82.8LEADING SCORE / 100

PERFORMANCE, UNDER TEST.

Writing load tests is easy.
Getting them right is the benchmark.

An open benchmark for AI on performance test scripting and engineering.

#1 overallDeepSeek V4 Pro
$0.0367estimated run cost
9 / 12tasks passed
REAL RUNS. NO SYNTHETIC RANKINGS.2 models · 12 tasks · 1 executable tool17 NOV 2026 / SEED PACK 01

01 / THE LEADERBOARD

Follow the flame.

WIDER IS BETTER ↗

Every segment is earned. Explore where each model picked up its points.

authoring 25correlation 15workload 15analysis 15repair 15 · not yet benchmarkedmigration 10 · not yet benchmarkedconcepts 5
  1. 01DeepSeek V4 Prodeepseek / direct API
    82.8
  2. 02DeepSeek Flashdeepseek / direct API
    80.0

Points = category score × weight ÷ present weights. Five active categories; 75 total weight. Hover, focus or tap a segment.

02 / THE COST FRONTIER

More score. Less spend.

A full seed run, priced in dollars. Up and left wins.

Score versus estimated run costDeepSeek V4 Pro: 82.8 points at $0.0367. DeepSeek Flash: 80.0 points at $0.0128. Both models are on the cost frontier.0255075100$0.00$0.02$0.04SCORE / 100ESTIMATED USD / RUNDeepSeek V4 ProDeepSeek Flash

A SMALL BUDGET. A REAL CHOICE.

What's your ceiling?

Find the highest score within your per-run budget.

DeepSeek V4 Pro

82.8 points · $0.0367 / run

Two models. One run each. A starting signal, not a final verdict.

03 / UNDER THE HOOD

No place to hide a miss.

12 TASKS / PASS OR FAIL

The same task, the same prompt. Focus or tap a result to inspect its grading detail.

Seed task results by model, grouped by category
Category / taskDeepSeek V4 ProDeepSeek Flash
authoring1 tasks
shop-flowjmeter
correlation2 tasks
session-isolationjmeter
token-refreshjmeter
workload4 tasks
arrivalreasoning
concurrencyreasoning
little-lawreasoning
pacingreasoning
analysis2 tasks
errorsreasoning
meanreasoning
concepts3 tasks
closedreasoning
correlationreasoning
percentilereasoning

Repair and migration: not yet benchmarked. Results are binary, with no partial credit.

04 / PICK YOUR INSTRUMENT

Built for the tools you use.

Executable plans first. More engines next.

JMeter

Real scripts.
Real checkouts.

Executable task pass rate

  1. DeepSeek V4 Pro66.7
  2. DeepSeek Flash66.7

k6

Next in
the pipeline.

Graders coming

Gatling

Next in
the pipeline.

Graders coming

Locust

Next in
the pipeline.

Graders coming

JMeter score = percentage of executable tasks passed. No scores are assigned to tools without graders.

05 / SHOW YOUR WORK

A benchmark.
Not a black box.

Explore the harness on GitHub ↗

Direct provider APIs. Temperature 0. One fixed system prompt and identical task prompts for every model. No routing layer.

Generated JMeter plans run against a loopback demo shop. Passing requires zero sample errors and verified session-aware checkouts. Numeric and choice tasks use deterministic answer keys.

Category means are weighted, then renormalized over categories with tasks. Unmeasured categories stay unmeasured, never zero.

authoring25correlation15workload15analysis15repair15migration10concepts5

Cost is estimated from published peak list prices, including cache-hit pricing. It is not an invoice. Latency measures provider calls, excluding grading.

This public seed pack is small: one run per model, not a confidence interval. DeepSeek thinking is disabled. Output cap: 8,192 tokens. Tool scores and overall scores measure different slices.

THE PAPER TRAIL

DeepSeek V4 Pro2026-11-17 / 10:02:15 UTC10,588 tokens · 91.61s API latencyRun artifacts ↗Harness 7383b4e
DeepSeek Flash2026-11-17 / 10:01:23 UTC12,072 tokens · 43.29s API latencyRun artifacts ↗Harness 7383b4e