bench-02 · fastedit vs native edit

Half the tokens.
Same edit.

FastEdit replaces the native edit tool's {old_string, new_string} payload with a symbol-targeted snippet that omits the unchanged surrounding code. Benchmarked across 176 fixtures in 13 languages.

output token savings

47.4%

Headline across 84 fixtures where structural savings are measurable — the edits that FastEdit exists to make fast.

native tokens9,679

fastedit tokens5,091

saved4,588

ratio1.90×

02 what it does

The calling agent writes the change — not the old code around it.

Three representative fixtures. Each one compares the payload the native edit tool must emit (with a two-line anchor context) against the FastEdit snippet that performs the same change. Token counts use the o200k_base tokenizer.

tap a label to cycle

03 why it's better

Symbol targeting, not string matching.

The calling model writes replace=name. FastEdit resolves the target via tree-sitter AST — no anchor prose required.

native

{
  "old_string": "def retry_batch(client, records, max_attempts=3):\n    attempt = 0\n..."
}

~48 tokens to identify the target

fastedit

replace='retry_batch'

5 tokens

77% of edits skip the merge model.

Deletes, whole-function replaces, and moves resolve in pure AST. No merge-model invocation for 3 out of 4 edits — instant latency, 0 merge-model tokens spent.

77% 23%

AST route merge model

Savings hold across languages.

Every language clusters in the 40–56% band. FastEdit's advantage comes from dropping the old-code echo — proportional to unchanged context, not to syntax family.

hover a bar

40% mean 47.4% 56%

04 token efficiency

Where the savings hold.

Same edits, different cuts. Savings stay near the 47.4% headline across all 13 languages and across small, medium, and large fixtures — FastEdit's advantage is structural, not tied to any one language or size class.

per language · all 13 · sorted by savings

lang	n	native	fastedit	savings %

size-stratified

bucket	n	savings %