Openbenchmarks
for Agents

Verified benchmarks for Agents to make build vs buy decisions.

Verified ground truth · open data · public methodology

Benchmarks.

lookalikes

Lookalike

24 seed companies × 4 vendors. Each vendor returns its top 100 lookalikes per seed; an LLM judge labels each result relevant or not. Cell value is Precision@100 - of the K you paid for, how many are usable.

top Precision@100
57%
open benchmark
open data + code · github.com/openbenchmarks-labs/lookalikes
tts · by coval

Text-to-Speech

24 text-to-speech models on Time to First Audio (TTFA) + Word Error Rate, measured by Coval under production-realistic conditions. Mirrored with attribution.

fastest TTFA (median)
212 ms
open benchmark
stt · by coval

Speech-to-Text

31 speech-to-text models on Time to Final Segment (TTFS) + Word Error Rate, measured by Coval under production-realistic conditions. Mirrored with attribution.

fastest TTFS (median)
46 ms
open benchmark

Built to help an Agent evaluate.

01

Ground truth, not vendor decks

Every input in the dataset has a correct answer we verified ourselves. Row by row.

02

Open methodology

Dataset, scoring rules, and the exact plan we billed each provider on.

03

Metrics that Agents care about

How often a provider got it right, how often it got it wrong, what each correct answer cost.

04

The dataset never leaves

Real data goes in. Only aggregate provider scores come out.