Load testing your HTTP service from code: k6, NBomber, Goose and pepe
k6 is our Pick for planned arrivals and CI failure rules, with a narrow lead over NBomber. Consider NBomber for .NET, Goose for Rust or pepe for configured request flows.
k6
k6 leads NBomber narrowly, 4.175 to 3.95. Arrival-rate executors keep starting work while the service slows, dropped work is counted, and thresholds on latency, errors, checks and dropped work set the exit code.
JavaScript but not Node.js, so Node-only packages need rewriting. A failed check alone does not fail the run: add a threshold. AGPL-3.0.
Why k6 is our PickTeam behind k6? Get your award badge
No Rising Star selected
Our November 1 review of the candidates in our research did not establish every requirement for a newcomer award.
Directory pages were not readable, so discovery was limited to the named candidates and their public product records.
Who is this article for? Developers and test engineers who want to load test their own HTTP service from code, on their own machine and in CI, and need the test to fail when the service misses its targets.
About 22 min read
What Preferred Pick, Rising Star and Ready to Rock mean
The quick list
k6 is our Pick, narrowly ahead of NBomber. Compare the four runners for traffic scheduling, CI failure rules and local results; choose the code or flow model your team can maintain.
Local k6, Goose and pepe are free software. For our two-author commercial team, NBomber Business was $1,500 a year per company on November 1, 2026: an estimated $4,500 over three years if the current price stays unchanged. Personal is free for personal use only. Existing machines and CI capacity add no new cash charge in this scenario; new infrastructure and staff time depend on your setup. Optional hosted services are separate.
The comparison at a glance
Compare how each runner schedules traffic, decides pass or fail in CI, keeps results and scales beyond one machine.
k6 is our Pick with a narrow lead over NBomber. Order follows the scores for planned-rate HTTP tests and CI failure rules; it is not a capacity or speed ranking. All tools remain selectable.
Hides matching answers for the selected tools. Unknowns and unverified claims stay visible.
| Feature | k6 | NBomber | Goose (Rust) | pepe |
|---|---|---|---|---|
| Best fit | Preferred Pick: a planned arrival rate with pass/fail rules in CI, run from your own machine; narrow lead over NBomber | Also consider: C# or F# journeys in a .NET test project | Also consider: HTTP journeys that reuse Rust code | Also consider: local request flows configured in a CLI |
| Workload scheduling | Closed user-count executors and open arrival-rate executors | Open simulations such as Inject; closed user count with KeepConstant | User-count scenarios; global open arrival scheduler not established | Rate-controlled starts with a concurrency cap; flow behavior depends on the pinned version |
| When the planned load is not reached | dropped_iterations counts iterations that could not start; it can carry a threshold | No k6-equivalent dropped-work gate established in the sources we read | No dropped-iteration metric established in the sources we read | Compare configured limits and report results; no achieved-rate measurement by us |
| Pass/fail rule for CI | Thresholds on latency, failure rate, checks or dropped iterations exit non-zero | Runtime thresholds record failures; documented xUnit assertion makes CI fail | Threshold failures exit 1 when main propagates execute() errors; early abort can be disabled | --fail-if returns exit 4 when a limit is crossed |
| Code reuse boundary | Not Node.js: packages needing fs or os fail; bundle other dependencies | Ordinary .NET code in scenarios; proprietary NBomber APIs | Rust code and libraries in scenarios | Flow configuration rather than a general-purpose script runtime |
| Results kept on your machine | Web dashboard with self-contained HTML export; end-of-test summary; JSON and CSV outputs | Local HTML, CSV, TXT and Markdown | Local HTML and JSON; optional PDF documented | Released README promises versioned JSON and terminal output; HTML availability conflicts with roadmap |
| Running across machines | k6 Operator on a Kubernetes cluster you run | Distributed feature availability depends on the commercial edition | Gaggle removed in 0.17; do not count a future restoration | Distributed runs are a roadmap item, not a current promise |
| Selected cost: local use; NBomber checked November 1, 2026 | Free local runner; optional Grafana Cloud priced separately | Business $1,500/year per company; three-year estimate $4,500 at current price | Free local software; no hosted plan selected | Free local software; no hosted plan selected |
k6 is our Pick for code-defined HTTP load tests of your own authorized service, with a narrow lead over NBomber (4.175 to 3.95). NBomber suits .NET code, Goose Rust, and pepe configured flows. Public-source research: k6 checked October 19, 2026; additional runners November 1. No installs, generated traffic or capacity measurement by us. Local k6, Goose and pepe are free software; commercial NBomber Business is $1,500/year per company, checked November 1, 2026. Compare planned and completed work, and confirm that a wrong response fails CI. Next substantive review: January 17, 2027, or sooner after a major release.
