# Koyeb Sandboxes

Official pricing: https://www.koyeb.com/pricing  
Category: agent-sandbox · Isolation: vm

## Pricing regimes (raw)

- **Eco instances (2 GB/vCPU, older Milan CPUs, was/fra/sin)** (sizes): eco-nano 0.1 vCPU/0.25 GiB $0.00216/h; eco-micro 0.25 vCPU/0.5 GiB $0.0036/h; eco-small 0.5 vCPU/1 GiB $0.0072/h; eco-medium 1 vCPU/2 GiB $0.0144/h; eco-large 2 vCPU/4 GiB $0.0288/h; eco-xlarge 4 vCPU/8 GiB $0.0576/h [burstable]
- **Standard instances (1 GB/vCPU up to 2xlarge, Genoa, all regions)** (sizes): nano 0.25 vCPU/0.25 GiB $0.0036/h; micro 0.5 vCPU/0.5 GiB $0.0072/h; small 1 vCPU/1 GiB $0.0144/h; medium 2 vCPU/2 GiB $0.0288/h; large 4 vCPU/4 GiB $0.0576/h; xlarge 8 vCPU/8 GiB $0.1152/h [burstable]
- **GPU instances (bundled vCPU/RAM/disk per GPU)** (resource): $0/vCPU-h, $0/GiB-h
- **Savings plan / long-term reservation (up to 50% off, best case)** (sizes): eco-xlarge 4 vCPU/8 GiB $0.0576/h; eco-2xlarge 8 vCPU/16 GiB $0.1152/h; xlarge 8 vCPU/8 GiB $0.1152/h; 2xlarge 16 vCPU/16 GiB $0.2304/h; 3xlarge 24 vCPU/32 GiB $0.4608/h; 4xlarge 32 vCPU/64 GiB $0.9216/h [sales]

## Features

Yes: snapshots, memory snapshots, fork/clone, pause/resume, persistent disk, volumes, idle auto-stop, ≥24 h sessions, custom image (Docker or snapshot), start from your own snapshot, your own Docker/OCI image, full VM (own kernel), Docker inside, preinstalled agents, HTTPS preview URLs, custom domains, raw TCP inbound, egress allowlist, open internet, private networking, self-hosting, BYOC, SOC 2, SSO, GPU, Python, Node.js, Go, Rust, Java, wake on request, memory fork, audit logs, EU data residency, MCP server, automatic snapshots, snapshots on demand, gVisor or VM (no shared kernel), inbound access rules, firewall inside (nftables), extra volumes

No: nested virtualization, browser, desktop GUI, computer-use API, code interpreter, browser + desktop control, anti-bot stealth, CAPTCHA solving, residential IPs, SSH, public IPv4, static egress IP, open source, HIPAA, arm64, Windows, macOS, GPU desktop, HTTP method/path egress rules, live resize, webhooks, spend limits, agent harness API, hosted agent API (their own agent), secret proxy, secret proxy for any API, shared volumes

Unknown: everything else. Evidence (source + quote) per feature: https://battleships.dev/data/providers/koyeb.json → feature_evidence

