Benchmarks
Up to 53% fewer tokens on Claude Code.
Up to 57% on Codex.
Measured by replaying real recorded Claude Code and Codex sessions; medians 36% and 32% on long sessions. Savings grow with session length; short sessions change little. See your own number: install, sign in, then run one command. It runs offline — nothing leaves your machine.
densilo benchmarkResults
What the model received, with and without Densilo.
Every recorded session, short ones too, added up.
| Agent | Tokens sent | Input cost |
|---|---|---|
| Claude Code | −39% | −34% |
| Codex | −31% | −27% |
Why it adds up: about 75% of what Claude Code costs is re-reading context it already sent. The longer a session runs, the more of that there is, and the more Densilo saves.
How we measured
Real sessions, replayed.
We took real recorded Claude Code and Codex coding sessions and replayed every request through Densilo on the same machine, then counted the tokens the model would receive with and without it and priced them as the provider bills input. No model was called.
These are one team's sessions so far. Run the benchmark on yours to see your own.
What it doesn't show
A replay measures tokens and cost, not task quality.
- In a replay, the model's answers are the ones already recorded, so it can't tell you whether the work would have gone as well. Use Densilo on real work and judge that yourself.
- We ran these ourselves, on one team's sessions. They are not an independent audit, and they are not your codebase.
- Short sessions change little. The numbers above come from long sessions, where there is the most to save.
- The amounts assume you pay per token (an API key or a cloud provider). On a flat subscription (Claude Pro or Max, ChatGPT) fewer tokens are not money back. On a Claude Pro or Max plan they mean 2× more work before you hit your usage limit. If your numbers say otherwise, tell us.
Measure yours
See your own number, offline.
densilo benchmark replays your recent Claude Code and Codex sessions for a repository through Densilo on your machine. No model is called, nothing is uploaded, and it takes a few minutes.
- Install and sign in.
curl -fsSL https://api.densilo.com/install.sh | sh, thendensilo login. The benchmark also runs signed out, but then it understates the saving. - Go to a repository you have worked in with Claude Code or Codex. Longer sessions show more; five or more sessions give a steadier result.
- Run
densilo benchmark, ordensilo benchmark --sessions 0to replay all of that repository's sessions, not just the newest 20. It finds them in~/.claude/projectsand~/.codex/sessionsand saves a report in~/.densilo/benchmarks/. - Read the report. It shows tokens sent and input cost without and with Densilo, for your own sessions.
| Option | What it does |
|---|---|
--sessions N | Newest N sessions (default 20; 0 for all) |
--since DAYS | Only sessions from the last DAYS days |
--agent claude|codex | Only one agent's sessions (default: both) |
--json | Print the numbers as JSON |
Questions about the results? Email support@densilo.com.
Measure it on your own work.
Install it, run densilo benchmark in a repository you work in, and see what it would have saved.