All features and sources 18 criteria
Open the groups that matter to you. They compare the same selected tools. “Planned” means future work, not an available feature. Sources include the published conditions.
Workload and pass/fail rules
Show 9 criteria
| Feature | k6 | NBomber | Goose (Rust) | pepe |
|---|---|---|---|---|
| Best fit | Preferred Pick: a planned arrival rate with pass/fail rules in CI, run from your own machine; narrow lead over NBomber | Also consider: C# or F# journeys in a .NET test project | Also consider: HTTP journeys that reuse Rust code | Also consider: local request flows configured in a CLI |
| Authoring | JavaScript or TypeScript; types are stripped, not checked | C# or F#; JSON configuration for common settings | Rust scenarios and transactions | CLI options and request-flow configuration |
| Code reuse boundary | Not Node.js: packages needing fs or os fail; bundle other dependencies | Ordinary .NET code in scenarios; proprietary NBomber APIs | Rust code and libraries in scenarios | Flow configuration rather than a general-purpose script runtime |
| Workload scheduling | Closed user-count executors and open arrival-rate executors | Open simulations such as Inject; closed user count with KeepConstant | User-count scenarios; global open arrival scheduler not established | Rate-controlled starts with a concurrency cap; flow behavior depends on the pinned version |
| What a rate counts | Iterations of your function, not individual requests | Injected scenario starts per interval, or maintained user count | Scenario users; requests depend on transaction code | Rate option documented per second; flows contain multiple requests |
| When the planned load is not reached | dropped_iterations counts iterations that could not start; it can carry a threshold | No k6-equivalent dropped-work gate established in the sources we read | No dropped-iteration metric established in the sources we read | Compare configured limits and report results; no achieved-rate measurement by us |
| Response checks | check() records pass or fail; on its own it does not fail the run | Scenario and step assertions in C# or F# | Response validation in Rust transaction code | Flow logic and error/latency limits documented; verify your response rules |
| Pass/fail rule for CI | Thresholds on latency, failure rate, checks or dropped iterations exit non-zero | Runtime thresholds record failures; documented xUnit assertion makes CI fail | Threshold failures exit 1 when main propagates execute() errors; early abort can be disabled | --fail-if returns exit 4 when a limit is crossed |
| Multi-step journeys | Scripted request sequences in JavaScript | Programmed scenario steps with an HTTP client | Rust scenario/transaction code with an HTTP client | Flows pass values from one response into later requests |
These selected tools have matching answers in this group. Turn off “Show differences only” to see them.
Sources for workload and pass/fail rules
k6 documentation checked October 19, 2026: models, dropped iterations, thresholds, modules. Additional runners checked November 1, 2026: NBomber source, simulations, thresholds and assertions; Goose book, thresholds; pepe released README, reference. Documented behavior, not a measured workload comparison.
