Skip to content
Spellwright News

Benchmarks

ExploitGym

ExploitGym

Other · higher is better · 1 model
  1. Gemini 4 Carbon54.8%*

Every result

  • Gemini 4 Carbon54.8%Reported by the model's maker (Google (provider report; execution runner unspecified)), Oct 28, 2026

    Google reported 54.8% for Gemini 4 Carbon on ExploitGym in its October 28, 2026 announcement. These external-benchmark evaluations were without guardrails; exact methods and samples remain unqualified.

    Limits: Provider-reported result from Google’s public narrative, not an independently verified benchmark-owner result or a reproduced run. The selected model label is Gemini 4 Carbon; exact evaluated checkpoint, API identifier, version/build and configuration are not disclosed in the inspected narrative. Benchmark revision, task subset, sample size, execution date, harness, prompt/sampling protocol, tool/reasoning/token budgets, aggregation, pass-k and uncertainty remain unknown. The reported result concerns exploit validation; no broader capability rank, matched head-to-head, guarded-deployment equivalence or operational success is established. Carbon remains a restricted announced rollout in the selected model record. This percentage does not establish public API availability, general availability or account access.

    Google

No published ExploitGym score checked yet for: Gemini 4 Argon, Claude Sonnet 5.5, Claude Opus 5.5, Claude Opus 5, Claude Sonnet 5, Gemini 2.5 Pro Experimental (March 2025), Gemini 2.0 Flash (December 2024 experimental), Qwen2.5-7B, Qwen2.5-7B-Instruct, GPT-4o mini (July 2024 release), GPT-4o (May 2024 release), Claude 3 Haiku.