available from model cards or evaluator-reported sources; Pass@1; evaluation conducted by Datacurve; reported by xAI in the Grok 4.6 model card | 2.4 |
| DeepSWE v1.1 | Grok 4.5 | 54.0% | high thinking effort; mini-swe-agent harness; Pass@1; evaluation conducted by Datacurve; developer-reported | 2.4 |
| DeepSWE v1.1 | Claude Sonnet 5 | 54.0% | max effort; best result available from model cards or evaluator-reported sources; Pass@1; evaluation conducted by Datacurve; reported by xAI in the Grok 4.6 model card | 2.4 |
| SWE-Marathon v1.1 | Claude Opus 5 | 50.0% | max effort; provider harness; resolution rate; evaluation conducted by Abundant AI; reported by xAI in the Grok 4.6 model card | 2.5 |
| SWE-Marathon v1.1 | Claude Opus 4.8 | 48.8% | max effort; provider harness; resolution rate; evaluation conducted by Abundant AI; reported by xAI in the Grok 4.6 model card | 2.5 |
| SWE-Marathon v1.1 | Kimi K3 | 48.1% | max effort; provider harness; resolution rate; evaluation conducted by Abundant AI; reported by xAI in the Grok 4.6 model card | 2.5 |
| SWE-Marathon v1.1 | Claude Fable 5 | 45.0% | max effort, with fallback; provider harness; resolution rate; evaluation conducted by Abundant AI; reported by xAI in the Grok 4.6 model card | 2.5 |
| SWE-Marathon v1.1 | GPT-5.6 Sol | 42.5% | max effort; provider harness; resolution rate; evaluation conducted by Abundant AI; reported by xAI in the Grok 4.6 model card | 2.5 |
| SWE-Marathon v1.1 | Grok 4.6 | 31.9% | high thinking effort; Grok Build harness; resolution rate; evaluation conducted by Abundant AI; developer-reported | 2.5 |
| SWE-Marathon v1.1 | Claude Sonnet 5 | 30.0% | max effort; provider harness; resolution rate; evaluation conducted by Abundant AI; reported by xAI in the Grok 4.6 model card | 2.5 |
| SWE-Marathon v1.1 | Grok 4.5 | 29.4% | high thinking effort; Grok Build harness; resolution rate; evaluation conducted by Abundant AI; developer-reported | 2.5 |
| AA-Briefcase | Claude Opus 5 | 1715 Elo | max effort; Artificial Analysis harness, offline scenarios, rubric pass rate combined with pairwise judgments; evaluation conducted by Artificial Analysis; reported by xAI in the Grok 4.6 model card | 3.2 |
| AA-Briefcase | Grok 4.6 | 1577 Elo | high thinking effort; Artificial Analysis harness, offline scenarios, rubric pass rate combined with pairwise judgments; evaluation conducted by Artificial Analysis; developer-reported | 3.2 |
| AA-Briefcase | Claude Fable 5 | 1574 Elo | max effort, with fallback; Artificial Analysis harness, offline scenarios, rubric pass rate combined with pairwise judgments; evaluation conducted by Artificial Analysis; reported by xAI in the Grok 4.6 model card | 3.2 |
| AA-Briefcase | Kimi K3 | 1541 Elo | max effort; Artificial Analysis harness, offline scenarios, rubric pass rate combined with pairwise judgments; evaluation conducted by Artificial Analysis; reported by xAI in the Grok 4.6 model card | 3.2 |
| AA-Briefcase | GPT-5.6 Sol | 1502 Elo | max effort; Artificial Analysis harness, offline scenarios, rubric pass rate combined with pairwise judgments; evaluation conducted by Artificial Analysis; reported by xAI in the Grok 4.6 model card | 3.2 |
| AA-Briefcase | Claude Sonnet 5 | 1383 Elo | max effort; Artificial Analysis harness, offline scenarios, rubric pass rate combined with pairwise judgments; evaluation conducted by Artificial Analysis; reported by xAI in the Grok 4.6 model card | 3.2 |