Results, scaling and adoption
Show 9 criteria
| Feature | k6 | NBomber | Goose (Rust) | pepe |
|---|---|---|---|---|
| Results kept on your machine | Web dashboard with self-contained HTML export; end-of-test summary; JSON and CSV outputs | Local HTML, CSV, TXT and Markdown | Local HTML and JSON; optional PDF documented | Released README promises versioned JSON and terminal output; HTML availability conflicts with roadmap |
| Running across machines | k6 Operator on a Kubernetes cluster you run | Distributed feature availability depends on the commercial edition | Gaggle removed in 0.17; do not count a future restoration | Distributed runs are a roadmap item, not a current promise |
| Platform qualification | Scripts importing k6/x extensions fetch a built binary by default; can be turned off for CI without network access | .NET test project; preserve the assertion that checks threshold failures | Rust toolchain and project configuration required | Windows, macOS and Linux binaries documented with checksums and provenance |
| Release evidence: k6 October 19; other runners November 1, 2026 | 2.3.0, September 21, 2026; 1.8.2 security update for the 1.x line, October 12, 2026 | 6.7.1, checked November 1, 2026 | 0.19 series, checked November 1, 2026 | 0.57.0 release README used; default branch advertises newer features |
| Change to plan for | 2.0 (May 11, 2026) removed old commands, the externally-controlled executor and the legacy summary format | Commercial use requires Business; keep test code and raw reports | Thresholds documented from 0.19; older distributed Gaggle removed | 1.0 flag/report stability contract not yet promised; pin the version |
| License | AGPL-3.0 | Commercial company license; Personal is for personal use only | Apache-2.0 | MIT |
| Selected cost: local use; NBomber checked November 1, 2026 | Free local runner; optional Grafana Cloud priced separately | Business $1,500/year per company; three-year estimate $4,500 at current price | Free local software; no hosted plan selected | Free local software; no hosted plan selected |
| Usage reporting | Anonymous usage report on by default; K6_NO_USAGE_REPORT turns it off | Data handling during execution not audited by us | Data handling during execution not audited by us | Released guardrails documented; data handling during execution not audited by us |
| Moving to another runner later | Keep scripts and results; rewrite runtime-specific helpers | Keep C#/F# logic and reports; replace NBomber-specific APIs | Keep Rust logic and reports; replace scenario APIs | Keep flows and JSON; translate configuration and CI exit handling |
These selected tools have matching answers in this group. Turn off “Show differences only” to see them.
Sources for results, scaling and adoption
k6 evidence checked October 19, 2026: dashboard, Operator, extensions, versioning, source, usage reporting. Other runners checked November 1, 2026: NBomber source, pricing, license, refund policy; Goose source, releases; pepe released README, roadmap. Where current documentation conflicts with released capability, use the more limited description. Prices are vendor statements; no purchase or security audit.
What matters when choosing
k6 is our Practical Pick for a developer or test engineer who wants a load test of their own HTTP service to hold a planned rate of traffic and to fail the CI run when the service misses its targets. Its lead over NBomber is narrow: 4.175 to 3.95 on our five criteria, a gap of 0.225. It is free to run on your own machine. Its arrival-rate executors start new iterations on schedule whether or not earlier ones have finished, so a slowing service does not quietly lower the load. Work it could not start is counted in a dropped_iterations metric, and its thresholds turn rules on latency, error rate, checks and dropped work into a failing exit code. Its web dashboard can save a self-contained HTML report when the run ends. Open and closed models, dropped iterations, thresholds, web dashboard
The most important compromise is that k6 runs JavaScript but is not Node.js. Packages that rely on Node APIs such as fs or os do not work, other dependencies must be bundled, and TypeScript is stripped of its types rather than checked. A second trap catches new users: a failed check() does not fail the run by itself, so you need a threshold on the checks metric. k6 2.0, released May 11, 2026, also removed old commands and one executor; the 1.x line now receives only major and security fixes. Modules, TypeScript, 2.0 release notes, versioning policy
Choose Locust if your test logic belongs in Python. Its simulated users are plain Python classes, a response can be marked failed when its content is wrong, and its HTML, CSV and JSON reports stay on your machine. Its wait times limit how often each user repeats a task; they cannot start extra users, so throughput falls when the service slows. A pass/fail rule on aggregate latency means writing a short Python handler. Consider Artillery for declarative YAML journeys extended with Node.js code, with distributed runs launched into your own AWS account. Its phases start new users at a set arrival rate, and its ensure plugin fails a run on metric rules. Its local HTML report command was removed in v2.0.22, however, which leaves console and JSON output unless you use its Cloud dashboard, and production use on Azure requires a commercial license. Locust tasks and wait times, Locust in CI, Artillery phases, ensure, report removal, repository licensing
These are researched judgments from documentation, release records, license files, pricing pages and public reports. We checked k6, Locust and Artillery on October 19, 2026 and the additional alternatives on November 1, 2026. We did not install a runner, generate traffic or measure capacity. Three conditions would change our recommendation: test logic that must stay in Python (choose Locust), Node.js helpers you cannot rewrite, or distributed runs on AWS without operating a Kubernetes cluster (consider Artillery). Whichever you choose, keep scripts and raw results in your repository, and confirm on a disposable copy of your service that a deliberately wrong response fails the CI job before you rely on it.
