TokenWarden across the latest benchmark run.
Benchmark results
TokenWarden sends far less context to the model.
Public benchmark snapshots compare opencode context usage with TokenWarden, no plugin, and other optimization plugins.
The latest published run includes 60 tests across Qwen 3.7 Max and Qwen 3.5 9B local models.
Compared with no optimization plugin.
Measured across nine latest-run samples.
Best TokenWarden task result in the latest run.
Latest run
Overall token results.
| Adapter | Runs | Median tokens | Average tokens | Median tokens kept out | Average tokens kept out |
|---|---|---|---|---|---|
| No plugins | 9 | 346,261 | 547,728 | 0 | 0 |
| TokenWarden | 9 | 116,570 | 189,402 | 250,047 | 358,326 |
| OpenSlimEdit | 9 | 257,426 | 303,820 | -6,288 | 243,908 |
| DCP | 9 | 241,324 | 281,644 | 150,174 | 266,084 |
| OpenRTK | 9 | 452,249 | 530,245 | -200,501 | 17,483 |
TokenWarden detail
Task-level reductions.
| Task | Median tokens | Average tokens | Median reduction |
|---|---|---|---|
| Routing ledger | 120,669 | 132,446 | 67.45% |
| Helper ledger | 67,724 | 186,143 | 71.67% |
| Core API ledger | 116,570 | 249,617 | 93.54% |
Earlier run
Same pattern, smaller sample.
| Adapter | Median tokens | Average tokens | Median tokens kept out |
|---|---|---|---|
| No plugins | 439,108 | 465,366 | 0 |
| TokenWarden | 83,104 | 77,064 | 376,028 |
| OpenSlimEdit | 262,551 | 239,407 | 176,557 |
| DCP | 212,342 | 209,273 | 226,766 |
| OpenRTK | 404,820 | 352,022 | 34,288 |