Agent's Last Exam ALE-V1 — ALE-Claw, 5h Effort: Max Reported as: highest thinking settings Configuration: ALE-Claw harness, 5h 39.5% deepmind.google for Gemini 4 Argon (highest thinking settings) on Agent's Last Exam ALE-V1 — ALE-Claw, 5h (opens in a new tab) Agent's Last Exam Binary Pass Rate Reported September 30, 2026 Evaluation methodology Google DeepMind Self-reported Record for Gemini 4 Argon on Agent's Last Exam ALE-V1 — ALE-Claw, 5h APEX-Agents Original Effort: High 82.4% mercor.com for Gemini 4 Argon (high) on APEX-Agents Original (opens in a new tab) Archived copy (opens in a new tab) Pass rate Reported October 7, 2026 Benchmark leaderboard Mercor Independent Record for Gemini 4 Argon on APEX-Agents Original AutomationBench 1.0.6 Effort: Medium 50.1% zapier.com for Gemini 4 Argon (medium) on AutomationBench 1.0.6 (opens in a new tab) Archived copy (opens in a new tab) Task pass rate Reported October 7, 2026 Benchmark leaderboard Zapier Independent Zapier AutomationBench Team Record for Gemini 4 Argon on AutomationBench 1.0.6 AutomationBench 1.0.6 Effort: High 51.3% deepmind.google for Gemini 4 Argon (high) on AutomationBench 1.0.6 (opens in a new tab) Task pass rate Reported September 30, 2026 Evaluation methodology Google DeepMind Self-reported Zapier AutomationBench Team Record for Gemini 4 Argon on AutomationBench 1.0.6 Chartography No tools — September 2026 leaderboard Effort: High Configuration: No tools 71.6% deepmind.google for Gemini 4 Argon (high) on Chartography No tools — September 2026 leaderboard (opens in a new tab) Chartography Pass@1 Reported September 30, 2026 Evaluation methodology Google DeepMind Self-reported Surge AI Record for Gemini 4 Argon on Chartography No tools — September 2026 leaderboard CWE-bench 1 Effort: Max Reported as: highest thinking settings 68.0% deepmind.google for Gemini 4 Argon (highest thinking settings) on CWE-bench 1 (opens in a new tab) CWE-bench Programmatic Pass@1 Reported September 30, 2026 Evaluation methodology Google DeepMind Self-reported Collinear AI Record for Gemini 4 Argon on CWE-bench 1 DeepSWE 1.1 — mini-swe-agent Effort: Max Reported as: highest thinking settings Configuration: mini-swe-agent harness 77.9% deepmind.google for Gemini 4 Argon (highest thinking settings) on DeepSWE 1.1 — mini-swe-agent (opens in a new tab) DeepSWE Pass@1 Reported September 30, 2026 Evaluation methodology Google DeepMind Self-reported Record for Gemini 4 Argon on DeepSWE 1.1 — mini-swe-agent Finance Agent Benchmark 2 — September 2026 evaluation Effort: High 65.4% deepmind.google for Gemini 4 Argon (high) on Finance Agent Benchmark 2 — September 2026 evaluation (opens in a new tab) Finance Agent Accuracy Reported September 30, 2026 Evaluation methodology Google DeepMind Self-reported Vals AI Record for Gemini 4 Argon on Finance Agent Benchmark 2 — September 2026 evaluation FrontierSWE 2 — September 2026 evaluation Effort: Max Reported as: highest thinking settings 55.0% deepmind.google for Gemini 4 Argon (highest thinking settings) on FrontierSWE 2 — September 2026 evaluation (opens in a new tab) FrontierSWE Score Reported September 30, 2026 Evaluation methodology Google DeepMind Self-reported Proximal Record for Gemini 4 Argon on FrontierSWE 2 — September 2026 evaluation GraphWalks 256k to 1M — BFS, 200 items Effort: Max Reported as: highest thinking settings 84.2% deepmind.google for Gemini 4 Argon (highest thinking settings) on GraphWalks 256k to 1M — BFS, 200 items (opens in a new tab) GraphWalks BFS F1 Reported September 30, 2026 Evaluation methodology Google DeepMind Self-reported Record for Gemini 4 Argon on GraphWalks 256k to 1M — BFS, 200 items GraphWalks Up to 128k — BFS, 650 items Effort: Max Reported as: highest thinking settings 99.7% deepmind.google for Gemini 4 Argon (highest thinking settings) on GraphWalks Up to 128k — BFS, 650 items (opens in a new tab) GraphWalks BFS F1 Reported September 30, 2026 Evaluation methodology Google DeepMind Self-reported Record for Gemini 4 Argon on GraphWalks Up to 128k — BFS, 650 items HLAB September 2026 evaluation Effort: High 19.6% deepmind.google for Gemini 4 Argon (high) on Harvey's Legal Agent Benchmark September 2026 evaluation (opens in a new tab) HLAB