Disclosure: Practical Picks articles are researched and written with AI, currently Anthropic’s Claude. Grafana’s official k6 MCP server documents setup for Claude, and pepe documents a Claude Code plugin and MCP server. We have no relationship with Grafana Labs or Anthropic over this article, and these Claude links carried no weight in our judgment. mcp-k6, pepe roadmap
k6
Preferred Pick · a planned arrival rate with CI pass/fail rules
Choose k6 when the test must keep sending work on schedule while the service slows, and CI must fail on latency, errors or wrong responses. Add a threshold on checks, and bundle any dependencies.
Reasons to choose it
- Open arrival-rate executors alongside closed user-count ones
- dropped_iterations shows when the planned load did not run, and can fail the run
- Thresholds on latency, errors, checks and dropped work set the exit code
- Web dashboard saves a self-contained HTML report
What to watch
- JavaScript but not Node.js: packages needing fs or os do not work
- A failed check alone does not fail the run
- 2.0 (May 2026) removed old commands and one executor; pin the version
- Running across machines yourself means operating Kubernetes
k6 offers two ways to schedule work. In its closed model, each virtual user starts its next iteration only after the previous one finishes, so slower responses mean fewer iterations; Grafana’s documentation notes that testing literature calls this problem coordinated omission. The constant-arrival-rate and ramping-arrival-rate executors use an open model instead and start iterations on schedule. When no virtual user is free, k6 drops the iteration and counts it in dropped_iterations, which can carry its own threshold. Check that metric beside latency: a good percentile for the requests that ran does not prove the planned load ran. A rate counts iterations of your function, not individual requests. Open and closed models, dropped iterations
Acceptance rules are thresholds. k6 exits with a non-zero code when a threshold fails, and abortOnFail can stop the run early. A check() only records a pass or a fail; the documentation says checks “do not affect the exit status”. Add a threshold such as checks: ['rate>0.99'] when CI must fail on wrong responses. This is the strongest documented reason to choose k6, and the check-versus-threshold distinction is also a surprise reported in the public discussion we read. Thresholds
Read the full k6 analysis and sources
Code reuse has a boundary. k6 does not resolve node_modules the way Node.js does, so dependencies must be bundled into a self-contained file, and packages that need Node APIs such as fs or os will not work. TypeScript has run by default since v0.57, but k6 strips types without checking them, so run a type checker separately. Remote modules are downloaded and executed at run time. In current releases, a script that imports an official or community k6/x/ extension makes k6 obtain a binary containing it from a build service by default; set K6_AUTO_EXTENSION_RESOLUTION=false and build your own binary with xk6 if your CI must not fetch code during a run. Modules, TypeScript, extensions, 2.0 release notes
Pin the version in CI. k6 2.0.0, released May 11, 2026, removed the externally-controlled executor (scripts that use it no longer start), the pause, resume, scale and status commands, the --no-summary flag and the legacy summary format, and it no longer starts its local REST API server unless you pass --address. Version 2.3.0, on September 21, 2026, added --scenario to run named parts of a script and --once for smoke runs. Grafana states a major release once a year and a minor release every six weeks; the previous major line stays in support mode, with major and security fixes only, and 1.8.2 shipped security updates on October 12, 2026. 2.0.0, 2.3.0, 1.8.2, versioning policy
