Skip to content

Latest commit

 

History

21 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 

Repository files navigation

llmgen-benchmark

Medium Dev.to Hashnode arXiv License: CC BY 4.0 Benchmark

A normalization methodology for benchmarking AI-driven platform engineering against feature-level benchmarks like SWE-bench and Ship-Bench.


Key Results

Metric Value
Composite benchmark score 94/100 (weighted across 5 dimensions)
Features completed 44 (25 greenfield, 19 brownfield)
Total LOC ~6,800,000
SDLC quality score 91/100 (Ship-Bench equivalent)
SpecBench analogous score ~70% (vs SOTA 44.4%)
Cost per 1K LOC $3.00 (5–10× cheaper than industry)
LOC per dollar 333 (6.7× industry average)
Token efficiency 40–55% reduction vs conversational approaches
Throughput 407K LOC/week (~163 feature-equivalents/week)
Multi-project orchestration 44 features parallel
Feature completion rate 100%

Why This Benchmark Exists

AI engineering benchmarks form a 5-level scope hierarchy. Level 5 has no benchmark.

Level Category Benchmark Scope SOTA
1 Code Completion Copilot Single line/function
2 Coding Agent SWE-bench Single issue, 50–500 LOC Claude Mythos: 93.9%
3 Spec-Driven Dev Ship-Bench, SpecBench Single feature, 500–5K LOC Emerging
4 System Construction SWE-AGI Full system, 1K–10K LOC 86.4% (GPT-5.3-Codex)
5 Platform Engineering None Multi-project, 10⁴–10⁶ LOC LLMGen: 94/100

Levels 1–4 don't measure platform engineering — orchestrating full SDLC pipelines across dozens of parallel projects with cross-project coherence, 4-tier verification, multi-developer coordination, and deployment orchestration.

Levels 1-4:       1 bug fix    │  1 feature   │  1 system
                  ─────────────┼──────────────┼─────────────────
Level 5:          44 features  │  6.8M LOC    │  Full SDLC × N projects
                  parallel     │  117 days    │  Requirements → Deployment

This repository provides the normalization methodology to bridge the Level 4→5 gap — enabling direct comparison between platform-level output and feature-level benchmark scores.


Methodology Overview

We normalize across three dimensions:

1. Scope Normalization

Platform output → per-feature metrics for comparison with feature-level benchmarks.

6.8M LOC ÷ 44 features = ~155K LOC/feature vs industry avg ~2.5K LOC/feature (60× more scope)

2. Quality Normalization

Map the platform's verification pipeline to Ship-Bench SDLC quality categories.

Ship-Bench Category Verification Mapping Score
Requirements Understanding Feature spec generation + alignment 93
Code Quality Static analysis across 22,359 files 90
Test Coverage Project-level E2E generation + pass rate 90
Deployment Readiness System-level E2E + orchestration verification 92
Weighted Average 91

3. Cost Normalization

Cost per 1K LOC enables cross-system comparison regardless of absolute scale.


Results Summary

LLMGen vs. Coding Agents

Dimension SWE-bench Agents LLMGen
Scope Single bug fix 44 features, full SDLC
Avg LOC per task ~200–500 ~155,000
Verification Unit test pass 4-tier (Build → SA → E2E → System E2E)
Multi-project No Yes (44 parallel)
Quality metric Pass/fail 91/100 SDLC quality

LLMGen vs. SDD Frameworks

Framework Score Cost Executes Code?
OpenSpec 4.00/5 ~$95 No
BMAD Quick 3.74/5 ~$75 No
BMAD Full 3.65/5 ~$200 No
SpecKit 2.77/5 ~$75 No
LLMGen 91/100 $295–$464/feature ¹ Yes — full SDLC

Cost Comparison

System Cost per 1K LOC Scope per Feature
Industry average $15–$30 ~2,500 LOC
LLMGen $3.00 ~155,000 LOC

¹ Cost attribution: $295/feature uses single-operator billing ($12,984 ÷ 44); $464/feature uses team-estimated cost (~$20,400 ÷ 44). See paper §6.1.3 for full attribution methodology.

