# Blaxel

Official pricing: https://blaxel.ai/pricing  
Category: agent-sandbox · Isolation: firecracker

## Pricing regimes (raw)

- **Sandbox (active time, RAM-only pricing, CPU bundled)** (sizes): 2 GB 1 vCPU/2 GiB $0.0828/h; 4 GB 2 vCPU/4 GiB $0.1656/h; 8 GB 4 vCPU/8 GiB $0.3312/h; 16 GB 6 vCPU/16 GiB $0.6624/h
- **Batch Jobs (non-interactive tasks, max 24 h)** (resource): $0/vCPU-h, $0.0216/GiB-h [alt]
- **Agent Runtime (announced; unpriced)** (resource) [beta]

## 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, root, browser, desktop GUI, computer-use API, browser automation API, desktop control API, browser + desktop control, preinstalled agents, HTTPS preview URLs, custom domains, egress allowlist, open internet, static egress IP, private networking, BYOC, SOC 2, HIPAA, SSO, Python, Node.js, credential injection, HTTP method/path egress rules, wake on request, live fork (no pause), memory fork, scheduled runs (cron), EU data residency, zero data retention, scoped API keys, MCP server, automatic snapshots, snapshots on demand, fast boot (benchmarked), gVisor or VM (no shared kernel), inbound access rules, firewall inside (nftables), secret proxy, secret proxy for any API, extra volumes, shared volumes

No: nested virtualization, systemd, code interpreter, anti-bot stealth, CAPTCHA solving, residential IPs, SSH, public IPv4, raw TCP inbound, self-hosting, open source, GPU, arm64, Windows, macOS, GPU desktop, live resize, spend limits, agent harness API, hosted agent API (their own agent)

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

## Caveats

- Billing is on allocated RAM x ACTIVE time, where active = has an open connection (not CPU utilisation). Engine running hours = wall-clock session hours; real bills can be much lower if clients disconnect between steps (auto-standby ~15 s after last connection) or higher if idle WS/TCP connections (up to 15 min) or keepAlive processes hold the sandbox active.
- Standby snapshot storage ($0.20/GB-month) appears to be billed on ~allocated memory size (vendor example: 2 GB sandbox in standby 1 month = $0.40); persistent fleets pay this for every standby hour.
- Archive (private preview) stops snapshot billing; archive storage price not published.
- Tier 'fee' is a rolling 30-day top-up volume deposited as wallet credit (not a fee); tier downgrades if top-ups lapse, even with balance. Promo credits do not count. Whether wallet credit expires is not documented.
- Sandbox-count quota includes standby sandboxes; an unpublished per-tier account memory quota also exists. Tiers 3-8 not public.
- Tier 0/1 force max TTL 7/30 days; unlimited persistence needs Tier 2+.
- Support add-ons carry a % of usage surcharge (3% / 10%) the card cannot model.
- Minimum billed seconds per resume not documented.
- Writable FS is tmpfs ~50% of RAM; extra disk = Volume at $0.12/GB-month on PROVISIONED size (disk_gib_month here = volume price). Images $0.045/GB-month not modelled.
- Sizes above 16 GB: core count undocumented (third-party cites XL 32 GB at $1.3248/h); up to 256 GB on Custom only.
- Batch Jobs mode is only valid for non-interactive tasks <= 24 h; MCP hosting ($0.0252/GB-h) omitted as a mode.
- No public inbound IPv4; dedicated egress IPs/gateway/proxy free during beta. Wildcard custom domains $20/domain-month.
- 'Up to $200' free credits: exact amount not specified.
- Blaxel acquired by Baseten (2026-09-10); product/pricing stated unchanged.
- unverified: batch jobs may bill active (used) RAM, per "billed by the memory it uses" (https://blaxel.ai/pricing); card keeps alloc as the upper bound.
- Features check 2026-09-28: nested_virt null -> false. Blaxel: "Each sandbox runs in a dedicated Firecracker microVM with its own kernel" (blaxel.ai/blog/soc-2-compliant-sandbox-providers); Firecracker test_nv.py asserts guests have no nested virtualization (/dev/kvm absent). Desktop = official blaxel/xfce-vnc hub image; Docker = official blaxel/docker-in-sandbox image.
- Standby delay conflict: pricing FAQ says sandboxes scale to zero after ~5 s of inactivity, but the lifecycle docs and blog still say ~15 s after the last connection. Card keeps 15 s (conservative); if 5 s is right, disconnected sessions bill ~10 s less each.

## How this provider charges


Blaxel (acquired by Baseten, announced 2026-09-10; "Blaxel will continue to operate and its product does not change") sells one
Pay-as-you-go plan plus a Custom/enterprise plan. **There is no CPU price.** Everything is metered in GB of *allocated* RAM × seconds of
*active* time; CPU cores are derived from memory (2 GB→1, 4 GB→2, 8 GB→4, 16 GB→6 cores). "Active" means "has an active connection",
**not** CPU utilisation. Idle sandboxes drop to standby (~15 s per docs, "~5 s" per marketing) and then pay only snapshot storage.