## Caveats

- Plan fees are only partly usage credit: Pro $29 includes $10, Scale $299 includes $100, Enterprise from $1,000 includes $500. Starter (no fee) is closed to new signups since about 2026-02-26 (flag legacy).
- Eco eligibility for sandboxes is undocumented. Eco-xlarge fits 4/8 exactly at $0.0576/h; standard needs xlarge 8/8 at $0.1152/h. Eco = older Milan CPUs, 20 GB disk max, no volumes, was/fra/sin only.
- Light Sleep (~200 ms wake) is free only during public preview; a GA price is announced but unpublished. Deep sleep and paused services are $0. The idle period before sleeping (min 5 min) is billed at the full rate.
- 'CPU concurrency' 100 (Pro) / 1,000 (Scale) may count vCPUs rather than instances; if vCPUs, 50 x 4-vCPU sandboxes require Scale.
- Sandbox snapshots (FILESYSTEM, or FULL with running processes) and spawn-from-snapshot exist only in the official SDK (koyeb-sdk >= 1.5.2, 2026-08-27, labelled experimental); not documented on koyeb.com and no snapshot price published (snapshot_gib_month null). Volumes (preview, 1-10 GB, was/fra, standard/GPU only, not eco) and volume snapshots were 'free during the public preview' (blog); an unverified search snippet cites 10 GB free then $0.08/GB-month.
- Bandwidth: pricing page = 1 TB/mo included on Pro/Scale then $0.02/GB US/EU, $0.04/GB Asia; the older FAQ says 100 GB free then $0.04/GB. Engine uses the pricing page.
- features.max_session_h set null (the features file's 24 h is the optional auto-delete maximum, not a forced session limit). delete_after_inactivity_delay max 12 h after scale-to-zero.
- Savings-plan mode uses the published best case (50% off); actual discounts are negotiated.
- CPU guarantee is 'at least 0.25 dedicated vCPU per GB memory, burstable to all cores', so a standard 1 GB/vCPU instance may get less than its nominal vCPUs under contention.
- Startup program up to $30k credits (not in free).
- No benchmark data for Koyeb in benchmarks.json.
- Verifier 2026-09-28 (pricing page embedded instance catalog): GPU regions corrected from us/eu/asia to us (na) + eu (RTX-4000-SFF-ADA only); L4 and B200 removed from the priced GPU map because their catalog status is RESTRICTED (no regions); eco instances have light_sleep_enabled=false (deep sleep only). Standard, eco and GPU hourly prices, plan fees/included usage, concurrency and bandwidth re-confirmed.
- Data residency (DPA): Customer Data stays in the selected region but may move between datacenters within it; global edge routing is not extra compute regions. Core regions FRA, PAR, SIN, TYO, WAS, aws-us-east-1; SFO preview on paid plans. Eco only FRA/WAS/SIN. Egress: $0.02/GB US/EU, $0.04/GB Asia after 1 TB/month (card egress_gib is the US/EU rate). eco-nano hour = per-second $0.0000006 x 3600 = $0.00216 (display rounds to $0.0022). (patches v10-regional, v5-calculators, 2026-09-28)
- Official pages conflict: the live pricing page omits Starter and says Enterprise starts at $1,000/month, while the organization reference (updated May 27, 2026) says Starter is available to all users with a valid payment method and Enterprise has a $700/month committed cost. Keep pricing-page values for ranking, but do not present Starter closure or $1,000 as undisputed.
- Aggregate cap: the General FAQ says 'With the Scale plan, you can use up to 1TB of memory and up to 1000 services.' So Scale is also capped at ~1 TB of concurrently running RAM (e.g. 125 x 8 GiB sandboxes), and 'CPU concurrency' 100/1,000 most likely counts vCPUs, not instances (unconfirmed). Large fleets need Enterprise.
- Sandbox SDK defaults: instance_type 'micro' (0.5 vCPU/512 MB, $0.0072/h), region 'na', idle_timeout 300 s -> deep sleep after 5 min (min scale 0); Light Sleep for sandboxes is only an '_experimental_enable_light_sleep' flag. idle_timeout=0 means always-on (min scale 1).
- Billing clock: 'Paid usage begins when an Instance has been created and the Deployment has a Starting status', so image pull/boot time is billed (boot_overhead_s unknown). Usage 'accounted per second and billing is rounded up to the nearest unit'; no minimum billed duration is documented.
- Terms: Fees depend on Authorized Users and Services; exceeding the plan's included users/services is 'charged based on usage' but no per-user/per-service price is published. Starter orgs are limited to 1 member (org doc).