Results stay local if you want them to. The web dashboard (K6_WEB_DASHBOARD) can write a self-contained HTML report at the end of a run (K6_WEB_DASHBOARD_EXPORT), but it includes graphs only when the test lasts more than three times the aggregation period. JSON and CSV outputs are covered by Grafana’s stability policy. To spread a test across machines yourself, the k6 Operator runs it as a TestRun resource on a Kubernetes cluster you operate and splits the load across pods; scripts kept in a ConfigMap are limited to 1 MiB. We could not establish from Grafana’s documentation whether a threshold is judged across all pods together or separately in each pod, so confirm that before relying on one pass/fail result from a distributed run. Grafana Cloud k6 is an optional hosted service: on October 19, 2026 its pricing page listed 500 virtual-user hours a month free, and a Pro plan at $19 a month including 500 hours, then usage from $0.15 per virtual-user hour. Web dashboard, distributed tests, Grafana pricing
Grafana Labs announced its acquisition of k6 in June 2021, and the project is licensed under AGPL-3.0, which some organizations’ license policies restrict; we give no legal opinion. Each run sends an anonymous usage report by default, containing a random installation ID, the k6 version, virtual-user counts, durations, operating system, imported modules and outputs; K6_NO_USAGE_REPORT or --no-usage-report turns it off. We did not audit the network traffic. Acquisition announcement, repository, usage collection
Grafana’s June 23, 2026 incident report says an attacker copied its repositories after a supply-chain attack. Grafana says customer production systems were not accessed without authorization and Grafana Cloud was unaffected, and it reports that a commissioned Mandiant investigation found no evidence of tampering in public organizations or production repositories delivered to users. We did not see the audit itself. The report, which had no later update when we rechecked it on October 19, 2026, does not mention k6; it is not evidence that a k6 release was compromised. Choose Locust if your team’s test logic is Python, and Artillery if Node.js helpers or AWS-hosted distributed runs decide the job. Post-incident report
Used k6? Share your experience to help other readers choose.
NBomber
Also consider · C# or F# HTTP scenarios
Choose NBomber when HTTP journeys belong in your .NET test project. Keep the assertion on returned threshold failures so CI fails, and budget for commercial licensing.
Reasons to choose it
- C# and F# scenario code
- Open arrival simulations and closed user-count simulations
- Scenario and step metric thresholds
- Local HTML, CSV, TXT and Markdown reports
What to watch
- A recorded threshold failure needs an assertion to fail the CI job
- Personal license is for personal use only
- Proprietary APIs and commercial entitlement affect continuity
- No k6-equivalent dropped-work gate established here
NBomber suits a team that wants to keep HTTP journeys in its .NET test project. Scenarios are ordinary C# or F# code. Its open simulations, including Inject, start work at a rate; closed simulations such as KeepConstant maintain a user count. That gives you both workload shapes, though the documentation we read does not establish a dropped-work gate equivalent to k6’s. Scenarios and HTTP example, load simulations (checked November 1, 2026).
Its threshold API checks scenario and step metrics during execution and can abort after repeated failures. The documented xUnit example reads the returned thresholds and asserts that none failed: keep that assertion in your test, rather than assuming a recorded threshold failure alone fails the CI process. It offers local HTML, CSV, TXT and Markdown reports. Thresholds and assertions, report formats (checked November 1, 2026).
Read the full NBomber analysis and sources
The material tradeoff is commercial licensing and dependence on NBomber’s APIs. On November 1, 2026, Business was $1,500 a year per company, while Personal was free for personal use only. For the two-author commercial team we considered, that is an estimated $4,500 over three years at the current price. NBomber LLC’s refund policy allows a full refund within 30 days; after that sales are final. Cancellation through its portal ends renewal, with access continuing to the end of the paid period. Choose k6 if its JavaScript runtime suits you and you want a free local runner with explicit dropped-iteration thresholds. Pricing, license agreement v3.2, refund and cancellation.
Used NBomber? Share your experience to help other readers choose.
