71% fewer tokens. Measured on a real external open-source project — reproducible.
The Formbricks measurement — reproducible on a real external codebase
The most credible measurement is the one anyone can reproduce. We cloned Formbricks cold from GitHub — a real external TypeScript survey platform we don't maintain — and measured token usage fixing DEF-FBK-001 (PII leak in survey exports) with and without SysEdge. Both paths used actual Anthropic API token counts from --output-format json, not file-size estimates, so you can clone the same repo and re-run it.
Without SysEdge: 14,512 input+output tokens (30 input + 14,482 output), plus 1,797,494 cache-read tokens to orient on the codebase
With SysEdge: 4,141 input+output tokens (22 input + 4,119 output), plus 473,649 cache-read tokens
Result: 71% fewer input+output tokens — ~73% if you count total tokens including cache reads
Internal measurements — four tasks
We also measured four tasks on our own live 44-module Go/TypeScript codebase. These use len(text)/4 estimates, not exact API counts.
| Task | Module size | Without SysEdge | With SysEdge | Saving |
|---|---|---|---|---|
| DEF-001 — MCP scope enforcement (licmcp · small service · 1 file) | small | 21,929 | 1,348 | −93% |
| ENH-103 — Remove legacy impl-planner surface (plan · 4 Go files + 1 TSX) | medium | 31,467 | 1,683 | −94% |
| ENH-330 — Delegation inbox API endpoint (manage · 2,889-line handler file) | large | 48,588 | 1,423 | −97% |
| ENH-200 — CreateActionDialog + reassign endpoint (core · Go + TSX + API) | large | 53,114 | 1,086 | −97% |
| Average (4 internal tasks) | — | 38,774 | 1,385 | −96% |
| DEF-FBK-001 — PII leak in survey exports (Formbricks · external open-source · actual API tokens) | real-world | 14,512 | 4,141 | −71% |
ROI calculation
Orientation cost only · implementation tokens are equal in both paths and excluded from this calculation.
Faster at higher session counts. Token savings alone — coordination and test-coverage benefits not included in this calculation.
Full measurement methodology
We measured the orientation phase only — the tokens a session reads BEFORE writing any code. Implementation tokens are equal in both paths and excluded.
Without SysEdge: we listed every file a Claude Code session would realistically read to orient on the task — the root CLAUDE.md, instance CLAUDE.md, the enhancements/defects file, and relevant source files. File sizes are actual bytes from the live codebase.
With SysEdge: we captured the actual output of briefing --instance X, worklog --instance X, and show-enhancement --id ENH-NNN for each task, plus the instance CLAUDE.md (read in both).
Token estimate: len(text) / 4 — a conservative approximation for mixed code and prose. Claude's actual BPE tokeniser gives slightly fewer tokens for code-heavy content.
Caveat: the "without" scenario reads whole source files. A skilled session using Read with offset/limit on targeted sections would read fewer tokens — we estimate the honest real-world saving at 60–80% for medium and large modules.
Formbricks measurement: a reproducible measurement on a real external open-source TypeScript project, cloned cold from GitHub. Unlike the four internal tasks, this uses actual Anthropic API token counts (from --output-format json). Without SysEdge, orienting on the defect cost 30 input + 14,482 output tokens (14,512 total), plus 1,797,494 cache-read tokens. With SysEdge, graph queries cost 22 input + 4,119 output (4,141 total), plus 473,649 cache-read tokens. That is a 71% reduction in input+output tokens — and ~73% if total tokens including cache reads are counted (1,812,006 → 477,790), so the result holds whichever basis you use.