Skip to main content

RepoContextBench v1

Repository research, measured against source-backed answers.

A public benchmark for agents answering practical questions about a real codebase. Compare source-grounded answer scores, speed, and price, then inspect the matched CodeAlive result behind the headline.

Lines of code
512.95K
Questions
20
Gold claims
82

512,954 source lines across 4,181 files, measured with cloc at the pinned agent-framework commit. Blank and comment lines are excluded.

Model comparison

Score, speed, and price on one screen

Each model appears once, using its highest-scoring complete run with a reported or estimated answerer price. Score is a 0–100 measure of source-grounded answer quality based on weighted gold-claim coverage, faithfulness, and evidence use.

Score vs. Cost per task

CodeAlive contextRepository toolsHarness:CodeAlive agentClaude CodeCodex CLI
ModelConfigurationHarnessScoreAvg time / taskCost / runCost / taskTotal tokens
Claude Fable 5Repository tools · maxClaude Code80.785.3s$14.08$0.703.59M
Qwen 3.6 35B A3BCodeAlive · deep · highCodeAlive agent77.660.7s$0.21$0.01051.95M
Claude Opus 4.8Repository tools · max · no subagentsClaude Code77.573.6s$7.71$0.394M
Claude Sonnet 4.6CodeAlive skill · maxClaude Code72.067.7s$3.90$0.194.44M
GPT-5.4 miniRepository tools · xhighCodex CLI70.0168.9s$3.12$0.1618.33M
Gemini 3.5 FlashCodeAlive · standard · mediumCodeAlive agent66.138.7s$2.93$0.152.39M
Gemini 3.1 Flash LiteCodeAlive · deep · mediumCodeAlive agent53.28.9s$0.26$0.01301.19M

Only complete 20-task runs with zero run-health failures and an available answerer price are shown. One best Score is retained per model; all no-price rows are excluded. Judge time, tokens, and cost are excluded. Prices are historical reported values or estimates, not quotes.

CodeAlive impact

CodeAlive raised Sonnet's Score by 6.5 points.

The matched comparison used the same Sonnet 4.6 model, 20 questions, maximum reasoning effort, and fixed judge contract. Only the context path changed.

With CodeAlive, Score increased from 65.45 to 71.95.

Claude Sonnet 4.6 max

20 of 20 tasks · fixed judge contract

+6.5 points

Repository tools

65.5

With CodeAlive

72.0

Score runs from 0 to 100 and combines weighted gold-claim coverage, faithfulness, and evidence use. Higher is better.

How to read the numbers

A narrow, auditable repository-research test

20 practical questions

The public dataset covers architecture, workflows, persistence, streaming, observability, fault tolerance, and calibrated abstention in one pinned repository snapshot.

A source-grounded Score

Answers are checked against 82 weighted atomic claims and 99 exact evidence spans. Score combines claim coverage, faithfulness, and evidence use instead of relying on exact text matching.

Static research only

Agents cannot use the internet, run tests, execute or modify code, or verify runtime behavior. The benchmark measures how well they research the repository as written.

The Sonnet comparison is a complete archival matched pair from the same 20-task line, rejudged under the fixed publication judge contract. It predates the final frozen dataset hash; the main leaderboard otherwise uses the frozen v1 publication rows.

Next dataset

RepoContextBench is expanding beyond a half-million-line repository.

We are working on the next RepoContextBench release. It will add a dataset built on a codebase with more than one million lines of code, so retrieval quality is tested under a meaningfully larger search space.

Read the v1 report