Pixel vs WarpGrep

Morph's search subagent: a model trained with reinforcement learning drives ripgrep and file reads on your code for a few turns and hands back the relevant spans.

Updated

Short answer

On 45 plain-English queries, Pixel 0.6.0 puts the right file first for 87%, WarpGrep for 69%, and has it in its top 10 for 96% against 71%; Pixel answers in 0.7 s on average, WarpGrep in 6.8 s. WarpGrep needs no index and hands the agent the code spans themselves; each search is a paid call, and the lines its tools read go to Morph's API. Its own claim is agent-level, on SWE-Bench Pro, which this benchmark does not measure.

Side by side

Measured head to head. Pixel and WarpGrep on the same cases, on the same machine.

MeasurePixelWarpGrep
Right file first (45 plain-English queries)87%69%
Right file in the top 1096%71%
Average answer time0.7 s6.8 s
Index to build firstYes, in .pixel/None
Where the search runsOn your machineTools local, model on Morph's API
Cost per searchNoneA paid API call

Every figure above is on the benchmarks page, with its sample size, its method and the cases where Pixel loses.

Where WarpGrep wins

Nothing to build before the first search, and an answer that is the code itself, in one or two files, rather than a ranked list.

How they differ

WarpGrep is Morph’s search subagent: a model trained with reinforcement learning (morph-warp-grep-v2.1) drives ripgrep, directory listings and file reads on your tree for up to six turns, then hands the agent the code spans it judged relevant. It keeps no index, so there is nothing to build, and each search is a paid call to Morph’s API. pixel search-meaning answers from an index in .pixel/: a ranked list of files with the matching chunk of each, computed on the machine.

What leaves the machine

WarpGrep’s SDK runs its tools locally, but their results, grep lines and reads of up to 800 lines, are the model’s next input and are sent to Morph’s API. Pixel’s search sends nothing: the embedding model is downloaded once, from Hugging Face, on first use, and the index never leaves the machine.

Reading the numbers

WarpGrep’s figures come from its one run, on 2026-09-27 beside Pixel 0.5.2, on the same queries and code; it was not re-run for Pixel 0.6.0, since each search is a paid call. It names about one file per answer, so its top-10 score is close to its first-answer score, and it is not deterministic: of 19 cases scored in two runs on the same day, 2 changed verdict. Its own claim is agent-level, on SWE-Bench Pro: a coding agent with WarpGrep solves more tasks with fewer input tokens. That is not what this benchmark measures, for either tool.