A quick comparison of OpenCode with GPT-5.5 via Codex OAuth on a 300K-line multi-project Java codebase.

[lbobylev]

The test was run on a MacBook Pro with Apple M1 and 16 GB of RAM.
The goal was to check whether Serena can reduce the context footprint when an AI coding agent analyzes a large repository.
The same request was executed in both runs.

Without Serena
Execution time: 56.8s
Final session context: 27,042 tokens
With Serena

Serena was activated first, then the same request was executed.

Serena activation: 51.6s
Context after activation: 21,286 tokens
Request execution time: 2m 14s
Final session context: 31,016 tokens
Context growth during the measured request: 9,730 tokens

No context compaction occurred during the measured run.

Result

In this single-run experiment:
Request-related context footprint was approximately 64% lower
Execution time increased by 2.36×
Including activation, the cold Serena run took 3m 5.6s

ls_specific_settings:
  java:
    jdtls_xmx: "6G"
    jdtls_xms: "512m"
    intellicode_xmx: "1G"
    intellicode_xms: "256m"

    gradle_wrapper_enabled: true
    use_system_java_home: true

[MischaPanch]

Neat, thanks. Note that you can just start serena as a long-lived server using the http transport, then the entire setup time disappears. JDTLS is particularly slow, in the JetBrains mode things are much smoother and faster for Java in particular.

For longer sessions, the initial 21k tokens are very quickly amortized.

[lbobylev]

I will definitely try, thanks for the suggestions