← @bgSponge

Grep achieves SOTA on major deep research benchmarks

Summary

Grep achieves state-of-the-art performance across three independent deep research benchmarks, outperforming systems from OpenAI, Google, Anthropic, and Perplexity.

  1. Architecture drives research performance.

    Grep outperforms Claude Opus 4.6 standalone by 18.8 percentage points on the same model, proving the delta comes from the research architecture.

    Same model. Radically different results. That delta is architecture.

  2. Planning prevents retrieval errors.

    The planner calibrates effort to the task to avoid excessive retrieval, which can hurt factual-correctness scores in DeepSearchQA.

    Excessive retrieval can actually hurt factual-correctness scores.

  3. Language tuning is unnecessary.

    Grep scored 56.42 on Chinese queries and 56.13 on English despite zero optimization for Chinese, demonstrating architectural robustness.

    When the architecture is right, you don't need to tune for every language separately.

  4. One agent replaces committees.

    Instead of negotiating specialized agents, Grep uses a single agent that morphs, loads skills, and spawns sub-agents via cell division.

    Think cell division, not a committee.

  5. Grep Brain enables persistence.

    The new Grep Brain architecture transforms the tool into a system that remembers work across sessions and produces deliverables like spreadsheets and reports.

    transforms Grep from a research tool into a persistent system

Highlights

  • In Grep, effort doesn't just mean doing more work - it means being more focused. Excessive retrieval can actually hurt factual-correctness scores.

  • Grep is one agent that reshapes itself across stages of research - and when the work demands it, splits into parallel copies of itself.

Read the original ↗Save to your Sponge

Summarized with Sponge. Quotes are the author’s own words.

More from @bg

All readings →