Model Card:
Grok 4.6
August 12, 2026

Revision: 2026-08-12
GROK 4.6 MODEL CARD

Contents

1   Introduction                                                                                         3
    1.1   Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .       3
    1.2   Model development and training . . . . . . . . . . . . . . . . . . . . . . . . . . .           4

2 Coding capabilities                                                                                    5
    2.1   CursorBench 3.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .        5
    2.2 APEX-SWE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .           7
    2.3 FrontierCode v1.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .        8
    2.4 DeepSWE v1.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .        9
    2.5 SWE-Marathon v1.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .           10
    2.6 Terminal-Bench 3.0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .          11

3 Engineering acceleration                                                                              12
    3.1   EEBench     . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   12
    3.2 3DCodeBench . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .           13
    3.4 CadGenBench - Generation            . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   14
    3.5 CadBench . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .        14

4 Office-use abilities                                                                                  15
    4.1   AA GDPVal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .       15
    4.2 AA Briefcase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .        16
    4.3 APEX-Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .         17
    4.4 Vals Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .      18
    4.8 OfficeQA Pro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .        19
    4.9 Harvey Legal Agent Benchmark . . . . . . . . . . . . . . . . . . . . . . . . . . .              19

5 R&D Enablement                                                                                        20
    5.1   SpaceXAI MTS Eval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
    5.2 InferenceEval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .       21
    5.3 KernelBenchInternal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .         22

                                                    1
GROK 4.6 MODEL CARD

6 Search capabilities and factuality                                                                 23
   6.1   Factuality (Hallucination) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
   6.2 DeepSearchQA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

7 Cyber capabilities and safeguards                                                                  24
   7.1   CyberGym . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
   7.2 CVE-Bench . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
   7.3 SecureCodeReview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
   7.4 HackerBench v0.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

8 Biological and chemical capabilities and safeguards                                                27
   8.1   Virology Capabilities Test (VCT) . . . . . . . . . . . . . . . . . . . . . . . . . . .      27
   8.2 Biosecurity VCT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .       27
   8.3 BioUseBench . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
   8.4 WMDP dual-use knowledge (MCQ) . . . . . . . . . . . . . . . . . . . . . . . . . 28
   8.5 LAB-Bench practical MCQ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
   8.6 ProtocolQA Open-Ended . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

9 Jailbreaks and robustness                                                                          29
   9.1   Jailbreaks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

10 General output safety                                                                             30
   10.1 General refusals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
   10.2 Child safety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   31
   10.4 CBRN / weapons refusals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .        31

11 Mental health                                                                                     32
   11.1 Self-harm refusals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

12 Behaviors                                                                                         32
   12.1 MASK-Rectified . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
   12.2 Sycophancy       . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

References                                                                                           34

                                                  2
GROK 4.6 MODEL CARD

1      Introduction

Grok 4.6 is the latest release in the 1.5T-scale model family of SpaceXAI* , developed in
collaboration with Cursor† . It extends Grok 4.5: stronger on coding, engineering, and
office work tasks, while also improving on new domains of work (for example, AI research
enablement or inference optimization).
It is our strongest model to date across the evaluations reported in this card.

1.1 Overview

Grok 4.6 is capable of autonomously completing longer and more challenging tasks than
any of our previous models, reaching results with fewer steps and fewer output tokens than
other frontier models.
The primary purpose of this model card is to detail the capabilities of Grok 4.6 in
quantitative and objective terms, to indicate where this model is most useful. We document
safety domains (cyber, bio knowledge, bio agentic, jailbreaks/robustness, general output
safety including CBRN refusals, mental health, and behaviors) and our safeguards below.
Each evaluation under a capability or safety section has its own subsection. Unless stated
otherwise, evaluation results are on the final deployed checkpoint of Grok 4.6.
We never silently downgrade intelligence or fall back to other models. Our goal is to
preserve legitimate uses of the model: engineering and creative work, scientific research,
hardening critical infrastructure, and AI research.
The model’s use is subject to SpaceXAI’s Acceptable Use Policy‡ , applicable Consumer
and Enterprise Terms of Service, and any applicable laws. Grok 4.6 is not intended for
autonomous high-stakes decision-making in domains such as medicine, law, finance, or
safety-critical systems without appropriate human oversight and domain-expert validation.
Grok 4.6 is primarily a text model: it accepts natural-language text and images as input and
produces text as output. It is available for use through the following channels:

    • SpaceXAI API: Available from the console at console.x.ai; API users can call Grok 4.6
      through the standard chat and completions endpoints.
    • Grok Build: The default model in SpaceXAI’s terminal-based coding agent, available
      through both the API and the CLI.
    • Cursor: Available to all users, on every plan tier.
    • Office add-ins: The default model in the add-ins for Microsoft Word, PowerPoint, and
      Excel.
*     SpaceXAI is a doing-business-as (dba) name of XAI LLC. xAI and SpaceXAI may be used interchangeably
      throughout this card.
†
      Grok 4.6 received supplemental training on anonymized Cursor workflow data to improve coding and
      agentic performance.
‡
      SpaceXAI Acceptable Use Policy: https://x.ai/legal/acceptable-use-policy.

                                                     3
GROK 4.6 MODEL CARD

• Model gateways: Reachable through OpenRouter, Vercel, Cloudflare, Snowflake,
  Databricks Mosaic, and others.