## How this provider charges


Koyeb Sandboxes (public preview since 2025-11-19) are Koyeb services running on Koyeb Instances: Cloud Hypervisor microVMs on bare metal. They are billed **per second on a fixed instance type**, plus a monthly plan subscription whose fee only partly returns as included usage. Koyeb was acquired by Mistral AI (announced 2026-02-17). Since about 2026-02-26 new users must subscribe to Pro/Scale/Enterprise, because the free Starter plan was removed.

## Regime table

| # | Regime | When it applies | How billed | Numbers | Source |
|---|---|---|---|---|---|
| 1 | Standard instances | All regions (fra, was, sin, tyo, par, sfo preview, aws-us-east-1 on Scale). AMD EPYC Genoa, NVMe | Per second on the instance while running. RAM is fixed at **1 GB/vCPU** (nano to 2xlarge) | nano 0.25/0.25 $0.0036/h · micro $0.0072 · small 1/1 $0.0144 · medium 2/2 $0.0288 · large 4/4 $0.0576 · **xlarge 8/8 $0.1152** · 2xlarge 16/16 $0.2304 · 3xlarge 24/32 $0.4608 · 4xlarge 32/64 $0.9216 · 5xlarge 40/128 $1.8432 | https://www.koyeb.com/pricing (Standard tab), https://www.koyeb.com/docs/reference/instances |
| 2 | **Eco instances** | Only was / fra / sin. Older AMD EPYC Milan, "half the cores" of the same-priced standard type, small disk (max 20 GB). **No volumes** | Per second. Same $/h ladder as standard, but **2 GB/vCPU** | eNano 0.1/0.25 $0.0022 · eMicro 0.25/0.5 $0.0036 · eSmall 0.5/1 $0.0072 · eMedium 1/2 $0.0144 · eLarge 2/4 $0.0288 · **eXLarge 4/8 $0.0576** · e2XLarge 8/16 $0.1152 | https://www.koyeb.com/pricing (Eco tab), https://www.koyeb.com/blog/new-eco-instances-the-most-affordable-way-to-deploy-apps-globally |
| 3 | GPU instances | GPU sandboxes supported ("CPU and GPU instances"). Plan GPU concurrency: Pro 20, Scale 40 | Per second, **bundled** vCPU/RAM/disk per GPU | RTX-4000-SFF-ADA 0.50 (6 vCPU/44 GB) · L4 0.70 (6/32) · RTX-A6000 0.75 · L40S 1.20 (15/64) · A100 1.60 (15/180) · A100 SXM 2.15 · RTX PRO 6000 2.20 · H100 2.50 · H200 3.00 · B200 5.50. Multi-GPU linear (8x H100 $20, 8x H200 $24, 8x A100 SXM $17.20). Tenstorrent N300s on request | https://www.koyeb.com/pricing (GPU tab) |
| 4 | Pro plan | Entry paid plan (required for new signups) | $29/month fee with **$10 included compute** (not a full credit) | 10 users, 100 services, CPU concurrency 100, GPU 20, 1 TB bandwidth, 5 builds. Light sleep window 5 min-3 h, deep sleep 5 min-6 h | https://www.koyeb.com/pricing, /docs/faqs/pricing, /docs/run-and-scale/scale-to-zero |
| 5 | Scale plan | Needs more than 100 services/instances, AWS regions, SLA | $299/month with **$100 included** | 50 users, 1,000 services, CPU concurrency 1,000, GPU 40, 99.9% SLA. Sleep windows up to 6 h light / 12 h deep | same |
| 6 | Enterprise | Sales | "Starting at $1,000/mo", $500 included usage | Custom CPU/RAM/GPU, 50 regions, BYOC, dedicated racks, SSO, ISO27001/SOC2, 99.99% SLA | https://www.koyeb.com/pricing |
| 7 | Starter (legacy) | Accounts created before about 2026-02-26 only | Pay-as-you-go, no fee, no included usage | Light sleep fixed 5 min, deep sleep ≤ 1 h 05 | https://www.koyeb.com/blog/koyeb-is-joining-mistral-ai-to-build-the-future-of-ai-infrastructure, https://github.com/robhunter/agentdeals/issues/2180 |
| 8 | Savings plans / long-term reservations | Contact sales | Committed capacity | "**up to 50% off** the on-demand price" | https://www.koyeb.com/pricing |
| 9 | Running, idle before sleep | Between the last request and the idle period expiring | Full instance rate | Minimum idle period 5 min (all plans) | /docs/run-and-scale/scale-to-zero |
| 10 | **Light Sleep** (snapshot-based, ~200 ms wake) | Scale-to-zero, first stage | **Free during public preview.** "Once it is Generally Available, there will be a cost associated" (amount unpublished) | $0 now, null later | /docs/run-and-scale/scale-to-zero |
| 11 | Deep Sleep (VM discarded, 1-5 s cold start) | After the deep-sleep window | No compute billed ("No charges for paused Services") | $0 | same; /docs/faqs/pricing |
| 12 | Auto-delete | `delete_after_delay` (60 s-24 h after create), `delete_after_inactivity_delay` (≤ 12 h after scale-to-zero) | Stops all billing | n/a | /docs/sandboxes/sandbox-lifecycle |
| 13 | Instance disk | Local NVMe SSD included in the instance price, ephemeral | Included | 20 GB (eXLarge) / 80 GB (xlarge) | pricing tabs |
| 14 | Volumes (preview) | was/fra only, 1-10 GB, single-instance services, **not on eco** | Price not published (the pricing page lists "NVMe Volumes and Snapshots" as a Pro feature) | null | /docs/reference/volumes |
| 15 | Volume snapshots (preview) | Volumes only. There is **no sandbox snapshot API** | "Free during the public preview". A search-engine snippet claims 10 GB free then $0.08/GB-month, but no official page confirms it | $0 now; $0.08 unverified | https://www.koyeb.com/blog/snapshots-create-a-point-in-time-copy-of-your-high-performance-volumes |
| 16 | Egress | Outbound to the internet | Pricing page: **1 TB/month included** (Pro and Scale), then $0.02/GB US/EU, $0.04/GB Asia. The older FAQ says 100 GB free then $0.04 | $0.02 / $0.04 | https://www.koyeb.com/pricing, /docs/faqs/pricing |
| 17 | Custom domains | Beyond 100 (Pro) / 500 (Scale) | "not charged yet, but will be" | $0.20/month each | /docs/faqs/pricing |
| 18 | Free instance | One per org: 0.1 vCPU/512 MB, fra/was | Free, sleeps after 1 h. Not a sandbox size. It is unclear whether it survives on paid plans only | $0 | /docs/reference/instances |
| 19 | Startup program | Application | Credits | up to $30k | https://www.koyeb.com/pricing |

