← All benchmarks
Benzi vs CodeGraph
6 repos · 1 run/arm · Claude Sonnet 5 · repos, questions, and methodology from CodeGraph's own published benchmark →
Each AI review below was prompted from a new chat on a new account — no shared context between them.
Gemini's read

Benzi Product comes out on top (9.5/10), followed by Benzi MCP (8.8), CodeGraph (8.2), and Bare CC (7.7).

Overview & comparative ranking

RankArmAvg scoreStrengthsWeaknesses
1Benzi Product9.5 / 10Exceptional depth, includes concrete line numbers and internal state flags (abortIndex, no_grammar gotchas, OwnedTasks shut-down race), and clear architecture diagrams/tables.Slightly higher output token overhead (though focused entirely on useful code details).
2Benzi MCP8.8 / 10Very accurate, complete step-by-step traces, excellent handling of asynchronous event loops and memory model details (e.g. Tokio Header/Trailer state transitions).Lacks some of the ultra-deep structural breakdowns and code-level nuances found in Benzi Product.
3CodeGraph8.2 / 10Clean formatting, clear summary sections, strong direct code citations. Very solid everyday technical answers.Slightly higher-level on complex runtime questions (e.g. missing some inner state-machine details in Tokio/Alamofire).
4Bare CC7.7 / 10Accurate code references, concise and direct.Can be overly terse, misses secondary runtime guarantees (memory management/shutdown details, edge-case invariants) compared to the top arms.
DeepSeek's read

Benzi Product comes out on top (9.5), followed by Benzi MCP (9.0), Bare CC (8.9), and CodeGraph (8.6).

Overall average and ranking

RankArmAvg scoreNotes
1Benzi Product9.5Most complete and precise overall. Consistently includes exact call paths, edge cases, and useful summaries.
2Benzi MCP9.0Very strong, accurate, and well-structured. Slightly less exhaustive than Benzi Product on several repos.
3Bare CC8.9Concise, accurate, and usually complete. Very competitive; fewer deep edge-case details.
4CodeGraph8.6Generally excellent, but the Gin answer contains a notable error: it says a middleware that never calls Next() stops the chain. In Gin, it does not necessarily stop; Abort() is what short-circuits.

Key scoring observations

Claude's read

Benzi Product comes out on top (avg 8.7). Bare CC and Benzi MCP tie at 8.0, and CodeGraph is last at 7.7.

Score by repo

RepoBare CCCodeGraphBenzi MCPBenzi Product
gin7689
alamofire8.57.578
okhttp898.58.5
excalidraw8.5879
tokio8798.5
django88.58.59
Average8.07.78.08.7

Bare CC is Claude Code, Sonnet 5. CodeGraph is bare CC + CodeGraph MCP. Benzi MCP is bare CC + Benzi MCP. Benzi Product is Benzi's own standalone agentic loop (running on Claude Sonnet 5 as well, but no Claude Code). Benchmark methodology replicated from CodeGraph's own published benchmarks: github.com/colbymchenry/codegraph

$0.2337
Bare CC
$0.1985
CodeGraph
$0.2655
Benzi MCP
$0.4720
Benzi Product

Our read: we've already shown Benzi is considerably cheaper (2x) and faster (40%) than bare Claude Code on bug-fixing tasks (click here). Code intelligence cutting the cost of localizing and fixing a bug, while raising the cost of an open-ended architectural question, tracks with intuition — the same tools that shortcut a targeted fix invite more exploration on a broad "how does this work" question. Benzi's answers also generally cover more depth, or connect systems that bare Claude Code and Claude Code with CodeGraph as an MCP tool miss.