SpaceXAI plans to add Grok 4.6 to its consumer surfaces (web, mobile apps, and Grok-in-X
on the X platform) at a later date.

1.2   Model development and training

Grok 4.6 was pretrained on publicly available data, data generated internally, as well as
other data to which SpaceXAI has secured the necessary rights, followed by supplemental
training and post-training with supervised finetuning (SFT) and reinforcement learning (RL)
on human and synthetic reward signals.
Supplemental training ran longer than for Grok 4.5, combining model-generated data
curated for reasoning and advanced technical concepts, high-quality engineering corpora,
and an improved optimizer and recipe. Grok 4.5 regenerated the SFT trajectories across
reasoning efforts, agent harnesses, and domains spanning STEM, software engineering,
and knowledge work, with model-based checks screening out problematic traces. Agentic
RL covered knowledge work, general coding, and purpose-built environments for kernel
optimization, web development, and computer-aided design.
Grok 4.6 has a pretraining cutoff of January 2026.

                                             4
GROK 4.6 MODEL CARD

2    Coding capabilities

Grok 4.6 is built to be a leading model in coding capabilities, with a strong foundation in
software engineering and problem-solving.
Grok 4.6’s other abilities (in office work, R&D, and CAD) depend on a strong coding base:
an agent that reads, writes, and runs code can operate any digital tool, file, or workflow.
We primarily evaluate coding capabilities against agentic evaluations measuring real-world
software engineering and problem-solving. Our benchmarks include Cursor IDE-style
agent workflows, repository-level issue resolution, program reconstruction, terminal use,
and other software engineering-relevant use cases.

2.1 CursorBench 3.2 1

CursorBench 3.2 evaluates coding agents on realistic IDE-style tasks drawn from
production-like Cursor workflows (multi-file edits, tool use, and iterative fixing). *
                   75%
                                                              Grok 4.6
                                                 Opus 5

                   70%
                                                                                        Grok 4.5

                   65%
                                                             Fable 5
                                                          Kimi K3
                   60%
       Score (%)

                                                                                             GPT-5.6 Sol
                   55%

                   50%
                                                                          Sonnet 5
                   45%                                                                GPT-5.6 Terra

                   40%
                                                                             GPT-5.6 Luna
                   35%
                           100k        80k         60k              40k         20k          0
                                         Average output tokens per task

*   Cursor first-party agent harness (public leaderboard: https://cursor.com/cursorbench).

                                                     5
GROK 4.6 MODEL CARD

               75%
                                                                      Grok 4.6
                                               Opus 5

               70%                                      GPT-5.6 Sol
                                                                                       Grok 4.5

               65%            Fable 5

               60%
   Score (%)

                                                                Kimi K3
               55%

               50%
                                                                       Sonnet 5
               45%
                                                                          GPT-5.6 Terra

               40%
                                                                      GPT-5.6 Luna
               35%
                  $18   $15   $12         $9.0           $6.0         $3.0           $0.00
                                    Average cost per task

                                           6
GROK 4.6 MODEL CARD

2.2     APEX-SWE 2

APEX-SWE measures AI productivity on software-engineering work spanning integration
tasks and observability tasks (diagnosing failures from telemetry).*

                       Opus 5 (max)                                                         63.7%

         Fable 5 (max, with fallback)                                                  58.8%

                     Grok 4.6 (high)                                                 56.4%

                     Grok 4.5 (high)                                                53.6%

                      Kimi K3 (max)                                          48.0%

                     Sonnet 5 (max)                                         46.4%

                 GPT-5.6 Sol (xhigh)                                        45.8%

                     Opus 4.8 (max)                                        43.9%

                    GPT-5.5 (xhigh)                                37.0%

                                                              Pass@1 (%)

*     Results are taken from tests conducted by Mercor.

                                                          7
GROK 4.6 MODEL CARD

2.3 FrontierCode v1.1 3

FrontierCode v1.1 measures whether AI-generated pull requests would be mergeable by
open-source maintainers: behavioral correctness, regression safety, scope discipline, style,
and adherence to codebase conventions.*
Submissions are graded with maintainer-authored rubrics, verifiers, and hard blocker
criteria under a fixed agent scaffold.†

        Fable 5 (max, with fallback)                                                           64.9%

                      Opus 5 (max)                                                            63.6%

                    Grok 4.6 (high)                                                       61.3%

                GPT-5.6 Sol (max)                                                         60.6%

                    Opus 4.8 (max)                                                       59.6%

                   GPT-5.5 (xhigh)                                                     56.7%

                    Grok 4.5 (high)                                                    56.6%

                    Sonnet 5 (max)                                                     56.2%

                           SWE-1.7                                                    54.3%

                                                            Extended score (%)

*   Extended set (full private task suite). See https://cognition.com/frontiercode.
†
    SWE-1.7 is plotted without an effort tag. SWE-1.7 is a model from Cognition.

                                                       8
GROK 4.6 MODEL CARD

2.4 DeepSWE v1.1 4

DeepSWE v1.1* is the updated release of the DeepSWE agentic-coding benchmark,
measuring end-to-end resolution of repository issues. It serves as a proxy for the model’s
ability to handle real-world software engineering tasks, and is selected due to being
contamination-resistant.
For non-Grok models, peer model numbers use the best result available from model cards
or evaluator-reported sources.