Goose (Rust)
Also consider · journeys that reuse Rust code
Choose Goose for journeys in Rust with a user-count workload. Metric thresholds can fail an unattended run; plan around the removal of its former distributed mode.
Reasons to choose it
- Full Rust code reuse
- Response validation in transaction code
- Threshold exit 1 when main propagates the execute() error
- Local HTML and JSON; optional PDF documented
What to watch
- Global open arrival scheduler not established
- Gaggle distributed mode removed in 0.17
- Rust toolchain and configuration burden
- No achieved-load or capacity measurement by us
Goose lets you build users, scenarios and HTTP request sequences in Rust. Choose it when your test logic already belongs in Rust and a user-count workload is suitable. We did not establish a global open arrival scheduler comparable to k6’s. It is Apache-2.0 software, maintained in public source associated with Jeremy Andrews and Tag1 Consulting. Goose book, source and license (checked November 1, 2026).
From version 0.19, metric thresholds can fail a run with exit code 1 when your Rust main function propagates the error from execute(), as the documented examples do. Rules can judge latency and failure rates; early abort is on by default and can be disabled. This makes an unattended CI gate possible without writing a separate report parser. Local reports include HTML and JSON, with optional PDF output documented by the project. Thresholds, reporting overview (checked November 1, 2026).
Read the full Goose (Rust) analysis and sources
Plan around the current distributed-execution limit: Gaggle was removed in 0.17, and a future restoration is not a feature you can rely on today. Keep Rust scenarios and raw reports so they remain portable. Choose k6 when independent scheduled arrivals and explicit dropped-work rules matter more than Rust reuse. Current source and release notes, releases (checked November 1, 2026).
Used Goose (Rust)? Share your experience to help other readers choose.
pepe
Also consider · local request-flow configuration
Consider pepe for compact response-dependent flows, versioned JSON and explicit CI failure limits. Pin the version and rely on released features.
Reasons to choose it
- Flows pass response values into later requests
- --fail-if returns exit 4 on crossed limits
- Versioned local JSON and terminal results
- Host, rate and request-count guardrails
What to watch
- Flow configuration rather than a general-purpose script runtime
- HTML export documentation conflicts with the roadmap
- Stable 1.0 flag and report contract still ahead
- Distributed execution is a roadmap item
pepe is an MIT-licensed Rust runner whose version 0.57.0 README documents response-dependent request flows, a terminal dashboard, versioned JSON reports and binaries for Windows, macOS and Linux. It suits a compact journey defined as a flow rather than a test written in a general-purpose language. Released README, releases (checked November 1, 2026).
The documented --fail-if rules cover latency and errors and return exit code 4 when a limit is crossed. Rate and concurrency controls help bound the run; host, request-count and rate guardrails can also limit what a saved test may send. These are documented controls, not measurements of achieved load. Command reference, released CI and guardrail description (checked November 1, 2026).
Read the full pepe analysis and sources
Pin the version and retain JSON. The current command reference advertises self-contained HTML export, while the roadmap calls it a feature for the next release and the released README promises JSON and terminal output. We use that more limited description. The roadmap also says a stable 1.0 flag and report contract is still ahead; do not assume an unchanged interface across updates. Choose k6 for its documented executors, dropped-iteration gate and established local HTML export. Roadmap, released README (checked November 1, 2026).
Used pepe? Share your experience to help other readers choose.
Which product are we missing?
Know a product that belongs in this article? Tell us what it does differently and why it deserves a closer look. Newcomers welcome.
Readers and product teams can suggest candidates. Suggestions inform our research; they do not guarantee inclusion or a recommendation.
Opens an email draft with this article attached as a link. Nothing is sent until you send it.
No email app? Send the product name, website, and this article’s title to admin@practicalpicks.guide.
