The feature flag bill nobody forecasts correctly is LaunchDarkly’s, and it’s not because the seat price is high. Twelve dollars per seat per month is fine. What gets teams is the second meter — client-side MAU at $8.33 per thousand on annual billing — because that one scales with your users, not your engineers. Ten developers shipping to a hundred thousand monthly users pay roughly $120 in seats and $833 in MAU. The seats are a rounding error.
That asymmetry is the whole game. Every vendor in this space meters you differently, and the model you pick determines whether your bill tracks your team size, your traffic, or nothing at all. Pick the one that’s misaligned with your shape and you’ll be renegotiating in eighteen months.
So here’s the actual cost math for the five platforms worth considering in 2026, with worked examples at two team sizes.
The four metering models
Strip away the marketing and there are only four ways these products charge you.
Per-seat is LaunchDarkly’s Pro tier ($12/seat/month) and Unleash Enterprise ($75/seat/month). Predictable, scales with headcount, and completely indifferent to your traffic. Good if you have a lot of users and few engineers.
Per-MAU is LaunchDarkly’s second meter and the one that bites. $8.33 per 1,000 client-side monthly active users annually, $10 per 1,000 if you pay monthly. Experimentation MAU is metered separately at $3 per 1,000. Session replay adds $3.50 per 1,000 sessions.
Per-request is PostHog and Flagsmith. PostHog gives you 1M flag evaluations free every month, then charges $0.0001/request from 1–2M, dropping to $0.000045 (2–10M), $0.000025 (10–50M), and $0.00001 above 50M. Flagsmith’s Start-Up plan is $45/month with 1M requests included and $7 per 100k after that.
Flat or free is the self-hosted open-source path — GrowthBook, Flagsmith OSS, Unleash OSS — where the software costs nothing and you pay in infrastructure and attention.
The trap is that “per-request” sounds cheap until you realize a badly configured client SDK polling every 30 seconds generates 86,400 requests per user per month all by itself. Request-based pricing rewards teams who cache and bootstrap properly. It punishes the ones who don’t, and it punishes them silently.
The price sheet, July 2026
| Platform | Cloud entry | Scale tier | Self-host | Meters on |
|---|---|---|---|---|
| LaunchDarkly | Free (limited) | $12/seat + $8.33/1k MAU | No | Seats + client MAU |
| Flagsmith | Free tier | $45/mo (1M req), $300/mo Scale-Up | Free (OSS) | Seats + requests |
| Unleash | Free OSS | $75/seat/mo Enterprise | Free, 1 project / 2 envs | Seats |
| PostHog | 1M req/mo free | Usage-based, no seat fee | Free (OSS) | Requests |
| GrowthBook | Free | $40/user/mo Pro | Free, unlimited | Seats (cloud only) |
Prices verified July 2026. They move — check the vendor pages before you build a business case on this table.
Two things stand out. PostHog charges nothing per seat, which is unusual and matters a lot if you want PMs and designers touching flags without a procurement conversation. And GrowthBook self-hosted has no usage limits at all — unlimited users, flags, and experiments — which makes it the only option here where growth is genuinely free.
Worked example: 10 devs, 100k MAU
The startup case. You’re shipping to a hundred thousand people and you want flags plus a bit of A/B testing.
LaunchDarkly Pro runs about $953/month — $120 in seats, $833 in client-side MAU. Turn on experimentation and you’re adding another $300 for the experimentation MAU meter. Call it $1,250.
Flagsmith Start-Up is $45/month if you stay inside 1M requests, which at 100k MAU means each user can trigger ten flag evaluations a month. That’s tight. Realistically you’re on Scale-Up at $300/month or eating overages at $7/100k.
PostHog depends entirely on your request volume, not your user count. If your SDK is well-behaved and you’re at 5M evaluations: 1M free, 1M at $0.0001 ($100), 3M at $0.000045 ($135) — about $235/month, with no seat cost. If your SDK polls aggressively, that number can be 10x higher, and that’s on you, not them.
GrowthBook self-hosted is $0 plus a Postgres instance and whatever your time is worth. Unleash OSS is also $0 but you’re capped at one project and two environments, which is a real constraint the moment you want staging separated from production per-team.
Worked example: 50 devs, 5M MAU
This is where the models diverge violently.
LaunchDarkly at list price: 50 seats × $12 = $600, plus 5M MAU × $8.33/1k = $41,650/month. Nobody actually pays that — you’d negotiate an enterprise contract — but it tells you where the negotiation starts. Vendr’s marketplace data puts the median LaunchDarkly annual contract around $72,000, with observed deals ranging from $19,500 to $165,700. Budget six figures if you’re big and you like the product.
Unleash Enterprise at 50 seats is $3,750/month, flat, regardless of whether you have 5M users or 50M. If you’re a consumer product with a small platform team, per-seat pricing is your friend and this looks great next to LaunchDarkly.
GrowthBook Cloud Pro is 50 × $40 = $2,000/month. Self-hosted is still $0.
PostHog at 5M MAU is unanswerable without knowing your request volume, which is the honest answer and also the frustrating one. At 50M evaluations you’re paying roughly $1,000/month; at 500M you’re in volume-discount territory around $5,000.
The pattern: per-seat pricing gets cheaper as you grow (more users amortized over the same platform team), per-MAU pricing gets brutal, and per-request pricing is a function of your engineering discipline rather than your business size.