## Regime table

| Regime | When it applies | How billed | Numbers | Source |
|---|---|---|---|---|
| Sandbox — active | Sandbox has an open connection / keepAlive process / WS/TCP not idle >15 min | Allocated RAM GB × active seconds; CPU bundled (not charged) | $0.0000115/GB-s = **$0.0414/GB-h**; 2 GB (1 core) $0.0828/h, 4 GB (2) $0.1656/h, 8 GB (4) $0.3312/h, 16 GB (6) $0.6624/h | https://blaxel.ai/pricing ; https://docs.blaxel.ai/Sandboxes/Overview |
| Sandbox — standby (auto scale-to-zero) | No active connection for ~15 s; full memory+FS+process snapshot, ~25 ms resume | $0 compute; snapshot storage only | $0.20/GB-month (~$0.000274/GB-h). Vendor example: "a 2 GB sandbox in standby for one entire month would cost $0.40" ⇒ snapshot billed ≈ allocated memory size | https://blaxel.ai/pricing ; https://blaxel.ai/blog/persistent-ai-sandboxes-blaxel |
| Sandbox — archived (private preview, feature flag) | Explicit `archive()`; memory dropped, only FS diff over image kept, processes restarted on unarchive | Snapshot-storage billing stops; archive storage price **not published** | null | https://docs.blaxel.ai/Sandboxes/Archive ; Sandboxes/Overview "Archived mode" |
| Sandbox — deleted / TTL-expired | Expiration policy (ttl, ttl-idle, ttl-max-age, date) or forced tier TTL | Nothing (manual snapshots survive deletion and keep billing) | Tier 0 max TTL 7 d, Tier 1 30 d, Tier 2+ unlimited | https://docs.blaxel.ai/Sandboxes/Expiration |
| Manual snapshots / fork (Carbon = Mark 3.1, private preview) | `snapshot()` / fork | Snapshot storage; never expire; count against snapshot quota | $0.20/GB-month (assumed same "Snapshots storage" line) | https://docs.blaxel.ai/Sandboxes/Fork ; /Infrastructure/Gens |
| Batch Jobs | Workload run as a Job (max 24 h/task, ~30 s boot, no standby/state) | "Active compute across every task in a job, billed by the memory it uses" | $0.000006/GB-s = **$0.0216/GB-h** (−47.8 % vs sandbox) | https://blaxel.ai/pricing |
| MCP hosting | Workload run as an MCP server (max 10 min per request) | Memory-based active compute | $0.000007/GB-s = $0.0252/GB-h | https://blaxel.ai/pricing |
| Agent Runtime | Long-running agent sessions | "Coming soon" | null | https://blaxel.ai/pricing |
| Images | Every stored template/deployment image (billed to *source* workspace even if shared) | GB-month | $0.045/GB-month (charged since 2025-12-01) | https://blaxel.ai/pricing ; changelog 2025-10-21 |
| Volumes | Block volume attached at creation (single-attach, cannot detach) | **Provisioned** size, for as long as it exists, regardless of data written | $0.12/GB-month | https://blaxel.ai/pricing |
| Agent Drive (private preview, us-was-1) | Shared RWX FUSE FS | Free during beta | $0 (beta) | https://blaxel.ai/pricing |
| Network | Internet egress, internal traffic, LLM gateway | Included | $0/GB; dedicated egress IPs, egress gateway runtime, proxy data: free during beta; no public inbound IPv4 offered | https://blaxel.ai/pricing |
| Wildcard custom domain | Sandbox preview URLs on *.your.domain (tier-gated) | Per domain-month (billing dimension `custom_domain_active_time_hours`) | $20/domain-month; FQDN domains "coming soon" | https://blaxel.ai/pricing ; /api-reference/billing |
| Quota Tier 0 (free) | Default; free credits only | No fee | 10 sandboxes; forced TTL ≤7 d; free accounts cannot request quota increases | https://blaxel.ai/pricing ; /Security/Quotas |
| Quota Tier 1 | ≥$20 real top-ups in trailing 30 days | Top-up = prepaid wallet credit, not a fee | 50 sandboxes; TTL ≤30 d | same |
| Quota Tier 2 | ≥$50 top-ups / 30 d | same | 200 sandboxes; unlimited standby persistence | same |
| Quota Tiers 3–8 | higher top-ups | same | **not public** ("view in console") | same |
| Quota Tier 9 | Contact | same | 100,000+ sandboxes | same |
| Custom (enterprise) | Sales | Quote; "volume pricing" | Up to 256 GB RAM/sandbox, private networking, bring-your-own-metal, custom SLA — prices null | https://blaxel.ai/pricing |
| Add-on: Email support | Opt-in | $800/month **+ 3 % of usage** | | https://blaxel.ai/pricing |
| Add-on: Dedicated support (Slack, SAML) | Opt-in; required for SAML SSO | $1,600/month **+ 10 % of usage** | | https://blaxel.ai/pricing |
| Add-on: HIPAA BAA | Opt-in | $250/month flat | | https://blaxel.ai/pricing |
| Free credits | New account | "Up to $200" one-time; promo credits do NOT count toward tier top-up volume | | https://blaxel.ai/pricing ; /Security/Quotas |

### Dated changes
- 2025-02-12: Beamlit (former name) launched plan-based pricing.
- 2025-08-01: "Launched our new usage-based pricing model" (current GB-s model). Third-party pages still cite the older named sizes XS (2 GB, $0.0828/h) … XL (32 GB, $1.3248/h) — same $0.0414/GB-h slope.
- 2025-11-07: started charging snapshot and volume storage (previously free). 2025-12-01: started charging image storage.
- 2026-09-10: Baseten acquisition; no price change announced.