What people are saying
Public discussion is a check on project claims, not a measured reliability score. In a June 2026 comparison, performance engineer NaveenKumar Namachivayam called k6 his “first recommendation in 2026” for HTTP and gRPC API workloads, citing its single binary in CI and thresholds that fail a pipeline, and flagged its AGPL-3.0 license for commercial use. He described Locust’s built-in reporting as minimal, although its documentation lists HTML, CSV and JSON reports. He reports no measurements of his own, and his article discloses that he builds commercial JMeter plugins. QAInsights comparison
A February 2026 performance-testing discussion includes positive k6 experience and confusion about checks not failing a run, which the threshold documentation explains. An April 2021 comparison discussion shows a preference for Python and uncertainty about free local tools versus hosted services. We read both on September 7, 2026; Reddit blocked our automated recheck on October 19. Neither is representative.
The project issue trackers show recent fixes rather than patterns of failure. Since September 2026, k6 fixed a report of digest authentication losing credentials when a script reused a shared URL object, and Locust fixed HTML report file names that truncated fractional spawn rates and an invalid spawn rate that let a run start with no users. An open October 2026 Artillery report, not yet answered or confirmed by the project, says its AWS Lambda functions crashed with Runtime.ExitError when a back-to-back run reused warm environments; if you plan distributed AWS runs, test consecutive runs before relying on them (Artillery issue 3782). Our October 19 search for vulnerabilities, incidents, lawsuits, shutdowns and ownership changes since September 7 found nothing else that changed our conclusions; a search that finds nothing is not proof that nothing exists. k6 issue 6397, Locust issue 3554, Locust issue 3556
How we researched
This is public-source research. We did not install a runner, create an account, generate local or remote load, measure capacity, test CI exit codes or audit security. No project supplied access, payment or input, and we use no affiliate links. We first compared these three runners on September 7, 2026, and rechecked every material claim on October 19, 2026: releases, runtime requirements, licenses, documented workload and pass/fail behavior, reporting, distributed execution, hosted prices, usage reporting and incident records. We read project documentation, GitHub release and issue records, the PyPI and npm package registries, license files, vendor pricing pages, the OSV vulnerability database and Grafana’s incident report.
We wrote this for a developer or test engineer who wants to load test their own HTTP service from code, locally and in CI. We fixed five criteria and their weights before comparing: control over the planned workload (25%), pass/fail rules for CI (25%), scripting and code reuse (15%), results kept on your own machine (15%) and scaling and adoption risk (20%). k6 scored highest. Giving scripting much more weight for a team whose test logic must be Python favors Locust; reuse of existing Node.js helpers narrows k6’s lead over Artillery but does not reverse it. Browser-level load tests and hosted-only services serve a different workflow.
On November 1, 2026, we screened the newer tools named in our research and their public repositories and release records. Directory pages were not readable, so discovery was limited to those candidates. None established all the requirements for a Rising Star award. We selected no Rising Star; the ordinary comparison includes pepe for its documented local flow and CI features. pepe releases, released README (checked November 1, 2026).
Under the same five criteria, k6 scored 4.175, NBomber 3.95, Goose 3.7 and pepe 3.675. The lead over NBomber is narrow. Locust also scored 3.675; pepe ranks ahead on workload control, the first tie-break. Locust remains the useful alternative for Python test logic, with a custom aggregate CI handler; Artillery scored 3.575 and remains relevant for YAML/Node.js journeys and distributed AWS runs, with its local HTML-report removal and Azure license restriction. Locust CI, Artillery ensure, report removal.
Other scored alternatives were JMeter (3.425), handicap and Vegeta (3.4 each), wrk2 (3.325), autocannon (3.275) and wrk (3.075). JMeter offers CLI runs, local HTML reports and script extensions, with results parsing needed for an aggregate CI gate. Vegeta, wrk and autocannon expose programmable requests and results but need custom gate plumbing. We could not assign complete scores to Gatling, Tsung, Barrage, bombard or LoadDensity: the public material available to us did not establish every criterion. JMeter CLI, handicap, Vegeta, wrk2, autocannon, wrk scripting (checked November 1, 2026).