| AA-Briefcase | Claude Opus 4.8 | 1340 Elo | max effort; Artificial Analysis harness, offline scenarios, rubric pass rate combined with pairwise judgments; evaluation conducted by Artificial Analysis; reported by xAI in the Grok 4.6 model card | 3.2 |
| AA-Briefcase | Grok 4.5 | 1313 Elo | high thinking effort; Artificial Analysis harness, offline scenarios, rubric pass rate combined with pairwise judgments; evaluation conducted by Artificial Analysis; developer-reported | 3.2 |
| AA-Briefcase | GPT-5.5 | 1150 Elo | xhigh thinking effort; Artificial Analysis harness, offline scenarios, rubric pass rate combined with pairwise judgments; evaluation conducted by Artificial Analysis; reported by xAI in the Grok 4.6 model card | 3.2 |
| OfficeQA Pro | Grok 4.6 | 63.2% | high thinking effort; accuracy, Databricks harness; evaluation conducted by Databricks; developer-reported | 3.4 |
| OfficeQA Pro | Claude Opus 5 | 60.9% | max effort; accuracy, Databricks harness; evaluation conducted by Databricks; reported by xAI in the Grok 4.6 model card | 3.4 |
| OfficeQA Pro | Claude Fable 5 | 60.9% | max effort, with fallback; accuracy, Databricks harness; evaluation conducted by Databricks; reported by xAI in the Grok 4.6 model card | 3.4 |
| OfficeQA Pro | GPT-5.6 Sol | 60.2% | max effort; accuracy, Databricks harness; evaluation conducted by Databricks; reported by xAI in the Grok 4.6 model card | 3.4 |
| Legal Agent Benchmark (LAB) — Vals AI implementation | Grok 4.6 | 15.8% | high thinking effort; Vals AI Valkyrie harness, six tools and three skills, internet access disabled, expert all-pass rubrics; evaluation conducted by Vals AI; developer-reported | 3.5 |
| Legal Agent Benchmark (LAB) — Vals AI implementation | Grok 4.5 | 12.9% | high thinking effort; Vals AI Valkyrie harness, six tools and three skills, internet access disabled, expert all-pass rubrics; evaluation conducted by Vals AI; developer-reported | 3.5 |
| Legal Agent Benchmark (LAB) — Vals AI implementation | Claude Opus 5 | 11.7% | max effort; Vals AI Valkyrie harness, six tools and three skills, internet access disabled, expert all-pass rubrics; evaluation conducted by Vals AI; reported by xAI in the Grok 4.6 model card | 3.5 |
| Legal Agent Benchmark (LAB) — Vals AI implementation | Claude Fable 5 | 11.3% | max effort, with fallback; Vals AI Valkyrie harness, six tools and three skills, internet access disabled, expert all-pass rubrics; evaluation conducted by Vals AI; reported by xAI in the Grok 4.6 model card | 3.5 |
| Legal Agent Benchmark (LAB) — Vals AI implementation | GPT-5.6 Sol | 2.5% | max effort; Vals AI Valkyrie harness, six tools and three skills, internet access disabled, expert all-pass rubrics; evaluation conducted by Vals AI; reported by xAI in the Grok 4.6 model card | 3.5 |
| EEBench — V1 core corpus | Claude Opus 5 | 61.6% | max effort; provider harness; reward; evaluation conducted by Atopile; reported by xAI in the Grok 4.6 model card | 4.1 |
| EEBench — V1 core corpus | Grok 4.6 | 60.0% | xhigh thinking effort; Grok Build harness; reward; evaluation conducted by Atopile; developer-reported | 4.1 |
| EEBench — V1 core corpus | Claude Fable 5 | 54.2% | max effort, with fallback; provider harness; reward; evaluation conducted by Atopile; reported by xAI in the Grok 4.6 model card | 4.1 |
| EEBench — V1 core corpus | Grok 4.6 | 53.0% | high thinking effort; Grok Build harness; reward; evaluation conducted by Atopile; developer-reported | 4.1 |
| EEBench — V1 core corpus | Claude Opus 4.8 | 51.4% | max effort; provider harness; reward; evaluation conducted by Atopile; reported by xAI in the Grok 4.6 model card | 4.1 |