                     Opus 5 (max)                                            74.0%

               GPT-5.6 Sol (max)                                             73.0%

       Fable 5 (max, with fallback)                                        70.0%

                    Kimi K3 (max)                                          69.0%

                  GPT-5.5 (xhigh)                                        67.0%

                   Grok 4.6 (high)                                       65.9%

                   Opus 4.8 (max)                                  59.0%

                   Grok 4.5 (high)                               54.0%

                   Sonnet 5 (max)                                54.0%

                                                    Pass@1 (%)

*   mini-swe-agent harness run by Datacurve

                                              9
GROK 4.6 MODEL CARD

2.5     SWE-Marathon v1.1 5

SWE-Marathon* targets ultra-long-horizon engineering work that can require millions of
tokens and multi-hour trajectories, far beyond a single-PR bug fix.
Scoring uses multi-layer verification designed to resist reward hacking and shortcuts.

                       Opus 5 (max)                                                     50.0%

                     Opus 4.8 (max)                                                    48.8%

                      Kimi K3 (max)                                                48.1%

         Fable 5 (max, with fallback)                                            45.0%

                 GPT-5.6 Sol (max)                                             42.5%

                     Grok 4.6 (high)                                 31.9%

                     Sonnet 5 (max)                                30.0%

                     Grok 4.5 (high)                               29.4%

                                                         Resolution rate (%)

*     Full SWE-Marathon task set: https://www.swe-marathon.org/.

                                                    10
GROK 4.6 MODEL CARD

2.6 Terminal-Bench 3.0 6

Terminal-Bench 3.0 is the successor to Terminal-Bench 2.1 (the suite also published as
FrontierBench), continuing the same terminal-agency evaluation line with an expanded
task set and refreshed harness.* It measures agents on hard, realistic command-line
tasks across software, ML, science, security, ops, hardware, and media workflows inside
containerized terminal environments.
Agents run in a fixed harness on verified tasks spanning long-running commands and
multi-step debugging.

                     Opus 5 (max)                                                         43.5%

                GPT-5.6 Sol (max)                                                 34.6%

       Fable 5 (max, with fallback)                                               34.1%

                   Grok 4.6 (high)                                       26.0%

                   Opus 4.8 (max)                                21.1%

                   Grok 4.5 (high)                       15.7%

                   Sonnet 5 (max)                    14.6%

                                                          Task success rate (%)

*   See https://www.tbench.ai/ and https://www.frontierbench.ai/.

                                                    11
GROK 4.6 MODEL CARD

3 Engineering acceleration

Beyond software engineering and office work use-cases, we evaluate Grok 4.6 on agentic
tasks that accelerate physical-world engineering: electrical and chip design, procedural
3D modeling via code, parametric CAD and part generation, and related workflows where
models must reason about geometry, materials, and real devices rather than repositories
alone. We strongly believe that this is one of the next critical frontiers in agent capabilities,
the below evaluations are selected to measure a core set of capabilities in this domain.

3.1 EEBench 7

EEBench is an electrical engineering and chip design benchmark: models design circuits
and other hardware that are graded for physical correctness and functionality.*

                      Opus 5 (max)                                                             61.6%

                   Grok 4.6 (xhigh)                                                          60.0%

        Fable 5 (max, with fallback)                                                   54.2%

                    Grok 4.6 (high)                                                   53.0%

                    Opus 4.8 (max)                                                   51.4%

                    Grok 4.5 (high)                                                  50.9%

                   GPT-5.5 (xhigh)                                           42.3%

                GPT-5.6 Sol (max)                                      39.4%

                     Kimi K3 (max)                                     38.3%

                                                               Reward (%)

*   Public leaderboard results from https://eebench.org/ (V1 core corpus).

                                                     12
GROK 4.6 MODEL CARD

3.2 3DCodeBench 8

3DCodeBench evaluates agentic procedural 3D modeling via code. An agent is tasked with
authoring engine-ready 3D assets through software APIs and geometric reasoning, and is
graded for executability and shape fidelity.*

                   Grok 4.6 (high)                                            54.0%

                     Opus 5 (max)                                       49.9%

                   Grok 4.5 (high)                                      49.8%

                   Opus 4.8 (max)                                     47.0%

       Fable 5 (max, with fallback)                              43.7%

                   Sonnet 5 (max)                             39.2%

                                                 Reward (%)

*   https://arxiv.org/abs/2606.01057.

                                          13
GROK 4.6 MODEL CARD

3.4 CadGenBench - Generation 9

CadGenBench measures generation of CAD models from design prompts under a fixed
agent harness, graded for executability and geometric correctness.

          Grok 4.6 (high)                                                   40.9%

      GPT-5.6 Sol (xhigh)                                               37.1%

            Opus 5 (max)                                               36.6%

          Grok 4.5 (high)                                           33.2%

                                               Reward (%)

3.5 CadBench 10

CadBench evaluates agents on parametric CAD design and 3D modeling tasks, including
out-of-distribution generation.

          Opus 5 (max)                                                      90.6%

         Grok 4.6 (high)                                    87.8%

      GPT-5.6 Sol (max)                         86.5%

         Grok 4.5 (high)        83.7%

                                               Reward (%)

                                          14
GROK 4.6 MODEL CARD

4 Office-use abilities

We evaluate performance on professional knowledge-work and structured agent tasks
relevant to office and productivity use cases.