## Gotchas

1. **The plan fee is mostly not credit.** Pro's $29 returns only $10 of usage, and Scale's $299 only $100, so there is a $19 and a $199 dead-weight floor. Since Feb 2026 new users can't use Starter, so the floor is unavoidable.
2. **Eco vs standard at the same price.** eXLarge (4 vCPU/8 GB, $0.0576/h) matches 4/8 exactly. On standard, 8 GB of RAM forces xlarge (8 vCPU/8 GB) at **twice** the price.
   - Eco uses older Milan CPUs and has a 20 GB disk max.
   - Eco exists only in was/fra/sin, and volumes can't attach to it.
   - Whether sandboxes may use eco types is **not documented**.
3. **Light Sleep is free only "during public preview"** and will be charged at GA (rate unknown). Keeping sandboxes warm will get more expensive.
4. **"CPU concurrency" of 100 (Pro) is ambiguous.** It could count instances or vCPUs. If it means vCPUs, 50 × 4-vCPU sandboxes need Scale. Every sandbox is also a *service*, capped at 100 (Pro) / 1,000 (Scale).
5. **No sandbox snapshots or forks.** State survives only in (preview, 10 GB max, non-eco) volumes. The 50 GiB snapshot requirement can't really be met.
6. **CPU is not fully dedicated.** Per the docs, "at least 0.25 dedicated vCPU per GB memory, burstable to all cores". A standard 4 vCPU/4 GB large is guaranteed only 1 vCPU.
7. **Bandwidth rules conflict.** The pricing page says 1 TB included then $0.02/GB. The FAQ says 100 GB then $0.04.
8. "Billing is rounded up to the nearest unit" (FAQ): the unit is not defined.
9. Mistral acquisition: the roadmap is tilting to inference/enterprise. Sandboxes are still "public preview".