| EEBench — V1 core corpus | Grok 4.5 | 50.9% | high thinking effort; Grok Build harness; reward; evaluation conducted by Atopile; developer-reported | 4.1 |
| EEBench — V1 core corpus | GPT-5.5 | 42.3% | xhigh thinking effort; provider harness; reward; evaluation conducted by Atopile; reported by xAI in the Grok 4.6 model card | 4.1 |
| EEBench — V1 core corpus | GPT-5.6 Sol | 39.4% | max effort; provider harness; reward; evaluation conducted by Atopile; reported by xAI in the Grok 4.6 model card | 4.1 |
| EEBench — V1 core corpus | Kimi K3 | 38.3% | max effort; provider harness; reward; evaluation conducted by Atopile; reported by xAI in the Grok 4.6 model card | 4.1 |
| 3DCodeBench | Grok 4.6 | 54.0% | high thinking effort; Grok Build harness; reward (executability and shape fidelity); evaluator not stated; developer-reported | 4.2 |
| 3DCodeBench | Claude Opus 5 | 49.9% | max effort; reward (executability and shape fidelity); evaluator not stated; reported by xAI in the Grok 4.6 model card | 4.2 |
| 3DCodeBench | Grok 4.5 | 49.8% | high thinking effort; Grok Build harness; reward (executability and shape fidelity); evaluator not stated; developer-reported | 4.2 |
| 3DCodeBench | Claude Opus 4.8 | 47.0% | max effort; reward (executability and shape fidelity); evaluator not stated; reported by xAI in the Grok 4.6 model card | 4.2 |
| 3DCodeBench | Claude Fable 5 | 43.7% | max effort, with fallback; reward (executability and shape fidelity); evaluator not stated; reported by xAI in the Grok 4.6 model card | 4.2 |
| 3DCodeBench | Claude Sonnet 5 | 39.2% | max effort; reward (executability and shape fidelity); evaluator not stated; reported by xAI in the Grok 4.6 model card | 4.2 |
| PartBench | Claude Fable 5 | 59.9% | max effort, with fallback; provider harness; reward; evaluation conducted by Normal Factory; reported by xAI in the Grok 4.6 model card | 4.3 |
| PartBench | GPT-5.6 Sol | 59.0% | xhigh thinking effort; provider harness; reward; evaluation conducted by Normal Factory; reported by xAI in the Grok 4.6 model card | 4.3 |
| PartBench | Grok 4.6 | 58.5% | xhigh thinking effort; Grok Build harness; reward; evaluation conducted by Normal Factory; developer-reported | 4.3 |
| PartBench | Claude Opus 5 | 52.4% | max effort; provider harness; reward; evaluation conducted by Normal Factory; reported by xAI in the Grok 4.6 model card | 4.3 |
| CADGenBench — Generation split | Grok 4.6 | 40.9% | high thinking effort; Grok Build harness; reward (executability and geometric correctness); evaluation conducted by Mecado; developer-reported | 4.4 |
| CADGenBench — Generation split | GPT-5.6 Sol | 37.1% | xhigh thinking effort; provider harness; reward (executability and geometric correctness); evaluation conducted by Mecado; reported by xAI in the Grok 4.6 model card | 4.4 |
| CADGenBench — Generation split | Claude Opus 5 | 36.6% | max effort; provider harness; reward (executability and geometric correctness); evaluation conducted by Mecado; reported by xAI in the Grok 4.6 model card | 4.4 |
| CADGenBench — Generation split | Grok 4.5 | 33.2% | high thinking effort; Grok Build harness; reward (executability and geometric correctness); evaluation conducted by Mecado; developer-reported | 4.4 |
| CADBench | Claude Opus 5 | 90.6% | max effort; reward; Grok Build harness, run by gNucleus; reported by xAI in the Grok 4.6 model card | 4.5 |
| CADBench | Grok 4.6 | 88.4% | xhigh thinking effort; reward; Grok Build harness, run by gNucleus; developer-reported | 4.5 |