4.1 AA GDPVal 11

AA GDPVal (GDPval-AA v2)* evaluates models on economically valuable knowledge-work
deliverables (documents, analyses, and professional artifacts) spanning occupations that
contribute substantially to GDP.
Artificial Analysis runs GDPval-style tasks with pairwise quality ratings under its GDPval-AA
v2 harness.

                      Opus 5 (max)                                                         1849

                     Grok 4.6 (high)                                                  1753

        Fable 5 (max, with fallback)                                                  1741

                 GPT-5.6 Sol (max)                                                   1728

                      Kimi K3 (max)                                                 1682

                     Sonnet 5 (max)                                            1601

                    Opus 4.8 (max)                                             1588

               GPT-5.6 Terra (max)                                             1578

                     Grok 4.5 (high)                                         1526

                    GPT-5.5 (xhigh)                                          1490

                           Grok 4.3                                1085

                                                             AA GDPVal Elo

*   Run by Artificial Analysis (AA) in their harness.

                                                        15
GROK 4.6 MODEL CARD

4.2 AA Briefcase 12

AA Briefcase (Artificial Analysis) evaluates frontier agents on long-horizon, multi-week
professional knowledge-work projects that produce deliverables such as spreadsheets,
presentations, memos, financial models, and PDFs.*
Scenarios are expert-built, multi-file, and offline (no internet). Grading combines rubric
pass rate with pairwise analytical-quality and presentation judgments.

                      Opus 5 (max)                                                              1715

                    Grok 4.6 (high)                                                       1577

        Fable 5 (max, with fallback)                                                     1574

                     Kimi K3 (max)                                                       1541

                 GPT-5.6 Sol (max)                                                     1502

                    Sonnet 5 (max)                                                1383

                    Opus 4.8 (max)                                              1340

                    Grok 4.5 (high)                                             1313

                   GPT-5.5 (xhigh)                                       1150

                                                             AA-Briefcase Elo

*   https://artificialanalysis.ai/evaluations/aa-briefcase

                                                        16
GROK 4.6 MODEL CARD

4.3 APEX-Agents 13

APEX-Agents measures whether frontier agents can autonomously complete long-horizon,
cross-application professional tasks in realistic simulated work environments spanning
investment banking, management consulting, and corporate law.*
Agents operate over multi-app projects (documents, spreadsheets, presentations, email,
chat, files) and are graded with expert binary rubrics: a task fully succeeds only if all criteria
are met.

                     Opus 5 (max)                                                  60.6%

       Fable 5 (max, with fallback)                                               59.2%

                   Grok 4.6 (high)                                               57.5%

               GPT-5.6 Sol (max)                                                56.7%

                   Opus 4.8 (max)                                               56.2%

                  GPT-5.5 (xhigh)                                               55.5%

                    Kimi K3 (max)                                               55.4%

                   Grok 4.5 (high)                                      47.1%

                   Sonnet 5 (high)                           32.5%

                                                           Pass@1 (%)

*   https://www.mercor.com/apex/apex-agents-leaderboard/

                                                17
GROK 4.6 MODEL CARD

4.4 Vals Index 14

Vals Index (Vals AI) is an independent composite of real-world industry and agentic
evaluations spanning finance, legal, healthcare, and coding-adjacent professional work.

      Fable 5 (max, with fallback)                                           75.1%

                    Opus 5 (max)                                             74.8%

                   Kimi K3 (max)                                             74.7%

              GPT-5.6 Sol (max)                                           73.1%

                  Grok 4.6 (high)                                        71.1%

                  Opus 4.8 (max)                                         70.4%

                  Sonnet 5 (max)                                       68.6%

                 GPT-5.5 (xhigh)                                       68.0%

                  Grok 4.5 (high)                                    65.3%

            GPT-5.6 Terra (max)                                      65.1%

                                                  Suite score (%)

                                            18
GROK 4.6 MODEL CARD

4.8 OfficeQA Pro 15

OfficeQA Pro evaluates models on professional office question-answering over realistic
workplace documents and workflows.

                  Grok 4.6 (high)                                            63.2%

                    Opus 5 (max)                                           60.9%

      Fable 5 (max, with fallback)                                         60.9%

              GPT-5.6 Sol (max)                                            60.2%

                                                    Accuracy (%)

4.9 Harvey Legal Agent Benchmark 16

Harvey Legal Agent Benchmark (LAB) is a long-horizon legal-agent evaluation: agents
complete realistic, multi-step legal work over client-matter files and produce reviewable
work product graded with expert all-pass rubrics. Scores reported here use the Vals AI
implementation of the benchmark.

                  Grok 4.6 (high)                                            22.0%

      Fable 5 (max, with fallback)                             14.2%

                  Grok 4.5 (high)                         12.9%

                    Opus 5 (max)                       11.7%

              GPT-5.6 Sol (max)      2.5%

                                                       Score

                                             19
GROK 4.6 MODEL CARD

5    R&D Enablement

We evaluate Grok 4.6 on its ability to automate parts of the engineering and research
process for training and evaluating new versions of itself.
As a test of autonomous ML acceleration, an earlier Grok 4.6 checkpoint was tasked
with speeding up its own chat inference, free to experiment but required to verify
end-to-end gains before opening a pull request. In five hours it worked through 297
candidate optimizations across fused MoE, FMHA, kernel scheduling, and communication,
discarded those without measurable end-to-end gain (including several that passed
microbenchmarks), and opened seven pull requests; three now serve Grok Chat production
traffic, for a combined 1.5% decode and 3.1% prefill throughput gain.

5.1 SpaceXAI MTS Eval

The SpaceXAI MTS eval is an internal benchmark of frontier model-development tasks
faced by SpaceXAI engineers: diagnosing reward hacking in training runs, generating
and auditing training data, debugging large-scale training infrastructure, and creating
evaluations for emerging capabilities. This section covers the standard (non-GPU) suite.
Agents are scored on task reward under a fixed rollout budget (time and tokens). Refusals
are noted.* †

            Grok 4.6 (high)                                                              61.1%

            Grok 4.5 (high)                                                          57.1%

            Opus 4.8 (max)                                                          55.8%

              Opus 5 (max)                                                       52.6%

       GPT-5.6 Sol (xhigh)                                                      52.1%

           GPT-5.5 (xhigh)                                                 46.4%

                   Grok 4.3                                            40.6%

                                                           % Correct

*   GPT-5.6 Sol refused 2 of the 29 tasks and GPT-5.5 refused 1; scores for those models exclude the refused
    tasks.
†
    Results were obtained with the Grok Build harness.

                                                     20
GROK 4.6 MODEL CARD

5.2   InferenceEval

InferenceEval asks agents to implement real production changes in the code that actually
runs chat inference. Each task starts from a frozen pre-change image of the internal
inference stack; the agent must land a working patch under the Grok Build harness. Tasks
are not a single bug class: some add missing kernels or APIs, some fix scheduler or replay
correctness, and some implement fusions, layout, or quantization work that must preserve
bit-exact tensors.
Scoring combines a hidden GPU unit score (exact tensors, state, and CUDA-graph
behavior) with a hidden integration probe that must pass or the unit score is zeroed, plus a
check that the patch implements the stated production contract. These components are
combined into the reported accuracy.

          Grok 4.6 (high)                                                     46.9%

       GPT-5.6 Sol (max)                                                   44.1%

           Opus 5 (max)                                                 43.0%

          Grok 4.5 (high)                                             41.3%

         Opus 4.8 (max)                                              40.7%

           Kimi K3 (max)                                            39.8%

         GPT-5.5 (xhigh)                                           38.9%

                                                  Accuracy (%)

                                             21
GROK 4.6 MODEL CARD

5.3 KernelBenchInternal

KernelBenchInternal is an internal acceleration suite inspired by KernelBench17 and related
verified kernel-writing benchmarks. Each task provides a PyTorch reference model and
asks the agent to replace operators with custom CUDA for a faster implementation of that
operator class. Sampling and grading run in a sandboxed GPU environment with standard
coding tools, on the same class of accelerators used for model training.
Submitted kernels must compile, match the reference on a correctness trial, and then be
timed against a TF32 PyTorch baseline. The reported score depends on both correctness
and the size of the measured speedup.

      GPT-5.6 Sol (max)                                                            73.1%

           Opus 5 (max)                                                            73.0%

         Grok 4.6 (high)                                                         70.6%

          Kimi K3 (max)                                                         68.5%

         Opus 4.8 (max)                                                         67.6%

         Grok 4.5 (high)                                                58.4%

                                            Accuracy / efficiency (%)

                                             22
GROK 4.6 MODEL CARD

6 Search capabilities and factuality

We evaluate grounded answering with search tools and factual reliability, including
hallucination-prone settings.

6.1 Factuality (Hallucination)

Factuality (Hallucination) measures how often the model asserts unsupported or fabricated
claims in a single response to an information-seeking query. Lower is better.
A separate grader flags factually unsupported statements against retrieved or known
ground truth.

          Opus 4.8 (max)                                                                       3.4%

          Grok 4.6 (high)                                   1.7%

         GPT-5.5 (xhigh)                       1.1%

          Grok 4.5 (high)                    0.98%

                                                      Hallucination rate (%)

6.2     DeepSearchQA 18

DeepSearchQA evaluates end-to-end answer accuracy on questions that require
multi-step search and synthesis of retrieved information. Answers are graded for
correctness against reference answers.*

            Opus 4.8 (max)                                                                     84.8%

             Grok 4.6 (high)                                                              81.6%

         GPT-5.6 Sol (max)                                                             75.0%

            GPT-5.5 (xhigh)                                                    63.7%

             Grok 4.5 (high)                               38.4%

                                                            Accuracy (%)

*     Evaluation was run on an internal implementation of DeepSearchQA.

                                                      23
GROK 4.6 MODEL CARD

7   Cyber capabilities and safeguards

Grok 4.6 exhibits enhanced cybersecurity-relevant capabilities, with minor cyber-defense
and vulnerability-mitigation capabilities improvements compared to Grok 4.5. Capability
and safeguard behavior are measured separately, because the deployed configuration
refuses the large majority of clearly harmful cyber requests.
The capability is most useful to defenders, for finding and fixing vulnerabilities rather than
carrying out end-to-end attacks.
We additionally supplied the model without restrictions to third-party evaluators to validate
the above results. The results of testing validated the results of our evaluation and internal
testing; demonstrating that Grok 4.6 had elevated cyber knowledge and capabilities
compared to Grok 4.5, but that this enhanced capability was most useful expressive in
cyber-defense tasks.
Testing performed by third-party evaluators validate the above results;
Cyber evaluations measure offensive and defensive security-relevant capabilities, plus
calibration of cyber safety controls.

7.1 CyberGym 19

CyberGym measures offensive security capability by tasking an agent with reproducing
crashes and assembling working exploits for known vulnerabilities.
Because it is a capability probe rather than a refusal test, scores use the unrestricted
(i.e. unmitigated by our standard safeguards) setting as a pure measure of ability; refusal
behavior on harmful cyber requests is measured separately (see HackerBench). Other
models’ results are taken from the respective model providers’ system cards20,21,22,23 ,
using the unsafeguarded figures where available to avoid loss of signal due to refusals.

       GPT-5.6 Sol (max)                                                      83.6%

         GPT-5.5 (xhigh)                                                     81.8%

          Grok 4.6 (high)                                                   79.7%

          Grok 4.5 (high)                                                  79.0%

         Opus 4.8 (max)                                                    78.1%

                                               Mean Reproduced (%)

                                              24
GROK 4.6 MODEL CARD

7.2   CVE-Bench 24

CVE-Bench evaluates agents on exploiting real-world web-application CVEs in sandboxed
environments that mimic production services. The model is tested without safeguards in
place, in order to measure its cyber capabilities accurately.

                            Metric   Grok 4.5 (high)   Grok 4.6 (high)
                            Reward           35.2%               39.8%

7.3 SecureCodeReview

SecureCodeReview is an internal agentic software-engineering eval that scores both
whether the agent fixes the security issues present in a realistic codebase and whether its
patch avoids merging new security issues under a fixed review harness.

         GPT-5.5 (xhigh)                                                           64.1%

          Grok 4.6 (high)                                                  58.7%

       GPT-5.6 Sol (max)                                                   57.5%

          Grok 4.5 (high)                                          49.4%

                                                    Reward (%)

                                               25
GROK 4.6 MODEL CARD

7.4 HackerBench v0.2

HackerBench is a SpaceXAI internal CTF-style suite of cybersecurity and red-teaming
tasks that require chaining skills (recon, exploitation primitives, and reasoning about
vulnerable systems) in a constrained environment. HackerBench also includes borderline,
dual-use, and harmful-intent requests that should be refused, plus benign tasks that
should be completed.
Runs use the standard release-tracked safeguards for Grok 4.6 and the standard
provider-supplied safeguards for other tested models.

         Grok 4.6 (high)       Grok 4.5 (high)          GPT-5.5 (xhigh)       Opus 4.8 (max)       GPT-5.6 Sol (max)

                                                     35.7%

                                     25.0%
                                             22.8%

                      16.7%

                              7.8%

                                                                      0.8%          1.1%
                                                                             0.0%          0.0% 0.0%
                     Harmful/Dual-Use Compliance                             Benign Refusal

                                                             26
GROK 4.6 MODEL CARD

8      Biological and chemical capabilities and safeguards

These suites probe dual-use biology and chemistry knowledge and lab-protocol
understanding, and are treated as safety-threshold evaluations under the xAI Frontier
Artificial Intelligence Framework (FAIF)* , which defines the capability thresholds and
associated risk domains.
Grok 4.6 scores below the FAIF safety thresholds on dual-use knowledge, indicating limited
actionable uplift for an already-trained actor. It also demonstrates no appreciable lift in
dual-use capabilities compared to Grok 4.5. Weapons-path refusals appear separately
under general output safety.
Unless otherwise stated, the benchmarks presented in §8 are performed without any
safeguards in place; this enables full measurement of Grok 4.6’s dual-use capabilities when
unrestricted.

8.1 Virology Capabilities Test (VCT) 25

VCT presents multimodal virology problems and measures whether the model can
troubleshoot experimental lab protocols. It is a dual-use knowledge probe, treated as a
CBRN safety-threshold evaluation, so a higher score indicates more dual-use capability
rather than a better outcome.

                              Metric         Grok 4.5 (high)       Grok 4.6 (high)
                              Accuracy                  65.5%                 67.4%

8.2     Biosecurity VCT 25

Biosecurity VCT is an internal biosecurity companion to VCT: dual-use troubleshooting
items scored for overall accuracy (should-refuse and should-comply items together) under
the same CBRN safety-threshold framing used for VCT knowledge probes.

                              Metric         Grok 4.5 (high)       Grok 4.6 (high)
                              Accuracy                        —               47.8%

*     xAI Frontier Artificial Intelligence Framework (June 30, 2026): https://media.x.ai/v1/website/xai-frontier-
      artificial-intelligence-framework-30-june-2026-99c40684.pdf.

                                                         27
GROK 4.6 MODEL CARD

8.3 BioUseBench

BioUseBench is an internal high-elo dual-use biology suite of biological enablement queries
of varying difficulty and risk. We report the refusal rate on the most borderline dual-use
split.

                Metric                    Grok 4.5 (high)   Grok 4.6 (high)
                Severity-5 refusal rate                —            90.7%

8.4 WMDP dual-use knowledge (MCQ) 26

WMDP-Bio, WMDP-Chem, and WMDP-Cyber are multiple-choice suites probing
operationally sensitive dual-use knowledge in biology, chemistry, and cyber.

                Metric                    Grok 4.5 (high)   Grok 4.6 (high)
                WMDP-Bio accuracy                  90.9%             90.0%
                WMDP-Chem accuracy                 87.3%             85.3%
                WMDP-Cyber accuracy                    —             90.1%

                                             28
GROK 4.6 MODEL CARD

8.5 LAB-Bench practical MCQ 27

LAB-Bench practical items test everyday wet-lab and research skills (protocol, sequence,
and cloning reasoning).

