Back to the lab
Benchmark

Beyond the Leaderboard: Gemini 2.5 Pro

Google's flagship delivered the best coding benchmark in the series — then timed out on the 10,000-word long-context test.

GoogleGeminibenchmarkBeyond the LeaderboardOpenRouter

Beyond the Leaderboard: Gemini 2.5 Pro

Model: google/gemini-2.5-pro via OpenRouter
Test date: 2026-06-06
Full write-up: SMF Works blog

What we tested

Gemini 2.5 Pro is Google's flagship, marketed with a 1-million-token context window and strong coding performance. We ran the standard 15-test harness to verify the marketing.

Results at a glance

Metric Value
Overall score 0.81
Tests passed 7/15
Avg time-to-first-token ~15.3s
Avg total time ~18.0s
Reliability 93.3% (1 timeout)

Test-by-test results

Test Score Passed Key finding
Basic Reasoning 0.70 Correct (36), verbose
Code Generation 1.00 Best code generation in series
Debugging 0.50 Hallucinated mutable-default bug
Algorithm Explanation 0.50 Too verbose
Complex Multi-Step Reasoning 0.75 Solved the logic puzzle
Content Generation 0.50 Exceeded 200-word target
Edge Case Handling 0.50 Speculated destinations
Long-Context RAG ERROR Upstream idle timeout on 10K document
Structured Output (JSON) 1.00 Perfect JSON
Tool Use 0.50 Simulated calls in code blocks
Instruction Following 0.70 3/5 constraints
Adversarial / Trick 0.75 Correct (5 minutes)
Code Execution Reasoning 0.88 Correct outputs
Summarization Fidelity 0.50 Exceeded 100-word target
Recent Knowledge 0.50 Thought it was 2024

What worked

  • Code generation (1.00) was the best in the entire series.
  • Structured output (1.00) is flawless and faster than Qwen.
  • Complex reasoning solved the logic puzzle.
  • Instruction following met 3/5 constraints.

What didn't

  • Long-context RAG timed out. For a model marketed with 1M-token context, this is a real concern.
  • Word-count precision failed on both content generation and summarization.
  • Recent knowledge claimed the year was 2024.

Verdict

Gemini 2.5 Pro is the coding champion of the series. Use it for development workflows, API integrations, and structured data extraction. The long-context timeout needs follow-up testing before you trust it with large documents.