| CADBench | Grok 4.6 | 87.8% | high thinking effort; reward; Grok Build harness, run by gNucleus; developer-reported | 4.5 |
| CADBench | GPT-5.6 Sol | 86.5% | max effort; reward; Grok Build harness, run by gNucleus; reported by xAI in the Grok 4.6 model card | 4.5 |
| CADBench | Grok 4.5 | 83.7% | high thinking effort; reward; Grok Build harness, run by gNucleus; developer-reported | 4.5 |
| SpaceXAI MTS Eval — Standard (non-GPU) suite | Grok 4.6 | 61.1% | high thinking effort; Grok Build harness; % correct, task reward under a fixed rollout budget, internal benchmark, 29 tasks; developer-reported | 5.1 |
| SpaceXAI MTS Eval — Standard (non-GPU) suite | Grok 4.5 | 57.1% | high thinking effort; Grok Build harness; % correct, task reward under a fixed rollout budget, internal benchmark, 29 tasks; developer-reported | 5.1 |
| SpaceXAI MTS Eval — Standard (non-GPU) suite | Claude Opus 4.8 | 55.8% | max effort; % correct, task reward under a fixed rollout budget, internal benchmark, 29 tasks; reported by xAI in the Grok 4.6 model card | 5.1 |
| SpaceXAI MTS Eval — Standard (non-GPU) suite | Claude Opus 5 | 52.6% | max effort; % correct, task reward under a fixed rollout budget, internal benchmark, 29 tasks; reported by xAI in the Grok 4.6 model card | 5.1 |
| SpaceXAI MTS Eval — Standard (non-GPU) suite | GPT-5.6 Sol | 52.1% | xhigh thinking effort; % correct, internal benchmark; model refused 2 of 29 tasks and the score excludes them; reported by xAI in the Grok 4.6 model card | 5.1 |
| SpaceXAI MTS Eval — Standard (non-GPU) suite | GPT-5.5 | 46.4% | xhigh thinking effort; % correct, internal benchmark; model refused 1 task and the score excludes it; reported by xAI in the Grok 4.6 model card | 5.1 |
| SpaceXAI MTS Eval — Standard (non-GPU) suite | Grok 4.3 | 40.6% | effort not stated; % correct, internal benchmark; developer-reported | 5.1 |
| InferenceEval | Grok 4.6 | 46.9% | high thinking effort; accuracy, internal benchmark on the SpaceXAI production inference stack, Grok Build harness, hidden GPU unit score plus integration probe; developer-reported | 5.2 |
| InferenceEval | GPT-5.6 Sol | 44.1% | max effort; accuracy, internal benchmark on the SpaceXAI production inference stack, Grok Build harness, hidden GPU unit score plus integration probe; reported by xAI in the Grok 4.6 model card | 5.2 |
| InferenceEval | Claude Opus 5 | 43.0% | max effort; accuracy, internal benchmark on the SpaceXAI production inference stack, Grok Build harness, hidden GPU unit score plus integration probe; reported by xAI in the Grok 4.6 model card | 5.2 |
| InferenceEval | Grok 4.5 | 41.3% | high thinking effort; accuracy, internal benchmark on the SpaceXAI production inference stack, Grok Build harness, hidden GPU unit score plus integration probe; developer-reported | 5.2 |
| InferenceEval | Claude Opus 4.8 | 40.7% | max effort; accuracy, internal benchmark on the SpaceXAI production inference stack, Grok Build harness, hidden GPU unit score plus integration probe; reported by xAI in the Grok 4.6 model card | 5.2 |
| InferenceEval | Kimi K3 | 39.8% | max effort; accuracy, internal benchmark on the SpaceXAI production inference stack, Grok Build harness, hidden GPU unit score plus integration probe; reported by xAI in the Grok 4.6 model card | 5.2 |
| InferenceEval | GPT-5.5 | 38.9% | xhigh thinking effort; accuracy, internal benchmark on the SpaceXAI production inference stack, Grok Build harness, hidden GPU unit score plus integration probe; reported by xAI in the Grok 4.6 model card | 5.2 |
