Koyeb vs Sevalla in Late 2026: Serverless Scale-to-Zero vs $5 Instances vs a Flat Hetzner Box

The Render-alternatives shortlist looks different than it did a year ago. Koyeb, the Paris-founded serverless platform, now sits inside Mistral's compute empire after a February 2026 acquisition — and its famous free tier is closed to new signups. Sevalla, Kinsta's usage-based PaaS, pitches the opposite model: no platform fee, no seat fees, just per-second pods starting at $5 a month. Put the same production app on each, and the cheaper unit price does not always win.

I priced one reference topology — a small production app with a web service, managed Postgres, Redis, and 100 GB of egress — on both platforms plus an owned Hetzner box as the none-of-the-above baseline. Here is the short version, with the full math below.

Two surprises hide in that table. First, at this small tier Sevalla's more expensive compute still loses to Koyeb on units but the gap is far narrower than Koyeb's 3.7x unit-price advantage at larger sizes suggests — because Koyeb's $29 Pro base fee (minus a $10 compute credit) eats the savings on small topologies. Second, the database dominates everything: managed Postgres alone costs roughly twenty times the entire Hetzner box. The rest of this post unpacks where each platform stands in late 2026, shows every line of the math, and maps which workload shape each one wins.

Where each platform stands in late 2026​

Koyeb was founded in Paris in 2020, raised $8.6 million, and built a serverless platform on its own stack and bare metal: per-second billing, scale-to-zero on CPU and GPU alike, serverless Postgres (GA since May 2025), and a GPU range stretching to 8x H200s. In February 2026 Mistral AI acquired it and folded the platform into Mistral Compute, its sovereign-European AI cloud play. The platform keeps operating, but the strategy visibly tilted toward enterprise and AI workloads — sandboxes, GPUs, inference — over the hobbyist funnel.

That tilt changed the entry price. Multiple independent trackers confirm Koyeb closed its free tier to new signups after the acquisition: existing accounts keep the free nano instance, but a new account now enters on Pro at roughly $29 a month (billed upfront) before deploying anything. Any 2026 cost comparison that prices Koyeb from $0 for a new team is stale. The honest floor for a newcomer is the Pro base plus metered compute — and the cheapest metered compute is the eco-nano at $0.0000006 per second, about $1.61 a month always-on.

Sevalla is Kinsta's PaaS product, running application hosting on Google Kubernetes Engine with Cloudflare on the edge — 25 GCP regions, 260-plus points of presence. The pricing philosophy is the anti-Koyeb: usage-based pods billed per second, no platform fee, no seat fees, no feature gates.

Your invoice is three lines: pod usage, egress at $0.10 per GB, and build time at $0.02 per minute. New accounts get a $20 credit against the first month's server resources.

The headline "$5 a month" needs one asterisk, and it matters for this comparison: the $5 Hobby pod cannot take a custom domain. Hobby is fine for previews, staging, and side projects, but a production app with its own domain starts at the Standard tiers — $18 a month for a 1 vCPU / 1 GB CPU-optimized pod, $40 for 1 CPU / 2 GB. I priced the reference topology on production-capable tiers throughout, and treat Hobby where it belongs: the bursty/preview analysis later.

The reference topology (so you can check my math)​

Every number in this post prices the same small production app:

• One web service, ~1 vCPU / 1 GB, always-on, behind a custom domain
• One managed Postgres for primary data
• One Redis for cache and queues
• 100 GB of monthly egress — a realistic small-app figure, not a landing page and not a video host
• Regular CI builds — about 100 build-minutes a month

List prices are current as of September 2026, cross-checked between the vendors' own docs and GetDeploying's provider comparisons. Where a platform has no managed equivalent (Koyeb publishes no managed Redis; Hetzner manages nothing), I priced the closest real option and labeled it.

Head-to-head: compute, Postgres, Redis, egress, builds​

Compute. At the small end, Koyeb's Standard small (1 vCPU, 1 GB RAM, 10 GB SSD) lists at $10.71 a month against Sevalla's CPU-optimized C1 (1 vCPU, 1 GB) at $18 — Koyeb about 41% cheaper per unit. The gap widens with size: 4 vCPU / 8 GB is $42.85 on Koyeb's eco-xlarge versus $160 on Sevalla's Standard S4, and 8 vCPU / 16 GB is $85.71 versus $320 — roughly 3.7x in Koyeb's favor at both steps. But units are not bills. A new Koyeb account pays the $29 Pro base (with a $10 compute credit, so $19 net before usage), while Sevalla charges no base at all. On the reference web service alone, the all-in compute lines read $29.71 for Koyeb ($19 net base + $10.71) versus $18 for Sevalla. Koyeb's unit advantage only overcomes its base fee once the topology grows past roughly $50 of metered compute — a crossover the sensitivity section below makes precise.

Postgres. This is the closest race in the comparison and the largest line in the bill: Koyeb's Large (1 vCPU, 4 GB RAM, 50 GB storage) at $144.04 a month versus Sevalla's Database 4 (2 vCPU, 8 GB) at $145 — within a percent of each other, with Sevalla handing you double the vCPU and RAM at the matched price. Neither platform currently documents a smaller production Postgres tier at these reference points, so small apps overpay for database headroom on both. If your data layer is modest, this single line is the strongest argument in the whole post for the self-hosted baseline.

Redis. Sevalla sells managed Redis directly — Database 2 with 2 GB at $34 a month. Koyeb lists no managed Redis, so the honest equivalent is a second small service running your own Redis image at $10.71 a month: cheaper, but you own persistence, failover, and upgrades. Score it as managed-versus-self-hosted inside the PaaS bill, not as a like-for-like win.

