Code-first, AI-first.
Because everything is code, Gatling handles AI systems the same way it handles everything else. It writes your tests, runs them against real systems, and tells you what happened, all in the workflow you already use.
Test-as-code is winning the AI race.
AI reads and writes real code, not a proprietary binary file or a recorded UI flow.
AI is making test-as-code itself easier to write, closing the gap even further.
Test AI systems the way they behave in production
AI-powered applications don't fail the way traditional APIs do. They're slow on purpose, they hold state, every call costs money, and a lot of them run over long-lived streaming connections. Most load testing tools were built for a different shape of problem. Gatling was built for this one.
The result: you ship AI applications that hold up in production, and you find out what they cost before your users do.
ScenarioBuilder llmStreaming = scenario("LLM Streaming (SSE)").exec( sse("Stream chat completion") .post("/chat/completions") .header("Authorization", "Bearer " + apiKey) .body(StringBody(""" { "model": "claude-sonnet-5", "stream": true, "messages": [{ "role": "user", "content": "..." }] } """)) ).asLongAs(session -> !session.contains("done")).on( sse.processUnmatchedMessages((messages, session) -> messages.stream().anyMatch(m -> m.message().contains("message_stop")) ? session.set("done", true) : session ) ), sse("close").close() );
The first autonomous load tester agent
Teams don't skip load testing because they stopped caring about performance. They skip it because writing the tests is slow, reading the results takes a specialist, and every API change breaks what they already wrote. We're building an autonomous load testing agent, brick by brick, and those three are where we started.
Each brick removes a reason teams skip load testing. Together, they add up to an autonomous load testing agent: one that designs the right test, runs it, reads what happened, and keeps your suite correct, without anyone opening a ticket.
Build performance tests at the speed of development
Gatling thinks AI-first to simplify the developer workflow, from writing your first test to keeping it correct as your system evolves.
What we offer for scripting:
The result: tests get written faster, and more people on the team can write them, not just the two who know how.
- ✓ Registered source repository: github.com/acme/checkout-service
- ✓ Created build-from-sources test: "Checkout Load Test"
- ✓ Started run from Paris location
AI turns your results into decisions at every level
A single run tells you what happened today. A history of runs tells you whether your system is getting better or worse. A comparison between runs tells you exactly what changed after a deployment.
AI Analysis covers:
The result: Performance testing becomes something you do continuously, instead of something you get to when there's time.
Health
Insights
Recommendations- Mean = 210ms vs P95 = 340ms shows a moderate right tail, consistent with normal AI-app latency variance.
- P95 ≈ P99 (340ms vs 355ms) indicates no extreme outliers in this run.
- Standard deviation = 48ms indicates stable performance under the tested concurrency.
- No timeouts observed, but duration and volume are too small to detect saturation under sustained load.
- Error ratio = 0.0% across 500 requests, 0 failed responses.
- P95 latency assertion (<500ms) passed on every run.
- 50 concurrent users, constant load, no ramp-up phase configured.
- P95 = 340ms, P99 = 355ms, tight tail, low dispersion.
- Configure a 5-minute ramp-up phase before sustained load to observe stabilization.
- Increase concurrency to 100-200 virtual users to exercise contention and reveal tail latencies.
- Add SLA-based assertions matching production requirements (P95 target, error rate threshold).
- Collect downstream service metrics (DB, cache, LLM API) to correlate latency spikes with resource events.
Tests that survive your API changes
Broken tests are the biggest hidden cost in load testing, but not just writing them, maintaining them. We're working toward tests that point straight to the fix, not just the failure.
After a load test reveals a regression, AI traces it back to the code most likely responsible and recommends what to change, so the fix starts in your codebase, rather than in your test suite.
Want to learn more about how Gatling brings AI into load testing?
Your all-in-one load testing platform
Design complex tests, manage global infrastructure, and turn results into action on one powerful platform.
Need technical references and tutorials?
Minimal features, for local use only