## Gotchas
1. **CPU utilisation is irrelevant; connection state is everything.** A sandbox at 1 % CPU with an open WebSocket pays 100 % of RAM rate; an idle WS/TCP connection keeps billing for up to 15 min before forced standby. keepAlive processes (timeout 0 = forever) keep it billing indefinitely.
2. **Tail billing on every session:** ~15 s (docs) of active billing after the last connection closes before standby. Minimum billed duration per resume is not documented.
3. **CPU is capped by memory:** you cannot buy 4 vCPU with 2 GB. 4 vCPU requires 8 GB; 6 cores requires 16 GB. Above 16 GB the core count is undocumented; 256 GB only on Custom.
4. **Writable disk lives in RAM** (tmpfs ≈ 50 % of memory). Needing more disk means either more RAM (billed per second active) or a Volume billed on provisioned size forever.
5. **Standby is not free**: snapshot ≈ allocated memory (vendor: 2 GB standby month = $0.40). 50 persistent 8 GB sandboxes idle all month ≈ 400 GB × $0.20 = $80/month.
6. **Sandbox quota counts standby sandboxes too** ("keeping inactive sandboxes in standby … counts against your quota limits"). A persistent fleet of 50 sits exactly at the Tier 1 cap; any extra forces Tier 2. There's also an account-wide memory quota per tier (QUOTA_EXCEEDED "maximum memory limit for the account"), values unpublished.
7. **Tiers are a rolling 30-day top-up volume**, not balance. Stop topping up and you are downgraded even with a large wallet balance, which may block creation. Promo/startup credits never unlock tiers.
8. **Tier 0/1 force TTL deletion** (7/30 d): "perpetual standby" really requires Tier 2 (≥$50/mo top-ups).
9. **Support add-ons are percent-of-usage surcharges** (3 % / 10 %) on top of the flat fee — SAML SSO is only via the $1,600 + 10 % add-on.
10. Batch Jobs are 48 % cheaper per GB-s than sandboxes for the same microVMs, but have no standby/state and 24 h max.
11. Manual snapshots outlive their sandbox and never expire, so they bill until deleted.
12. Archive storage rate unpublished; archive is private-preview only (403 without flag).

## Worked example
Workload: 4 vCPU / 8 GiB → Blaxel 8 GB sandbox (= 4 cores, exact fit). 50 concurrent × 8 h/day × 22 days = 8,800 sandbox-hours.
30 % CPU util (irrelevant to billing). 50 GiB snapshots retained. 100 GiB egress (free). Tier 1 needed for 50 concurrent (the $20 top-up is
consumed by usage, so adds $0). Image storage excluded (image size unknown; $0.045/GB-month).

| Regime | Compute | Storage | Egress | Add-on | **Monthly total** |
|---|---|---|---|---|---|
| A. Sandbox, connection held for whole 8 h session (Tier 1) | 8,800 × $0.3312 = $2,914.56 | 50 × $0.20 = $10.00 | $0 | — | **$2,924.56** |
| B. Sandbox, aggressive scale-to-zero — *illustrative*: only 30 % of session wall time has an active connection (2,640 h) | $874.37 | $10.00 | $0 | — | **$884.37** |
| C. Persistent fleet: 50 sandboxes never deleted, standby 554 h/month each at ~8 GB snapshot (instead of 50 GiB) | $2,914.56 | 400 × $0.20 × 554/730 = $60.71 | $0 | — | **$2,975.27** (needs Tier 2 for >30 d retention; top-up absorbed) |
| D. Batch Jobs (non-interactive; RAM billed as allocated 8 GB — upper bound) + 50 GB volume for retained state | 8,800 × 8 × $0.0216 = $1,520.64 | 50 × $0.12 = $6.00 | $0 | — | **$1,526.64** |
| E. MCP hosting rate (for reference only; 10-min request cap makes it unfit) | 8,800 × 8 × $0.0252 = $1,774.08 | $10.00 | $0 | — | $1,784.08 |
| F. A + Email support | $2,914.56 | $10.00 | $0 | $800 + 3 % × $2,924.56 = $887.74 | **$3,812.30** |
| G. A + Dedicated support (needed for SAML) | $2,914.56 | $10.00 | $0 | $1,600 + 10 % = $1,892.46 | **$4,817.02** |
| H. A + HIPAA BAA | $2,914.56 | $10.00 | $0 | $250 | **$3,174.56** |
| I. Custom / enterprise | quote (volume pricing) | | | | null |

Free "up to $200" one-time credit would reduce month 1 by ≤$200 (does not count toward tier).