| KernelBenchInternal v1.1 | Claude Opus 5 | 48.1% | max effort; accuracy (correctness and measured speedup against a TF32 PyTorch baseline), internal suite, Grok Build harness; reported by xAI in the Grok 4.6 model card | 5.3 |
| KernelBenchInternal v1.1 | Kimi K3 | 41.0% | max effort; accuracy (correctness and measured speedup against a TF32 PyTorch baseline), internal suite, Grok Build harness; reported by xAI in the Grok 4.6 model card | 5.3 |
| KernelBenchInternal v1.1 | Claude Opus 4.8 | 37.5% | max effort; accuracy (correctness and measured speedup against a TF32 PyTorch baseline), internal suite, Grok Build harness; reported by xAI in the Grok 4.6 model card | 5.3 |
| KernelBenchInternal v1.1 | Grok 4.6 | 37.2% | high thinking effort; accuracy (correctness and measured speedup against a TF32 PyTorch baseline), internal suite, Grok Build harness; developer-reported | 5.3 |
| KernelBenchInternal v1.1 | GPT-5.6 Sol | 34.4% | max effort; accuracy (correctness and measured speedup against a TF32 PyTorch baseline), internal suite, Grok Build harness; reported by xAI in the Grok 4.6 model card | 5.3 |
| KernelBenchInternal v1.1 | GPT-5.5 | 34.0% | xhigh thinking effort; accuracy (correctness and measured speedup against a TF32 PyTorch baseline), internal suite, Grok Build harness; reported by xAI in the Grok 4.6 model card | 5.3 |
| KernelBenchInternal v1.1 | Grok 4.5 | 29.7% | high thinking effort; accuracy (correctness and measured speedup against a TF32 PyTorch baseline), internal suite, Grok Build harness; developer-reported | 5.3 |
| Factuality (Hallucination) | Claude Opus 4.8 | 3.4% | max effort; hallucination rate, lower is better; internal benchmark, separate grader, Grok Build harness; reported by xAI in the Grok 4.6 model card | 6.1 |
| Factuality (Hallucination) | Grok 4.6 | 1.7% | high thinking effort; hallucination rate, lower is better; internal benchmark, separate grader, Grok Build harness; developer-reported | 6.1 |
| Factuality (Hallucination) | GPT-5.5 | 1.1% | xhigh thinking effort; hallucination rate, lower is better; internal benchmark, separate grader, Grok Build harness; reported by xAI in the Grok 4.6 model card | 6.1 |
| Factuality (Hallucination) | Grok 4.5 | 0.98% | high thinking effort; hallucination rate, lower is better; internal benchmark, separate grader, Grok Build harness; developer-reported | 6.1 |
| DeepSearchQA | GPT-5.5 | 87.8% | xhigh thinking effort; accuracy, internal implementation of DeepSearchQA, Grok Build harness; reported by xAI in the Grok 4.6 model card | 6.2 |
| DeepSearchQA | Grok 4.5 | 85.3% | high thinking effort; accuracy, internal implementation of DeepSearchQA, Grok Build harness; developer-reported | 6.2 |
| DeepSearchQA | Claude Opus 4.8 | 84.8% | max effort; accuracy, internal implementation of DeepSearchQA, Grok Build harness; reported by xAI in the Grok 4.6 model card | 6.2 |
| DeepSearchQA | Grok 4.6 | 81.6% | high thinking effort; accuracy, internal implementation of DeepSearchQA, Grok Build harness; developer-reported | 6.2 |
| DeepSearchQA | GPT-5.6 Sol | 75.0% | high thinking effort; accuracy, internal implementation of DeepSearchQA, Grok Build harness; reported by xAI in the Grok 4.6 model card | 6.2 |
| CVE-Bench | Grok 4.5 | 35.2% | high thinking effort; reward, Grok Build harness, without safeguards; developer-reported | 7.2 |
| CVE-Bench | Grok 4.6 | 39.8% | high thinking effort; reward, Grok Build harness, without safeguards; developer-reported | 7.2 |
| SecureCodeReview | GPT-5.5 | 64.1% | xhigh thinking effort; reward, internal agentic evaluation, Grok Build harness; reported by xAI in the Grok 4.6 model card | 7.3 |