Egress and builds. Sevalla meters both explicitly: 100 GB of egress adds exactly $10, and 100 build-minutes add $2. Koyeb publishes no separate per-minute build meter in any pricing source I checked, so builds add no visible line at this volume — and community pricing trackers report a 100 GB free bandwidth allowance with per-GB overage beyond it, which puts the reference egress at ~$0. Egress is worth watching as you scale — it is the line most likely to flip a comparison past a few hundred gigabytes — but at 100 GB it is $10 on one side and noise on the other.

Add it up: ~$184 a month on Koyeb, ~$209 on Sevalla — and $144–145 of each total is the database. Strip Postgres out and the compute-plus-cache-plus-egress subtotals are roughly $40 on Koyeb against $64 on Sevalla, which is the fair statement of the platforms' differing compute economics once the shared Postgres tax is removed.

Where each one wins: bursty versus steady​

The always-on totals above describe one workload shape. Change the shape and the winner changes — this is the actual decision, and it matters more than the $25 gap between the headline totals.

Bursty and preview workloads favor Koyeb's model, with a Sevalla caveat. Koyeb's scale-to-zero is the real thing: idle services stop billing entirely, per-second granularity, on GPUs as well as CPU. A preview environment that serves traffic two hours a day on a small instance costs under a dollar a month in metered compute, with no hibernate policy to configure and no cold-start babysitting beyond the platform's own light-sleep behavior. Sevalla answers two ways: per-second billing means deleting Friday's preview env on Monday costs three days, not a month — and the $5 Hobby pod, domainless but perfect for previews, undercuts Koyeb's unit price for always-on-tiny. The honest summary: Koyeb wins bursty on automation (zero is automatic), Sevalla Hobby wins always-on-tiny on unit price ($5 flat beats $10.71 flat). Teams running dozens of preview envs should price both: Koyeb's zero-idle against Sevalla's $5-always-on, multiplied by env count, with Koyeb's $29 base amortized across the fleet.

Steady always-on workloads split by size. Below roughly $50 of metered compute, Sevalla's zero base fee beats Koyeb's cheaper units — the reference topology demonstrates it. Past that crossover, Koyeb's 3.7x unit advantage at 4 vCPU and up compounds fast: a 4 vCPU / 8 GB service tier is $42.85 against $160, a $117 monthly gap no base fee survives. The break-even rule of thumb: if your steady-state metered compute fits in one small instance, Sevalla; if you run two or more production services, or anything above 2 vCPU sustained, Koyeb's units take over. Utilization is the second axis: anything averaging under ~20% CPU over the month belongs on scale-to-zero economics (Koyeb, or aggressive hibernation on Sevalla), because you are paying the always-on rate for idle time neither platform forces you to buy.

Neither wins the database. At ~$145 a month for entry production Postgres, both managed offerings cost six to eight times the compute they serve in the reference topology. Teams already self-hosting Postgres — on Hetzner, on RDS, anywhere — should mentally subtract that line from both PaaS totals before comparing. It is the single biggest lever in this entire post.

The none-of-the-above baseline: a flat Hetzner box​

The same topology fits comfortably on one Hetzner Cloud CX23 — 2 vCPU, 4 GB RAM, at $6.92 a month on current list prices — with the web service, Postgres, and Redis sharing the box behind Caddy or Traefik. One step up, the CX33 (4 vCPU, 8 GB) at $10.38 buys headroom most small apps never fill. Egress at 100 GB is noise against a multi-terabyte included allowance, and block storage only enters the picture past local disk at about $5 per 100 GB. All-in cash cost: under $10 a month, or roughly one-twentieth of either PaaS total.

Three honesty notes keep this baseline representative rather than flattering. First, Hetzner raised prices twice in 2026 — an April tier restructuring and June hikes that nearly tripled some dedicated and cloud list prices for new orders, citing hardware supply-chain pressure — so the $6.92 figure is the post-hike price, and dedicated-server pricing is currently volatile enough (one popular Ryzen box lists anywhere from €40 to €139 depending on facility and stock) that cloud VMs are the only stable baseline. Second, the $7 box manages nothing: Postgres backups, Redis persistence, OS patching, and 3 a.m. pages are yours, and pricing your own ops time at anything above zero narrows the gap fast. Third, the comparison assumes you can operate what you deploy — the moment the answer is "not yet," the PaaS premium is tuition, not waste.

Even with those caveats, the baseline reframes the PaaS decision correctly: you are not choosing between $184 and $209. You are choosing whether managed deploys, managed Postgres, and never touching a node are worth ~$175–200 a month over raw infrastructure — and, if yes, which billing shape matches your workload.

Which should you pick?​

For a side project or preview fleet where every env idles most of the day: Koyeb's scale-to-zero, with the $29 Pro base amortized across all your services — or Sevalla Hobby at $5 per always-on-tiny env if you prefer flat and predictable. For one small steady production app on a custom domain: Sevalla, where the absence of a platform fee beats cheaper units. For two-plus production services or anything GPU-adjacent: Koyeb, where per-second billing and a 3.7x unit advantage at size compound, and the Mistral-era GPU and sandbox roadmap is the actual product direction. For the team that already operates infrastructure: the Hetzner box, with eyes open about the ops time.

The deeper pattern is that the Render-alternatives market has split by billing philosophy rather than features. Both platforms git-push, autoscale, and run workers; what differs is who pays for idle — Koyeb's answer is "nobody, automatically," Sevalla's is "you, but cheaply and by the second," and Hetzner's is "you, flat, forever." Match the answer to your utilization curve and the invoice takes care of itself.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own. Star the repo on GitHub or deploy your first app today.