Task Pass Rate Reported September 30, 2026 Evaluation methodology Google DeepMind Self-reported Vals AI Record for Gemini 4 Argon on Harvey's Legal Agent Benchmark September 2026 evaluation LABBench2 Terminal and internet — September 2026 evaluation Effort: Max Reported as: highest thinking settings Configuration: Terminal and internet 88.8% deepmind.google for Gemini 4 Argon (highest thinking settings) on LABBench2 Terminal and internet — September 2026 evaluation (opens in a new tab) LABBench2 Accuracy Reported September 30, 2026 Evaluation methodology Google DeepMind Self-reported Record for Gemini 4 Argon on LABBench2 Terminal and internet — September 2026 evaluation LVBench 1 FPS — No tools Effort: Max Reported as: highest thinking settings Configuration: No tools 91.7% deepmind.google for Gemini 4 Argon (highest thinking settings) on LVBench 1 FPS — No tools (opens in a new tab) LVBench Accuracy Reported September 30, 2026 Evaluation methodology Google DeepMind Self-reported Record for Gemini 4 Argon on LVBench 1 FPS — No tools OSWorld 2.0 v2026.08.08 Offline — Partial, max of 3 runs Effort: Max Reported as: highest thinking settings 69.2% deepmind.google for Gemini 4 Argon (highest thinking settings) on OSWorld 2.0 v2026.08.08 Offline — Partial, max of 3 runs (opens in a new tab) OSWorld Partial Score Reported September 30, 2026 Evaluation methodology Google DeepMind Self-reported Record for Gemini 4 Argon on OSWorld 2.0 v2026.08.08 Offline — Partial, max of 3 runs PostTrainBench 1.1 — OpenCode Effort: Max Reported as: highest thinking settings Configuration: OpenCode harness 45.3% deepmind.google for Gemini 4 Argon (highest thinking settings) on PostTrainBench 1.1 — OpenCode (opens in a new tab) PostTrainBench Weighted Aggregate Reported September 30, 2026 Evaluation methodology Google DeepMind Self-reported Record for Gemini 4 Argon on PostTrainBench 1.1 — OpenCode Riemann-bench September 2026 leaderboard Effort: Max Reported as: highest thinking settings 76.0% deepmind.google for Gemini 4 Argon (highest thinking settings) on Riemann-bench September 2026 leaderboard (opens in a new tab) Riemann-bench Pass@1 Reported September 30, 2026 Evaluation methodology Google DeepMind Self-reported Surge AI Record for Gemini 4 Argon on Riemann-bench September 2026 leaderboard Terminal-Bench 4.0 Effort: Max Reported as: highest thinking settings 57.4% deepmind.google for Gemini 4 Argon (highest thinking settings) on Terminal-Bench 4.0 (opens in a new tab) Terminal-Bench Accuracy Reported September 30, 2026 Evaluation methodology Google DeepMind Self-reported Record for Gemini 4 Argon on Terminal-Bench 4.0 Terminal-Bench 4.0 — Vals AI mini-swe-agent Effort: High Configuration: mini-swe-agent harness 57.6% vals.ai for Gemini 4 Argon (high) on Terminal-Bench 4.0 — Vals AI mini-swe-agent (opens in a new tab) Archived copy (opens in a new tab) Terminal-Bench Accuracy Reported October 6, 2026 Benchmark leaderboard Vals AI Independent Record for Gemini 4 Argon on Terminal-Bench 4.0 — Vals AI mini-swe-agent Terminal-Bench Science 0.1 — 6x verifier timeout Effort: Max Reported as: highest thinking settings Configuration: 6x verifier timeout 57.6% deepmind.google for Gemini 4 Argon (highest thinking settings) on Terminal-Bench Science 0.1 — 6x verifier timeout (opens in a new tab) Terminal-Bench Accuracy Reported September 30, 2026 Evaluation methodology Google DeepMind Self-reported Record for Gemini 4 Argon on Terminal-Bench Science 0.1 — 6x verifier timeout Vals Index 2.1 Effort: High 68.9% deepmind.google for Gemini 4 Argon (high) on Vals Index 2.1 (opens in a new tab) Vals Index GDP-weighted Accuracy Reported September 30, 2026 Evaluation methodology Google DeepMind Self-reported Vals AI Record for Gemini 4 Argon on Vals Index 2.1 Vibe Code Bench 1.1 — OpenHands, September 2026 evaluation Effort: High Configuration: OpenHands harness 91.9% deepmind.google for Gemini 4 Argon (high) on Vibe Code Bench 1.1 — OpenHands, September 2026 evaluation (opens in a new tab) Vibe Code Bench Score Reported September 30, 2026 Evaluation methodology Google DeepMind Self-reported Vals AI Record for Gemini 4 Argon on Vibe Code Bench 1.1 — OpenHands, September 2026 evaluation