## Worked example

Workload: 4 vCPU / 8 GiB, 50 concurrent × 8 h/day × 22 days = **8,800 sandbox-hours**, 30% CPU util, 50 GiB snapshots retained, 100 GiB egress.

Assumptions:
- Sandboxes scale to zero outside hours (deep sleep = $0).
- 100 GiB egress is inside the 1 TB allowance.
- Snapshots: none available for sandboxes, so $0 (the unverified $0.08 rate would add (50−10) × 0.08 = $3.20).

| Regime | Compute | Idle tail (5 min/session) | Plan fee − included | Egress | **Monthly total** |
|---|---|---|---|---|---|
| Eco eXLarge (4/8), Pro | 8,800 × 0.0576 = $506.88 | 91.7 h × 0.0576 = $5.28 | 29 − 10 = $19 | $0 | **$531.16** |
| Eco eXLarge, Scale (if "CPU concurrency" counts vCPU) | $506.88 | $5.28 | 299 − 100 = $199 | $0 | **$711.16** |
| Standard xlarge (8/8, nearest fit), Pro | 8,800 × 0.1152 = $1,013.76 | $10.56 | $19 | $0 | **$1,043.32** |
| Standard large (4/4): does **not** meet 8 GiB | $506.88 | $5.28 | $19 | $0 | $531.16 (under-sized) |
| Starter legacy (old accounts), eco | $506.88 | $5.28 | $0 | $0 | **$512.16** |
| Never sleeping (730 h), eco, Pro | 50 × 730 × 0.0576 = $2,102.40 | n/a | $19 | $0 | **$2,121.40** |
| Savings plan, best case 50% off, eco, Pro | $253.44 | $2.64 | $19 | $0 | **≈ $275.08** (sales; "up to") |
| Enterprise | from $1,000/mo with $500 included | | | | **≥ $1,012.16** ($1,000 + usage above $500; eco rates, before negotiation) |

- The 30% utilisation changes nothing, because billing is on the instance.
- Light Sleep is free today. If it gets a GA price, idle-warm time will cost extra.

Sources: https://www.koyeb.com/pricing (Standard/Eco/GPU tabs rendered on the boat VM 2026-09-28) · https://www.koyeb.com/docs/reference/instances · https://www.koyeb.com/docs/faqs/pricing · https://www.koyeb.com/docs/run-and-scale/scale-to-zero · https://www.koyeb.com/docs/sandboxes · https://www.koyeb.com/docs/sandboxes/sandbox-lifecycle · https://www.koyeb.com/docs/reference/volumes · https://www.koyeb.com/docs/reference/snapshots · https://www.koyeb.com/blog/koyeb-sandboxes-fast-scalable-fully-isolated-environments-for-ai-agents · https://www.koyeb.com/blog/new-eco-instances-the-most-affordable-way-to-deploy-apps-globally · https://www.koyeb.com/blog/snapshots-create-a-point-in-time-copy-of-your-high-performance-volumes · https://www.koyeb.com/blog/koyeb-is-joining-mistral-ai-to-build-the-future-of-ai-infrastructure · https://github.com/robhunter/agentdeals/issues/2180