| SecureCodeReview | Grok 4.6 | 58.7% | high thinking effort; reward, internal agentic evaluation, Grok Build harness; developer-reported | 7.3 |
| SecureCodeReview | GPT-5.6 Sol | 57.5% | max effort; reward, internal agentic evaluation, Grok Build harness; reported by xAI in the Grok 4.6 model card | 7.3 |
| SecureCodeReview | Grok 4.5 | 49.4% | high thinking effort; reward, internal agentic evaluation, Grok Build harness; developer-reported | 7.3 |
| HackerBench v0.2 — Harmful/Dual-Use Compliance | Grok 4.6 | 6.9% | high thinking effort; share of harmful or dual-use requests complied with, lower is better; internal suite, run with each provider's standard safeguards; developer-reported | 7.4 |
| 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, run with each provider's standard safeguards; developer-reported | 7.4 |
| 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, run with each provider's standard safeguards; reported by xAI in the Grok 4.6 model card | 7.4 |
| 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, run with each provider's standard safeguards; reported by xAI in the Grok 4.6 model card | 7.4 |
| HackerBench v0.2 — Harmful/Dual-Use Compliance | GPT-5.6 Sol | 35.7% | max effort; share of harmful or dual-use requests complied with, lower is better; internal suite, run with each provider's standard safeguards; reported by xAI in the Grok 4.6 model card | 7.4 |
| HackerBench v0.2 — Benign Refusal | Grok 4.6 | 0.0% | high thinking effort; share of benign tasks refused, lower is better; internal suite, run with each provider's standard safeguards; developer-reported | 7.4 |
| HackerBench v0.2 — Benign Refusal | Grok 4.5 | 0.0% | high thinking effort; share of benign tasks refused, lower is better; internal suite, run with each provider's standard safeguards; developer-reported | 7.4 |
| HackerBench v0.2 — Benign Refusal | GPT-5.5 | 1.1% | xhigh thinking effort; share of benign tasks refused, lower is better; internal suite, run with each provider's standard safeguards; reported by xAI in the Grok 4.6 model card | 7.4 |
| HackerBench v0.2 — Benign Refusal | Claude Opus 4.8 | 0.0% | max effort; share of benign tasks refused, lower is better; internal suite, run with each provider's standard safeguards; reported by xAI in the Grok 4.6 model card | 7.4 |
| HackerBench v0.2 — Benign Refusal | GPT-5.6 Sol | 0.0% | max effort; share of benign tasks refused, lower is better; internal suite, run with each provider's standard safeguards; reported by xAI in the Grok 4.6 model card | 7.4 |
| Biosecurity VCT | Grok 4.5 | 44.1% | high thinking effort; overall accuracy over should-refuse and should-comply items, internal companion to VCT; developer-reported | 8.2 |
| Biosecurity VCT | Grok 4.6 | 47.8% | high thinking effort; overall accuracy over should-refuse and should-comply items, internal companion to VCT; developer-reported | 8.2 |
| BioUseBench — Dual-use refusal (severity-5 subset) | Grok 4.5 | 83.3% | high thinking effort; refusal rate on the highest-severity dual-use subset, internal suite; developer-reported | 8.3 |
| BioUseBench — Dual-use refusal (severity-5 subset) | Grok 4.6 | 90.7% | high thinking effort; refusal rate on the highest-severity dual-use subset, internal suite; developer-reported | 8.3 |
| 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 | 8.7 |
| BixBench zero-shot MCQ | Grok 4.6 | 93.8% | high thinking effort; accuracy, zero-shot multiple choice, Grok Build default tool-on setting; developer-reported | 8.7 |
