Skip to content

How it's built

Methodology & sourcing

One honest, coding-only composite — and a clear paper trail for every number.

✓ Verified data — every figure is cited. See sources.
Last updated Jul 9, 2026

Vibe Coding Index weights

Each suite is normalised to 0–100 and combined with the weights below. We weight agentic, real-repo work highest — that's what an AI coding harness does all day. When a model is missing a suite, we renormalise over the suites it does have instead of inventing a zero.

  • SWE-bench Verified
    40%
  • Terminal-Bench
    30%
  • LiveCodeBench
    30%

SWE-bench Verified

40%

Real GitHub issues fixed end-to-end.

metric: resolved · scale 0–100

Terminal-Bench

30%

Agentic tasks in a real shell.

metric: resolved · scale 0–100

LiveCodeBench

30%

Fresh competitive-programming problems.

metric: pass@1 · scale 0–100

Data & citations

This dataset is verified. Every suite score links to its source (SWE-bench, LiveCodeBench, Aider, Terminal-Bench leaderboards), and model-level pricing/speed cite the vendor or Artificial Analysis. 42 suite figures currently carry a source link.

Anything we can't verify is left blank (—). We never fill a gap with a guess.

Where live data plugs in

The seeder reads a single verified file — .ovibe/knowledge/benchmark-data.json — verbatim, with per-figure source URLs. Swapping in a live benchmark-API feed means writing that same JSON on a schedule; no schema or UI changes. Suites, pricing, speed and latency all live in benchmark_models / benchmark_scores.

Primary sources

Snapshot: Jul 9, 2026

Each suite is pulled from a single independent leaderboard where one exists; model-level pricing and throughput cite the vendor and Artificial Analysis. Where no independent run exists yet, the vendor's self-reported figure is used and flagged in the caveats below.

SWE-bench Verified

Real GitHub issues resolved end-to-end. Independent runs from the Epoch AI Benchmarking Hub; a few 2026 frontier models are vendor self-reported (via llm-stats.com) until Epoch reruns them.

Epoch AI Benchmarking Hub ↗

Terminal-Bench

Agentic tasks in a real shell. Best public entry per model from the Terminal-Bench board. Note: 2.0 and 2.1 runs are mixed and are not directly comparable (2.0 scores higher).

tbench.ai leaderboard ↗

LiveCodeBench

Fresh competitive-programming problems (pass@1), single independent harness.

vals.ai · LiveCodeBench ↗

Pricing, speed & latency

Token pricing and context from each vendor own docs; output speed (tok/s) and time-to-first-token from Artificial Analysis.

Artificial Analysis ↗
Caveats we won't hide
  • · Terminal-Bench mixes 2.0 and 2.1 entries across models — read it as directional, not head-to-head.
  • · Independent SWE-bench runs typically land 3–5 points below vendor self-reports; mixed provenance is noted per figure.
  • · Anything we can't verify from a primary source stays blank (—). We never fill a gap with a guess.