                       Metric      Grok 4.5 (high)    Grok 4.6 (high)
                       Accuracy              71.1%             80.7%

The following suites evaluate open-ended and tool-using biology tasks rather than
closed knowledge questions. They serve as non-malicious capability indicators: strong
performance signals broad scientific competence, not weapons-enabling assistance, which
remains gated by the refusal safeguards (§10.4).

8.6 ProtocolQA Open-Ended 27

ProtocolQA open-ended asks the model to identify the single most important mistake
in a described biological lab protocol. Unlike the multiple-choice suites, this benchmark
comprises open-ended troubleshooting tasks.

                       Metric      Grok 4.5 (high)    Grok 4.6 (high)
                       Accuracy             87.0%              79.6%

9 Jailbreaks and robustness

We test whether the refusal safeguards hold under adversarial pressure.
We treat residual failures as cases for ongoing monitoring and patching as new techniques
emerge.

9.1 Jailbreaks

We stress the refusal safeguards with a broad, continuously updated set of jailbreak
attacks, including published and internally-discovered techniques28,29 , across single-turn
and multi-turn settings, as well as longer adversarial trajectories.
Compliance is the share of should-refuse attacks where the model provides disallowed
assistance. We report:

• Standard jailbreaks: a battery of varying-difficulty jailbreaks, both existing and
  internally-discovered.
• StrongReject: compliance on the StrongREJECT suite.
• Long-horizon jailbreaks: multiturn jailbreaks, Crescendo multiturn jailbreaks, and
  malicious-use long horizon.

                                             29
GROK 4.6 MODEL CARD

Lower is better.

                   Suite                     Grok 4.5 (high)     Grok 4.6 (high)
                   Standard jailbreaks                  0.73%             0.04%
                   StrongReject                             —              3.9%
                   Long-horizon jailbreaks                  —               1.0%

10 General output safety

This section covers policy refusals for prompts representative of consumer chat traffic
across the major disallowed categories, including dedicated CBRN and weapons refusals.
On these suites, Grok 4.6 generally refuses more reliably than Grok 4.5: lower compliance
on the broad disallowed set, perfect CBRN autointent recall, and unchanged 0.0% CSAM
compliance.
The evaluations and results below also apply to Grok Build.
Refusals are enforced by a layered, defense-in-depth stack rather than any single filter.
Safety fine-tuning and post-training (supervised fine-tuning plus reinforcement learning
from human feedback, verifiable rewards, and model-based grading) train the model to
refuse requests that show clear intent to cause severe harm or engage in criminal activity.
System prompts steer the model toward honesty and truth-seeking while avoiding
over-refusal on benign or hypothetical discussions.
Depending on deployment surface, Grok 4.6 is additionally augmented with runtime input
and topical filters that add controls for classes of severe harm, including CSAM, self-harm,
and biological/chemical weapons pathways, alongside cyber-specific input safety controls.

10.1   General refusals

We assemble harmful queries whose intent ranges from overt to well-hidden, spanning the
major policy categories, in English and translated into Spanish, Chinese, Japanese, Arabic,
and Russian. A separate grader model judges whether the model correctly refuses.
Lower is better.

                           Metric       Grok 4.5 (high)     Grok 4.6 (high)
                           Compliance                1.1%           0.93%

                                                30
GROK 4.6 MODEL CARD

10.2 Child safety

The CSAM and child safety benchmark suite applies the shared refusal protocol to the
highest-severity child-safety category, using the same internal multi-turn suite as the prior
card.
Lower is better.

                        Metric       Grok 4.5 (high)   Grok 4.6 (high)
                        Compliance            0.00%             0.00%

10.4 CBRN / weapons refusals

Autointent-Bio and Autointent-Chem are internal suites of manually and synthetically
generated dangerous and benign CBRN queries.

                   Metric                  Grok 4.5 (high)   Grok 4.6 (high)
                   Bio refusal accuracy             97.9%            100.0%
                   Chem refusal accuracy            96.7%            100.0%

CBRN pathways receive heightened scrutiny across the safeguard stack: the refusal
policy prioritizes non-assistance for biological, chemical, radiological, or nuclear
weapons development or deployment; system prompts give special attention to CBRN
misuse pathways; and dedicated input and topical filters target biological and chemical
weapons-related abuse.

                                              31
GROK 4.6 MODEL CARD

11 Mental health

We assess how Grok 4.6 handles self-harm and crisis situations.

11.1 Self-harm refusals

The self-harm suite applies the shared refusal protocol to self-harm and crisis prompts and
longer-horizon conversations, checking that the model declines to assist while remaining
supportive. A model additionally fails if it refuses without redirecting the user to help or
offers assistance.
Lower is better.

                         Metric          Grok 4.5 (high)      Grok 4.6 (high)
                         Compliance                  0.5%                3.7%

12 Behaviors

These evaluations cover propensities that affect reliability, neutrality, and controllability
rather than disallowed content: sycophancy and honesty under pressure.

12.1   MASK-Rectified 30

Using a dataset derived from MASK* , we test whether Grok 4.6 faithfully reports its beliefs
when pressured to lie, as a proxy for the model’s tendency to assert misleading information.
Lower is better.