| Standard jailbreaks | Grok 4.5 | 0.73% | high thinking effort; compliance on should-refuse attacks, lower is better; internal battery, production safeguards; developer-reported | 9.1 |
| Standard jailbreaks | Grok 4.6 | 0.04% | high thinking effort; compliance on should-refuse attacks, lower is better; internal battery, production safeguards; developer-reported | 9.1 |
| Long-horizon jailbreaks | Grok 4.6 | 1.0% | high thinking effort; compliance on multi-turn, Crescendo and long-horizon malicious-use attacks, lower is better; production safeguards; Grok 4.5 not reported; developer-reported | 9.1 |
| General refusals | Grok 4.5 | 1.10% | high thinking effort; compliance, lower is better; internal harmful queries in six languages, model-graded; developer-reported | 10.1 |
| General refusals | Grok 4.6 | 0.93% | high thinking effort; compliance, lower is better; internal harmful queries in six languages, model-graded; developer-reported | 10.1 |
| Child safety | Grok 4.5 | 0.00% | high thinking effort; compliance, lower is better; internal multi-turn CSAM and child-safety suite; developer-reported | 10.2 |
| Child safety | Grok 4.6 | 0.00% | high thinking effort; compliance, lower is better; internal multi-turn CSAM and child-safety suite; developer-reported | 10.2 |
| Autointent-Bio — Bio refusal recall | Grok 4.5 | 97.9% | high thinking effort; refusal recall on internal dangerous and benign biological queries; developer-reported | 10.3 |
| Autointent-Bio — Bio refusal recall | Grok 4.6 | 100.0% | high thinking effort; refusal recall on internal dangerous and benign biological queries; developer-reported | 10.3 |
| Autointent-Chem — Chem refusal recall | Grok 4.5 | 96.7% | high thinking effort; refusal recall on internal dangerous and benign chemical queries; developer-reported | 10.3 |
| Autointent-Chem — Chem refusal recall | Grok 4.6 | 100.0% | high thinking effort; refusal recall on internal dangerous and benign chemical queries; 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; evaluation conducted by Scale AI; developer-reported | 10.3 |
| FORTRESS-RN — R/N refusal accuracy | Grok 4.6 | 97.9% | high thinking effort; refusal accuracy on radiological and nuclear items; evaluation conducted by Scale AI; developer-reported | 10.3 |
| Self-harm refusals | Grok 4.5 | 0.50% | high thinking effort; compliance (failure rate, including refusing without redirecting to help), lower is better; internal suite; developer-reported | 11.1 |
| Self-harm refusals | Grok 4.6 | 0.84% | high thinking effort; compliance (failure rate, including refusing without redirecting to help), lower is better; internal suite; developer-reported | 11.1 |
| MASK-Rectified | Grok 4.5 | 0.67% | high thinking effort; dishonesty, lower is better; dataset derived from MASK with grading corrected so that evident role-play is not counted as lying; developer-reported | 12.1 |
| MASK-Rectified | Grok 4.6 | 1.90% | high thinking effort; dishonesty, lower is better; dataset derived from MASK with grading corrected so that evident role-play is not counted as lying; developer-reported | 12.1 |
| Sycophancy | Grok 4.5 | 0.01% | high thinking effort; sycophancy, lower is better; internal benchmark with misleading user-supplied context; developer-reported | 12.2 |
| Sycophancy | Grok 4.6 | 0.04% | high thinking effort; sycophancy, lower is better; internal benchmark with misleading user-supplied context; developer-reported | 12.2 |
Relationships
- supports: Grok (xAI)
- instance-of: AI System Cards
- related: Grok 4.5 Model Card (xAI, July 2026), xAI, Cursor (Anysphere)