We could not confirm all prices for the additional alternatives.
| Tool | Public sources | What remains uncertain |
|---|---|---|
| k6 | Workload, threshold, module, TypeScript, dashboard, Operator, usage and versioning documentation; 2.0.0, 2.3.0 and 1.8.2 release notes; repository license; Grafana pricing and incident report. | Achieved arrival rate and capacity on your hardware, how thresholds behave across Operator pods, what the build service receives, the Mandiant audit itself. |
| NBomber | Scenario source, open/closed simulations, thresholds/assertions, local report formats, commercial pricing, license v3.2 and refund policy, checked November 1, 2026. | Achieved load and distributed behavior in your environment; no equivalent dropped-work gate established here. |
| Goose | Source/license, release history, book and metric-threshold documentation, checked November 1, 2026. | Global open arrival scheduling not established; former Gaggle mode removed; achieved load not measured. |
| pepe | Version 0.57.0 README, release records, command reference, roadmap and repository security information, checked November 1, 2026. | HTML export availability conflicts between current reference and roadmap; no stable 1.0 interface promise yet. |
| Locust | Task, headless, distributed, configuration and library-compatibility documentation; release notes to 2.46.7; PyPI record; repository and shutdown-related change. | load-generating process scaling and Python library compatibility in your environment, what locust.cloud now offers. |
| Artillery | Test-script, HTTP, metrics, expect, ensure, run, report, distributed, Lambda, Azure and telemetry documentation; 2.0.34 release notes; npm record; license files; pricing page; issue tracker. | Whether Cloud plan limits apply to AWS runs outside Artillery Cloud, Lambda reliability after the open report, AWS clean-up and total operating cost. |
| Additional alternatives and newcomers | Project documentation, pricing, repositories and release records checked November 1, 2026; NBomber thresholds, simulations, license and refund policy; Goose thresholds; pepe released README and roadmap. | Achieved load on your hardware; exact feature availability where release and current documentation differ; complete criterion evidence for the alternatives without scores. No Rising Star selected. |
None of this establishes load capacity, measurement accuracy or security on your systems. Next substantive review is due January 17, 2027, or sooner after a major release, a license change or a change in hosted terms. Found a changed release or behavior? Contact the editors.
Sources and evidence
Sources were checked October 19, 2026 unless another date is given. Project documentation establishes documented capabilities, not behavior in your environment; license files establish published terms, without a legal opinion; vendor pricing pages are vendor statements that can change. Community posts are attributed opinions, not representative samples.
- k6 (Grafana Labs): open and closed models; dropped iterations; thresholds; modules; TypeScript; HTTP requests; extensions; web dashboard; distributed tests with the Operator; usage collection; versioning and stability guarantees; release notes for 2.0.0, May 11, 2026, 2.3.0, September 21, 2026 and 1.8.2, October 12, 2026; repository and license; Grafana pricing; 2021 acquisition announcement.
- Locust: tasks, wait times and HTTP client; headless and CI; distributed operation; configuration and reports; library compatibility; PyPI record, 2.46.7, October 4, 2026; releases, including 2.46.6, September 17, 2026; repository and license.
- Artillery (Artillery Software Inc): test definitions; HTTP engine; reported metrics; expect; ensure; run; removed report command; distributed testing; AWS Lambda; Azure Container Instances; telemetry; privacy policy; pricing; 2.0.34 release, August 14, 2026; repository and licensing; Azure license; issue 3782, opened October 16, 2026.
- Additional runners (checked November 1, 2026): NBomber source, thresholds, load simulations, pricing, license v3.2, refund policy; Goose book, thresholds, source, releases; pepe released README, reference, roadmap.
- Incident and continuity records: Grafana’s June 23, 2026 post-incident report, a vendor account citing an external audit we did not see; Locust’s December 23, 2025 merged change removing locust.cloud references, a project change record. Vulnerability searches used OSV.
- Outside the projects: QAInsights comparison, June 2026, by an author who discloses commercial JMeter plugin work; Reddit QualityAssurance discussions from February 2026 and April 2021, last read September 7, 2026; issue records k6 6397, Locust 3554 and Locust 3556.
This is a researched comparison, not a product review. Method: evidence standard, editorial picks. Original publication: not published.