Technology / WORLD

LightOn reports lower search-token use with ColGrep-Agent

The company’s coding-search test found a token saving alongside lower observed accuracy and longer run time.

Reading sizeSources ↓Share story ↗
Conceptual illustration of technology, processors and connected information.Illustration
Symbolic illustration

LightOn researchers introduced ColGrep-Agent-2B on October 12, a local model that searches code repositories and returns file locations and excerpts to a larger coding agent. In the team’s test, Claude Code with Sonnet 5.5 handled 52 code-search questions across 13 repositories, with each question run three times in four setups. The delegated setup used 93,000 API tokens per question on average, compared with 142,000 for grep.

LightOn reports 34.3% fewer API tokens and 40% lower cache-aware API cost, but observed correct answers were 80.8% versus 83.3% for grep, and average time was 21.6 seconds versus 13 seconds. The accuracy difference was not statistically significant, but the reported interval does not rule out a meaningful drop. The results cover code-search tasks, not complete coding work, and are the authors’ own measurements.

Data

Timeline

Announcement summary

The company’s coding-search test found a token saving alongside lower observed accuracy and longer run time.

  1. · Occurred

    LightOn reports lower search-token use with ColGrep-Agent

    The company’s coding-search test found a token saving alongside lower observed accuracy and longer run time.

Methodology

This event record summarizes the source announcement and adds no independent verification or generalization beyond it.

Updated:

Structured coverage

Subjects and places

Subject proposals

Geographic references

Only source-provided geographic references are shown. Opening an external map is optional. Optional story maps use OpenStreetMap. Opening a map sends your IP address and this site’s address to the map provider; location data remains readable as text without loading the map.

  • France · Place role: Coverage · Place type: Country or area · Location precision: Regional
    Evidence
    • Explicit preserved editorial country/region; no city or national centroid inferredhuggingface.co ↗
Taxonomy version: IPTC Media Topics 2026-Q2

Continue reading

Technology

AgentSWE benchmark tests what coding agents actually deliver

The arXiv preprint defines 25 tasks across agent creation, editing existing code and optimization toward a target.

arXiv preprint by researchers affiliated with Beijing Academy of Artificial Intelligence, Institute of Automation of the Chinese Academy of Sciences and Renmin University of China1 min read
Data included
Technology

SWE-Critix trains agent critics for code repair

The method trains a code agent to critique repair trajectories using differential reasoning and test rewards.

arXiv preprint by researchers affiliated with Huawei and Monash University Malaysia1 min read
Data included