Codex CLI cross-file navigation benchmark
A paired, end-to-end comparison on an anonymized 553-source-file TypeScript repository. This measured complete Codex CLI sessions—not estimated payload sizes.
Result
| Metric | Codex CLI alone | With Speeder | Difference |
|---|---|---|---|
| Average input tokens | 153,323 | 100,599 | 34.4% fewer |
| Average completion time | 50.8 seconds | 41.1 seconds | 19.0% faster |
| Total output tokens | 7,380 | 5,334 | 27.7% fewer |
| Tool calls | 20 | 14 | 30.0% fewer |
| Quality pass rate | 3 / 3 | 3 / 3 | Equal |
Across all three pairs, provider-reported input tokens fell from 459,970 to 301,798. Wall time fell from 152.302 to 123.421 seconds.
Method
- Codex CLI 0.146.1 with GPT-5.6 Terra at medium reasoning effort.
- Same user prompt, source snapshot, task, permissions, and completion criteria.
- Three fresh single-task sessions per variant, paired by run.
- The Speeder condition used a cold isolated comparison with cache and capture disabled.
- Usage came from Codex CLI's provider-reported turn completion data.
Quality gate
The task required planning a rejected-RPC error-handling change in a shared function. A passing answer had to identify the implementation symbol, all six direct callers, focused test coverage, and the risk boundary around the RPC call. All six sessions passed.
Limitations
This supports a scoped claim for cross-file TypeScript navigation. It does not establish universal token, latency, or cost savings. Input tokens include cached input, which may be priced differently. Small tasks with an already-known file or assertion can be faster without structural discovery and should bypass Speeder.
Data integrity
Aggregate and per-run values are retained with the product's benchmark records. Session logs and repository source are excluded to avoid publishing prompts, local paths, or unrelated project context.