Skip to content

gpt-5.4-fast-benchmark.md

Latest commit

 

History

History
34 lines (26 loc) · 1.61 KB

gpt-5.4-fast-benchmark.md

File metadata and controls

34 lines (26 loc) · 1.61 KB

GPT-5.4 Fast Benchmark

This benchmark was run against OpenCode using the multi-auth Codex plugin in a real repository, with continued-session prompts that read files, analyze code, and reuse the same session across multiple turns.

Summary Table

Mode Avg Time to First Text Avg Time to Complete Total Tokens/sec Output Tokens/sec Reasoning Tokens/sec Completed Turns/min Notes
GPT-5.4 Medium 12.85s 12.92s 8,361 37.7 3.5 4.87 Baseline
GPT-5.4 Medium Fast 12.01s 12.10s 8,594 40.1 4.4 5.01 6.4% faster overall
GPT-5.4 High 57.10s 57.21s 4,645 23.3 7.4 1.27 Baseline
GPT-5.4 High Fast 63.65s 63.72s 6,629 37.0 12.9 1.75 Higher throughput, mixed latency
GPT-5.4 XHigh 88.48s 88.56s 2,836 22.8 13.6 0.88 Baseline
GPT-5.4 XHigh Fast 69.45s 69.55s 3,540 32.5 20.0 1.10 21.5% faster overall

Headline Results

  • XHigh Fast reduced end-to-end continued-session latency by 21.5%.
  • XHigh Fast increased throughput by:
    • +24.8% total tokens/sec
    • +42.7% output tokens/sec
    • +47.2% reasoning tokens/sec
  • Medium Fast delivered a smaller but consistent 6.4% latency improvement.
  • High Fast improved throughput significantly, but did not produce stable latency gains.

Methodology

  • Same repository
  • Same prompt sequence
  • Same continued-session flow
  • Real file reads and tool usage
  • Metrics taken from live OpenCode runs

This benchmark is intended to capture realistic coding-session behavior, not cold-start single-turn chat latency.