5 Best Open-Source A/B Testing Tools in 2026 (Ranked by How Open They Actually Are)

I’ve watched a lot of teams search “open-source A/B testing tools” and end up disappointed. The reason is simple. Most tools in this space wear the open-source label, but very few are actually open the way you think they are. The SDKs are open. The platform isn’t. Or the code is MIT, but the vendor tells you not to run it yourself.

So this ranking judges these three tools on the thing that actually matters when you say “open source”: can you read the code, can you run it yourself, and is that a supported path or a dare. Then I layer in the stuff you came for anyway - how deep the statistics really go, and what it costs.

Every price and license here was read from each vendor’s own tool page. Where a fact is third-party (funding, valuations), I say so.

How I ranked these

Three axes, in order of weight for this list:

• Openness that’s real. License is one thing. A supported self-host path is another. I weight both, and I dock points hard when “open source” is technically true but practically a trap.
• Stats-engine depth. A percentage rollout with a chart is not experimentation. I care about CUPED, sequential testing, Bayesian analysis, bandits, and sample-ratio-mismatch checks.
• What it costs to actually run. Free-tier generosity, and the hidden cost - DevOps time, a data warehouse, or a learning curve.

The short version

Tool	Really open source?	Stats engine	Self-host reality	Free path
GrowthBook	Yes - MIT core	Deepest here (CUPED, Bayesian, bandits, SRM)	Real, supported, unlimited users	$0 self-hosted
PostHog	Yes - MIT, but	Strong, analytics-integrated	Open but officially unsupported	1M requests/mo free
Statsig	No - SDKs only	Strongest engine overall	No real self-host	2M events/mo free

1. GrowthBook - the only one that’s open source the way you mean it

If “open source” is a hard requirement, this is the top of the list and it isn’t close. GrowthBook’s core is MIT-licensed, and the self-hosted edition runs unlimited users for $0. That’s a real, supported self-host path, not a technicality. You bring your own data warehouse, get the full stats engine, and rely on community support.

And the stats engine is the reason to pick it even if you didn’t care about openness. It’s the deepest of the open-source flag tools: CUPED for variance reduction, sequential testing, Bayesian analysis, multi-arm bandits, and SRM checks. It connects to BigQuery, Snowflake, Databricks and other sources - 11 in total. It’s warehouse-native, which means it queries the warehouse you already run instead of ingesting a copy of your events. Your analytics data never leaves your infrastructure.

GrowthBook was founded in 2020, went through Y Combinator, and has a reported Series A of $22.6M backed by Khosla and Nexus.

The gotcha is dependency and difficulty. Warehouse-native is a superpower only if you already have an instrumented warehouse - no warehouse, far less value. And the statistics have a genuine learning curve. Reviews consistently cite an engineer-first UI that PMs need a guided tour for, and docs that are thorough but hard to navigate. GrowthBook rewards data maturity and punishes teams that don’t have it. If you have both a warehouse and a data-literate person, GrowthBook is the best-value serious experimentation platform here, open source or not.

2. PostHog - MIT, but don’t buy it for the self-host

PostHog is genuinely open source under MIT, and I rank it second, but with a caution that matters more than most people realize. You can self-host it, but PostHog officially does not recommend it. Their own docs are blunt: self-hosted deployments are unsupported, no guarantees, no tagged releases, they don’t publish CVEs for self-host, and you “assume all responsibility and risk.” They steer almost everyone to Cloud.

So if your reason for searching “open-source A/B testing” is data residency or a supported on-prem deploy, PostHog is not that tool the way GrowthBook is. Read the license as transparency, not as a deployment target.

What PostHog does brilliantly is integration. It’s a product-analytics suite - events, funnels, session replay, surveys - with flags and experiments built in. Its experiments read directly from your product analytics, so the metrics you test against are the same events you already track. No piping data between two tools and hoping the definitions match. On raw experimentation depth it lands second only to Statsig in this set, and that analytics coupling is arguably a bigger practical win.

Flags and experiments share a free tier of 1M requests a month, then fractions of a cent per request. It was founded in 2020 (YC S20) and raised a Series E at a reported $1.4B valuation in October 2025.

The other catch is multi-product billing. Each product - flags, analytics, replay, surveys - meters and bills separately. Your flags line stays cheap while the total climbs across the suite. If you want analytics and experiments together and you’re happy on Cloud, PostHog is a strong pick. If you specifically need a supported self-host, it isn’t.

3. Statsig - the best engine, but not actually open source

I’m putting Statsig third on this specific list, and the reason is right there in the ranking criteria: Statsig is not open source. The SDKs are on GitHub, but the platform is proprietary SaaS. There’s no real self-host. Enterprise offers a warehouse-native deployment that runs compute on your Snowflake or BigQuery, which is great for data residency, but it is not running the application yourself.

I’m still including it, because if you loosen “open source” to “free and warehouse-flexible,” Statsig has the strongest experimentation engine of the three. It was built by ex-Facebook engineers to mirror Facebook’s internal stack, with sequential testing and CUPED. And its billing is uniquely friendly for testing: flag and config checks are unlimited and free on every tier - you pay on analytics events instead. The free Developer tier gives 2M events and 50k session replays with no credit card. Pro is a flat $150/mo.

Two honest cautions. The meter moves on events, so heavy analytics use is where the bill lives even though flags are free. And the bigger one: OpenAI acquired Statsig on 2 September 2025 for a reported ~$1.1 billion, and founder Vijaye Raji became OpenAI’s CTO of Applications. The product still ships independently, but adopting it is a multi-year bet on a roadmap that now sits inside OpenAI. Full detail on the Statsig page. Excellent engine - just don’t file it under “open source.”

So which one?

• You want genuinely open source and can self-host - GrowthBook. MIT core, unlimited users at $0, and the deepest stats engine here. You need a warehouse and someone who gets the statistics.
• You want analytics and experiments in one place, on Cloud - PostHog. The tightest flags-experiments-analytics integration, cheap flags. Don’t buy it for the self-host promise.
• You want the strongest experiment engine and don’t mind proprietary SaaS - Statsig, with the OpenAI-ownership asterisk priced in. Flag checks are free forever.

The honest bottom line: only one of these three is open source in the way most people mean it. If that word is load-bearing for you, start with GrowthBook. If it’s actually a proxy for “free and self-run,” all three have a path - just know which promise each one is really making.

---

Pricing and licensing read from each vendor’s tool page, verified 26 July 2026. Funding and acquisition figures (GrowthBook’s $22.6M Series A, PostHog’s $1.4B valuation, Statsig’s ~$1.1B OpenAI acquisition) are third-party-sourced and attributed as such.