The lever nobody tells you about
LaunchDarkly bills client-side MAU. Server-side SDK evaluations don’t generate MAU charges.
Read that again, because it’s worth real money. If you move flag evaluation from the browser to your backend — evaluate the flag server-side, render the result, ship HTML — the MAU meter mostly stops spinning. Same product, same flags, dramatically different bill.
This isn’t a billing hack, it’s better architecture anyway. Server-side evaluation means flag names and unshipped feature configs never reach the client, which is a genuine security consideration. Client-side flags leak your roadmap to anyone who opens devtools. I’ve seen unreleased pricing tiers discovered this way.
The catch is that server-side evaluation doesn’t work for flags that need to change UI without a round trip, and it complicates edge/CDN caching. But if your flag bill has a comma in it and 80% of your flags are backend logic, this is the first thing to look at.
Unleash’s OSS Edge problem
If you’re currently self-hosting Unleash and using Edge for flag evaluation at the edge, there’s a date on your calendar you may not know about: Unleash OSS Edge entered long-term support in December 2025 and reaches end-of-life on December 31, 2026.
After that, edge-based evaluation at scale means Enterprise Edge. Which means a paid dependency inside what you deployed specifically because it was fully open source. That’s a bait-and-switch in effect if not in intent, and it changes the calculus on Unleash as a “free self-hosted” option — the free tier’s 1-project/2-environment cap was already restrictive, and now the escape hatch for scale has a price tag.
I’d think hard before starting a new self-hosted Unleash deployment in the back half of 2026. GrowthBook and Flagsmith OSS don’t have this cliff. It’s the kind of thing that shows up in the release notes and nowhere near the pricing page — the same pattern we saw with HCP Terraform’s free tier changes.
What self-hosting actually costs
“Free” self-hosted is free like a puppy is free.
The container is easy — GrowthBook, Flagsmith, and Unleash all ship Docker images that run on a laptop. Production is a different conversation. You need Postgres with a real backup story, Redis for caching in most configurations, at least two application replicas behind a load balancer, and a plan for what happens when the flag service is down.
That last one is the important one. A feature flag service is a hard dependency on every request path that reads a flag. If it goes down and your SDKs aren’t configured with sane defaults and local caching, you don’t degrade — you fail. Every SDK worth using supports bootstrapping from a local file and serving last-known-good values from cache, and if you self-host you must configure both. The SaaS vendors have a global CDN and an SRE team. You have a Tuesday.
Realistic self-hosted cost: $150–400/month in infrastructure for a properly redundant setup, plus a few days of initial engineering and a couple of hours a month forever. Under about 15 developers that math usually favors the free cloud tiers. Over 50, self-hosting is clearly cheaper than LaunchDarkly and probably cheaper than everything else. In between, it depends on whether you have someone who enjoys this.
OpenFeature doesn’t save you
OpenFeature is a CNCF-backed vendor-neutral SDK spec, and the pitch is that you code against a standard interface and swap providers freely.
It’s real and it works, and it removes about 30% of the lock-in. Your application code stops caring who evaluates the flag. What OpenFeature does not portably move is everything that actually matters: your targeting rules, your user segments, your experiment definitions, your audit history, your approval workflows. Those live in the vendor’s data model and there’s no standard for them.
So OpenFeature makes the first migration cheaper — the SDK swap — while leaving you to hand-rebuild several hundred targeting rules in the new tool. Adopt it, because a 30% discount on a painful migration is still worth having. Just don’t let it convince you that vendor choice is reversible on a weekend.
Who has real experimentation
This is the fault line that splits the list.
Flags-only: Flagsmith and Unleash. They’ll flip bits, target segments, and do percentage rollouts. If you want to know whether the feature worked, you’re piping events to something else.
Flags plus real stats: GrowthBook, PostHog, LaunchDarkly, and Statsig. GrowthBook is the interesting one — it was built experimentation-first, does Bayesian and frequentist analysis, CUPED variance reduction, and reads from your existing warehouse rather than requiring you to double-write events. If you already have data in BigQuery or Snowflake, GrowthBook plugs into it instead of duplicating it. That’s a genuinely different architecture from everyone else here, and if you’ve already got a warehouse you’re paying for, it’s close to free experimentation.
PostHog bundles flags with product analytics and session replay, which is convenient if you want one tool and a liability if you wanted flags without shipping another analytics SDK to the browser. We covered that trade-off in more depth in the Sentry vs PostHog comparison.
Picking one
Small team, no budget, wants flags today: PostHog. The free tier is 1M evaluations and there’s no seat charge, so the whole team can use it without a purchase order.
Already have a data warehouse and care about experiment rigor: GrowthBook, self-hosted. Free, unlimited, and the stats are better than what the flag-first vendors bolted on afterward.
Large team, consumer-scale traffic, wants managed: Unleash Enterprise at $75/seat is flat and predictable, which beats watching a MAU meter spin. Just know about the Edge EOL.
Regulated, air-gapped, or data-residency constrained: Flagsmith self-hosted has the most mature on-prem story of the open-source three.
Enterprise with real money and a need for approval workflows, audit trails, and someone to call at 3am: LaunchDarkly, and negotiate hard. The product is genuinely the most polished. You’re paying for it twice.
If you’re currently on LaunchDarkly and the bill is uncomfortable, don’t start with a migration. Start by counting what fraction of your flags are evaluated client-side, and move the backend-only ones to server-side SDKs. That’s a week of work and it might cut the invoice more than switching vendors would.