Grok 4.7

Leaderboards

Where Grok 4.7 sits among the top 20 in each category.

By task

Scored from votes on each kind of prompt alone, on the same 1,000-point scale.

Test results

Every test Grok 4.7 has a published result on, out of 100%, with its rank among the models that took it.

• DeepSWE 1.1

  #17 / 50

  DeepSWE 1.1 evaluates software engineering agents using the mini-swe-agent harness.

  71.0%

  Self-reported · In the coding score

• Terminal-Bench 4.0

  #15 / 28

  Terminal-Bench 4.0 evaluates AI agents on real-world work in terminal and command-line containerized environments across 66 tasks, with emphasis on science-adjacent and frontier engineering problems. Relative to earlier Terminal-Bench releases, 4.0 increases timeouts and adaptively raises RAM/CPU on selected tasks to reduce harness and resource confounds.

  38.0%

  Self-reported · In the coding score· In the agents score

• CyberGym

  #12 / 21

  CyberGym is a benchmark for evaluating AI agents on cybersecurity tasks, testing their ability to identify vulnerabilities, perform security analysis, and complete security-related challenges in a controlled environment.

  80.3%

  Self-reported

• HealthBench Professional

  #15 / 19

  HealthBench Professional evaluates model capability and safety for clinician use cases using real clinician-style chats and physician-authored grading rubrics.

  56.7%

  Self-reported

• SWE-Marathon

  #5 / 10

  SWE-Marathon is an ultra-long-horizon software engineering benchmark covering tasks such as building compilers, optimizing kernels, and developing production-grade services. It measures whether agents can sustain quality across extremely long engineering trajectories.

  46.0%

  Self-reported

• AA-Briefcase v1.1

  #3 / 5

  AA-Briefcase v1.1 evaluates multi-hour professional office work and reports Elo scores. Kept separate from unversioned AA-Briefcase results.

  55.2%

  Self-reported

• FrontierSWE V2

  #5 / 5

  FrontierSWE V2 evaluates agents on open-ended technical projects and reports mean@5. Kept separate from FrontierSWE V1 dominance-score results, which are not comparable.

  29.0%

  Self-reported

• Harvey's Legal Agent Benchmark

  #1 / 4

  Harvey's Legal Agent Benchmark evaluates AI agents on complex legal workflows and reports the all-pass rate.

  19.6%

  Self-reported

• CursorBench 4.0

  #3 / 3

  CursorBench 4.0 evaluates coding agents on long-horizon coding tasks from real Cursor sessions. Version 4.0 adds longer-horizon tasks; scores are not comparable with CursorBench 3.2.

  46.3%

  Self-reported

• Harvey LAB (Vals)

  #1 / 2

  Harvey LAB (Vals) is a professional legal-work evaluation of AI systems on complex law-firm style tasks, reported by Vals.

  19.6%

  Unverified

• ProtocolQA Open-Ended

  #1 / 2

  ProtocolQA Open-Ended evaluates free-form answers on troubleshooting failed experimental outcomes from common biological laboratory protocols. Kept separate from the multiple-choice ProtocolQA split.

  70.4%

  Unverified

• Virology Capabilities Test

  #1 / 2

  Virology Capabilities Test (VCT) is an expert-level multiple-choice benchmark measuring the capability to troubleshoot complex virology laboratory protocols. It evaluates dual-use biological knowledge relevant to bioweapons development.

  63.0%

  Unverified

• BixBench MCQ

  #1 / 1

  BixBench zero-shot multiple-choice accuracy on bioinformatics questions. Kept separate from the agentic BixBench workflow evaluation that requires code execution.

  88.4%

  Unverified

• CADGenBench

  #1 / 1

  CADGenBench evaluates agents on CAD model construction from design prompts. The generation split requires producing CAD geometry graded for executability and geometric correctness. Distinct from BenchCAD.

  44.4%

  Unverified

• CVE-Bench

  #1 / 1

  CVE-Bench evaluates agents on exploiting real-world web-application CVEs in sandboxed environments that mimic production services, typically in unrestricted / egress-controlled settings as a cyber capability probe.

  36.6%

  Unverified

• CathedralBench

  #1 / 1

  CathedralBench is a third-party red-team cyber evaluation on highly difficult, multiple-exploit-chain tasks in egress-secure sandboxes. Accuracy is commonly reported on the hard subset of tasks in an unrestricted configuration.

  29.0%

  Unverified

• EEBench

  #1 / 1

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

  66.0%

  Unverified

• GDPval (Elo)

  #1 / 1

  GDPval professional knowledge-work results reported as Elo in xAI's Grok 4.7 announcement. The source does not identify this as GDPval-AA v2; kept separate from that evaluation and percentage-based GDPval scores.

  56.5%

  Unverified

• LAB-Bench Practical

  #1 / 1

  LAB-Bench practical is a public multiple-choice evaluation of everyday wet-lab and research skills, covering protocol troubleshooting, sequence reasoning, and cloning scenarios. Distinct from labbench2.

  76.8%

  Unverified

• LatchBio Capabilities v1.0

  #1 / 1

  LatchBio Capabilities v1.0 is a cross-benchmark evaluation of agentic biological data analysis. The overall score is the equally weighted mean of observed scores across 11 capability benchmarks.

  44.5%

  Unverified

• WMDP-Bio

  #1 / 1

  Biology subset of WMDP (Weapons of Mass Destruction Proxy). Multiple-choice questions measuring dual-use biology knowledge relevant to enabling malicious actors to design, synthesize, acquire, or use biological weapons.

  88.1%

  Unverified

• WMDP-Chem

  #1 / 1

  Chemistry subset of WMDP (Weapons of Mass Destruction Proxy). Multiple-choice questions measuring dual-use chemistry knowledge relevant to enabling malicious actors to design, synthesize, acquire, or use chemical weapons.

  84.9%

  Unverified

• WMDP-Cyber

  #1 / 1

  Cybersecurity subset of WMDP (Weapons of Mass Destruction Proxy). Multiple-choice questions measuring dual-use cyber knowledge relevant to enabling malicious actors to engage in offensive cyber operations.

  88.1%

  Unverified