WarpGrep v2 is a fast code search subagent that works via parallel tool calls. Runs up to 36 grep/read tool calls in under 5s.

SWE-Bench Pro: WarpGrep v2

Scores

• Opus 4.6 → 55.4
• GPT-5.3 Codex → 56.0
• MiniMax 2.5 → 55.4
• Opus 4.6 + WarpGrep → 57.5 (+2.1)
• Codex 5.3 + WarpGrep → 59.1 (+3.1)
• MiniMax 2.5 + WarpGrep → 57.6 (+3.7)

Efficiency

Tokens:

• Input tokens: ~17% less (Opus, Codex, MiniMax, etc..)
• Output tokens: 13% less (fewer turns)

Turns (MiniMax):

• Avg turns per task: 157 → 135

Runtime (Opus):

• Opus 4.6: 518s
• Opus 4.6 + WarpGrep: 545s
• 178s saved (12% faster)

Cost (Opus):

• Opus 4.6: $3.06
• Opus 4.6 + WarpGrep: $2.51
• $0.55 saved (15.6% cheaper)

TLDR

WarpGrep v2 raises SWE-Bench Pro scores (+2.1 Opus, +3.7 MiniMax) while using ~17% fewer input tokens, 13% fewer turns, running 12% faster, and costing 15.6% less. The median WarpGrep codebase search takes 5s end to end compared to 75s for the Claude Code Explore subagent

On normal usage on production repos these numbers improve across the board ~40% faster

install it in your favorite coding agent here https://www.morphllm.com/mcp

or integrate it into your product here https://docs.morphllm.com/sdk/components/warp-grep/tool