AI Benchmark Digest — 2026-07-09

View on AI Benchmark Hub

Daily

New Models (1)

• Grok 4.5 — ELO 2494, #13/1467, above Claude Opus 4.6, below Claude Sonnet 5
  • SWE-Marathon: 29.0 (#1/12)
  • Vals AI Harvey Legal Agent Bench: 12.92 (#1/16)
  • Vals AI SWE-bench Verified: 86.6 (#3/66)
  • VoxelBench: 1796.0 (#4/44)
  • Vals AI MedScribe: 86.88 (#4/67)
  • Vals AI GPQA: 92.93 (#4/118)
  • Ramp SWE-Bench: 81.01 (#5/21)
  • Vals AI Legal Research Bench: 37.98 (#5/15)
  • Vals AI (Vals Index): 65.3 (#6/32)
  • BridgeBench Security: 81.7 (#7/34)

New Scores From Top-10 Models (6)

• Qwen 3.7 Max on Agent Arena: 0.36 Net Improvement (%) (#30/32)
• Qwen 3.7 Max on Agent Arena - Bash Recovery: 7.91 Bash Recovery (%) (#20/32)
• Qwen 3.7 Max on Agent Arena - Confirmed Success: 1.43 Confirmed Success (%) (#29/32)
• Qwen 3.7 Max on Agent Arena - Praise vs Complaint: 0.52 Praise vs Complaint (%) (#31/32)
• Qwen 3.7 Max on Agent Arena - Steerability: 5.2 Steerability (%) (#22/32)
• Qwen 3.7 Max on Agent Arena - Tool Hallucination: 1.02 Tool Hallucination (%) (#14/32)

New #1 Leaders (7)

• LiveBench Simplify (Score): Gemini 1.5 Flash (Preview 0827) (84.75) beat Gemini 3.1 Pro (Preview) (High) (71.05) by 13.7.
• TabArena All Tasks (Score (%)): TabFM (default) (89.0) beat TabPFN-3 (default) (77.4) by 11.6.
• SWE-Marathon (Pass@1 (%)): Grok 4.5 (29.0) beat Claude Opus 4.8 (26.0) by 3.0.
• LiveBench Paraphrase (Score): Llama 3.1 70B Instruct (82.25) beat Gemini 3.1 Pro (Preview) (High) (79.87) by 2.38.
• LiveBench Story Generation (Score): Gemini 1.5 Flash (Preview 0827) (81.33) beat GPT-5.4 (xHigh) (79.02) by 2.31.
• Vals AI Harvey Legal Agent Bench (Accuracy (%)): Grok 4.5 (12.92) beat Claude Fable 5 (11.25) by 1.67.
• Tau3 Banking (Success Rate (%)): Grok 4.5 (High) (32.58) beat GPT-5.5 (xHigh) (31.34) by 1.24.