All numbers are real — measured by running target/release/tare on the committed corpus.
Tokenizer: tiktoken o200k_base (v0.13.0). Results committed to crates/tare-bench/results/;
reproduce with python3 crates/tare-bench/run_proof.py.
| Name | Content type | Command | Input tokens | Output tokens | Reduction |
|---|---|---|---|---|---|
| cargo_packages | json_array | tare compact-lossy |
6,906 | 3,625 | 47.5% |
| ps_aux | tabular | tare compact-lossy |
1,545 | 802 | 48.1% |
| app_log | logs | tare compact-lossy |
13,217 | 6,551 | 50.4% |
| agent_context | agent_context | tare compress |
15,130 | 8,499 | 43.8% |
| server_rs | code | tare skeletonize --path server.rs |
5,930 | 1,582 | 73.3% |
| json_crush_rs | code | tare skeletonize --path json_crush.rs |
3,937 | 1,607 | 59.2% |
| readme_prose | prose | tare compact-lossy |
5,732 | 2,727 | 52.4% |
Code reads are ~67–76% of a coding agent's tokens (SWE-Pruner, ACL 2026), so skeletonization is the single biggest lever.
The numbers above are tare-only (no competitor tools required). Head-to-head comparisons against
Headroom, LLMLingua-2, lean-ctx, and RTK live in crates/tare-bench/benchmarks/ (Python scripts
that require the competitor tools installed). Run the scripts there to reproduce; at equal
fidelity, tare matches or beats each — and is the only one with a lossless mode and cross-turn dedup.
!!! note
tare has been smoke-tested end-to-end against the live Anthropic API — through the proxy on a
Claude subscription (scripts/live-smoke-sub.sh) and via the MCP server over real stdio — but is
not yet production-hardened or load-tested.