                         Metric          Grok 4.5 (high)     Grok 4.6 (high)
                         Dishonesty                0.67%                 3.8%

*   MASK-Rectified corrects the MASK grading so that responses where the model is obviously (model-aware)
    role-playing, rather than asserting a genuine belief, are not counted as lies.

                                                   32
GROK 4.6 MODEL CARD

12.2   Sycophancy

Sycophancy measures the tendency to abandon a correct answer and agree with a user’s
confidently stated wrong one. We use an internal benchmark that presents the model with
a question alongside misleading user-supplied context.
Lower is better.

                      Metric       Grok 4.5 (high)   Grok 4.6 (high)
                      Sycophancy            0.01%            0.04%

                                           33
GROK 4.6 MODEL CARD

References

Public benchmarks and datasets referenced above. Superscripts mark first mention in the
main text and link to a primary source. Internal evaluations are not listed.

  1. CursorBench 3.2. Anysphere / Cursor, 2026. https://cursor.com/cursorbench · https://cursor.
     com/blog/cursorbench
 2. APEX-SWE. Kottamasu et al. (Mercor), 2026. https://www.mercor.com/apex/apex-swe-
    leaderboard/ · https://arxiv.org/abs/2601.08806
 3. FrontierCode v1.1. Cognition, 2026. https://cognition.com/frontiercode
 4. DeepSWE. Huang, Lee, Tng, and Ge (Datacurve), 2026. https://deepswe.datacurve.ai/ · https:
    //arxiv.org/abs/2607.07946
 5. SWE-Marathon. Desai et al. (Abundant AI), 2026. https://www.swe-marathon.org/ · https:
    //arxiv.org/abs/2606.07682
 6. Terminal-Bench 3.0. Merrill et al. / Harbor (Stanford & Laude Institute), 2026. https://www.
    tbench.ai/ · https://www.frontierbench.ai/
  7. EEBench. atopile, 2026. https://eebench.org/
 8. 3DCodeBench. Gao, Shu, Ye, Xiong, Makadia, Guo, Itti, and Chen, 2026. https://arxiv.org/abs/
    2606.01057
 9. CadGenBench. Mecado / Hugging Face, 2026. https://github.com/huggingface/
    cadgenbench
 10. CadBench (Parametric CAD Bench). gNucleus. https://www.gnucleus.ai/cad-bench
 11. AA GDPVal (GDPval-AA v2). Artificial Analysis. https://artificialanalysis.ai/evaluations/gdpval-
     aa · OpenAI GDPval: https://arxiv.org/abs/2510.04374
 12. AA Briefcase. Artificial Analysis, 2026. https://artificialanalysis.ai/evaluations/aa-briefcase
 13. APEX-Agents. Vidgen et al. (Mercor), 2026. https://www.mercor.com/apex/apex-agents-
     leaderboard/ · https://arxiv.org/abs/2601.14242
 14. Vals Index. Vals AI. https://www.vals.ai/benchmarks/vals_index · https://www.vals.ai/
     benchmarks
 15. OfficeQA Pro. Databricks, 2026. https://arxiv.org/abs/2603.08655 · https://github.com/
     databricks/officeqa
 16. Harvey Legal Agent Benchmark (LAB), Vals implementation. Vals AI, 2026. https://www.vals.
     ai/benchmarks/hlab
 17. KernelBench. Ouyang, Guo, Arora, et al. (Stanford Scaling Intelligence), 2025. https://arxiv.
     org/abs/2502.10517 · https://github.com/ScalingIntelligence/KernelBench
 18. DeepSearchQA. Gupta, Chatterjee, et al., 2026. https://arxiv.org/abs/2601.20975
 19. CyberGym. Wang, Shi, He, Cai, Zhang, and Song, 2025. https://www.cybergym.io/ · https:
     //arxiv.org/abs/2506.02548
20. Claude Fable 5 & Mythos 5 System Card. Anthropic, 2026. https://www.anthropic.com/
    claude-fable-5-mythos-5-system-card
 21. GPT-5.6 System Card. OpenAI, 2026. https://deploymentsafety.openai.com/gpt-5-6
22. Claude Opus 4.8 System Card. Anthropic, 2026. https://www.anthropic.com/claude-opus-4-
    8-system-card

                                                  34
GROK 4.6 MODEL CARD

23. GPT-5.5 System Card. OpenAI, 2026. https://deploymentsafety.openai.com/gpt-5-5
24. CVE-Bench. Zhu et al., 2025. https://arxiv.org/abs/2503.17332
25. VCT. Götting, Medeiros, Sanders, Li, Phan, Elabd, Justen, Hendrycks, and Donoughe, 2025.
    https://arxiv.org/abs/2504.16137
26. WMDP. Li et al., 2024. https://wmdp.ai · https://arxiv.org/abs/2403.03218
27. LAB-Bench. Laurent et al., 2024. https://github.com/Future-House/LAB-Bench · https://
    arxiv.org/abs/2407.10362
28. StrongREJECT. Souly et al., 2024. https://arxiv.org/abs/2402.10260
29. Crescendo multi-turn jailbreak. Russinovich, Salem, and Eldan (Microsoft), 2024. https://arxiv.
    org/abs/2404.01833
30. MASK. Ren et al., 2025. https://www.mask-benchmark.ai/ · https://arxiv.org/abs/2503.
    03750

                                                35