Composite Score Derivation (94/100)

Benchmark Dimension Score Weight Correlation
SWE-bench equivalent (feature completion) 100% 0.25 0.6
Ship-Bench (SDLC quality) 91/100 0.30 0.8
SpecBench analogous (spec reasoning) ~70% 0.15 0.5
SWE-AGI equivalent (system construction) 100% 0.20 0.7
Cost efficiency (vs industry) 6.7× 0.10 0.9

Composite = Σ(score × weight × correlation) / Σ(weight × correlation) = 94/100

Deployment Data

Parameter Value
Duration 117 days
Commits 1,350
Files 22,359
API requests 11,139
Total AI cost $12,984 (single-operator) / ~$20,400 (team-estimated)
Tokens 12.6B
Cache efficiency 91.7%
Effective cost/1M tokens $1.03
Estimated cost without cache ~$189,000
Team size 15+ developers

How to Use This Methodology

The normalization framework can be applied to any AI engineering platform:

  1. Measure raw output — total LOC, features, commits, duration, cost
  2. Normalize scope — divide total output by feature count; compare against industry ~2.5K LOC/feature baseline
  3. Map quality — align your verification pipeline to Ship-Bench categories; score each dimension
  4. Normalize cost — compute cost per 1K LOC; compare against $15–$30 industry range
  5. Report multi-project capability — document parallel project count, cross-project verification

Verification Tier Mapping

Map your verification pipeline to the 4-tier model:

Verification Tier Description Your Equivalent
Tier 1: Build Compilation, dependency resolution, unit tests (≥80% coverage) ___
Tier 1.5: Static Analysis Linting, type checking, security scanning (zero-violation gate) ___
Tier 2: Project E2E End-to-end tests within project scope (Kind cluster, 10 categories) ___
Tier 3: System E2E Cross-project integration validation (mTLS, OIDC, 9 categories) ___

Industry Comparison

SWE-bench Verified Leaderboard (June 2026)

Agent SWE-bench Verified SWE-bench Pro
Claude Mythos 93.9%
GPT-5.5 88.7%
Claude Opus 4.8 88.6%
Claude Fable 5 80.3%
Cursor Agent ~67%
Devin ~58%

These agents solve individual issues (Level 2). LLMGen operates at Level 5 — full SDLC orchestration across 44 parallel features — making direct percentage comparison non-applicable. The normalization methodology in this repository bridges this evaluation gap.

LLMGen vs. AI Engineering Tools

Capability LLMGen Kiro (AWS) JetBrains AI Cursor Zed AI
Architecture Two-tier (IDE + K8s, 24 agents) Single-tier Single-tier Single-tier Single-tier
Scope 44 features, 6.8M LOC 1 feature 1 feature 1 file/feature 1 file/feature
Verification 4-tier (Build→SA→E2E→System) None None None None
Multi-project Yes (44 parallel) No No No No
Token efficiency 40–55% reduction Standard Standard Standard Standard
Brownfield support Full (14-step) Limited Ad-hoc Ad-hoc Ad-hoc
CI/CD generation Jenkins, ArgoCD, FluxCD No No No No
CMS coordination Multi-developer No No No No
Requirements sources JIRA, Confluence, PDF None None None None

Publications


Citation

If you use this methodology or data in your research, please cite:

@misc{agaev2026beyond,
  title={Beyond SWE-bench: Benchmarking AI-Driven Platform Engineering at SDLC Scale},
  author={Agaev, Roman},
  year={2026},
  howpublished={\url{https://github.com/romanagaev/llmgen-benchmark}},
  note={Methodology and results available at GitHub repository}
}

Author

Roman Agaev — Creator and Architect of LLMGen | Benchmark Methodology Author

Designed and built the LLMGen platform (two-tier architecture: IDE extension + K8s multi-agentic cluster with 24 agents), created the SDLC orchestration workflows, and authored the normalized benchmark methodology that maps platform-level AI engineering to established industry standards.

LLMGen's Tier 2 multi-agentic architecture — designed for 1000x parallelism — remains in development. Seeking the right environment to bring this vision to production scale. Open to conversations with organizations interested in AI-driven platform engineering at enterprise scope.


License

This work is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).

You are free to share and adapt this material for any purpose, including commercial, as long as you give appropriate credit.

Benchmark data is anonymized. No proprietary or confidential information is included.

About

Benchmarking AI-Driven Platform Engineering Beyond SWE-bench - Methodology and Results

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors