# Upstash Box

Official pricing: https://upstash.com/pricing/box  
Category: agent-sandbox · Isolation: container

## Pricing regimes (raw)

- **Pay as You Go, Small box (2 vCPU / 4 GB), active CPU core-hours** (resource): $0.1/vCPU-h, $0/GiB-h
- **Pay as You Go, Medium box (4 vCPU / 8 GB), active CPU core-hours** (resource): $0.2/vCPU-h, $0/GiB-h
- **Pay as You Go, Large box (8 vCPU / 16 GB), active CPU core-hours** (resource): $0.4/vCPU-h, $0/GiB-h
- **Fixed / Keep-Alive Small box, always on ($8/box/month)** (sizes): small 2 vCPU/4 GiB $0.010959/h
- **Fixed / Keep-Alive Medium box, always on ($16/box/month)** (sizes): medium 4 vCPU/8 GiB $0.021918/h
- **Fixed / Keep-Alive Large box, always on ($32/box/month)** (sizes): large 8 vCPU/16 GiB $0.043836/h
- **Keep-Alive Small fleet sized to peak concurrency (part-time use)** (pool)
- **Keep-Alive Medium fleet sized to peak concurrency (part-time use)** (pool)

## Features

Yes: snapshots, fork/clone, pause/resume, persistent disk, idle auto-stop, ≥24 h sessions, custom image (Docker or snapshot), start from your own snapshot, browser, preinstalled agents, SSH, HTTPS preview URLs, egress allowlist, open internet, SOC 2, arm64, Python, Node.js, Go, Rust, credential injection, wake on request, live fork (no pause), webhooks, scheduled runs (cron), MCP server, snapshots on demand, agent harness API, hosted agent API (their own agent), inbound access rules, secret proxy, secret proxy for any API

No: memory snapshots, volumes, your own Docker/OCI image, full VM (own kernel), Docker inside, nested virtualization, desktop GUI, computer-use API, code interpreter, browser + desktop control, anti-bot stealth, CAPTCHA solving, residential IPs, public IPv4, raw TCP inbound, static egress IP, private networking, self-hosting, BYOC, open source, HIPAA, SSO, GPU, Windows, macOS, GPU desktop, HTTP method/path egress rules, live resize, memory fork, EU data residency, automatic snapshots, fast boot (benchmarked), gVisor or VM (no shared kernel), firewall inside (nftables), extra volumes, shared volumes

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

## Caveats

- 'Active CPU hour' resolved as a core-hour of actual consumption priced by box size (Small $0.10, Medium $0.20, Large $0.40 per core-hour): the FAQ ($0.20 for 100% of 2 cores for 1 h; $0.01 for 10% of 1 core) and Upstash's own comparison blog (0.167 core-h x $0.10 = $0.017 on a 2 vCPU box) multiply core-hours by the size rate. The FAQ/blog only give Small examples, so the Medium/Large per-core rate is the table value applied per core (not confirmed by a worked example). Alternative readings: uniform $0.10/core-h (Medium full load $0.40/h) or price per box-hour at full load (Medium $0.20/h = $0.05/core-h). The engine uses the per-core size rate (most expensive, $0.80/h for Medium at 100%).
- Active basis: engine bills vcpu x util x rate using the rounded box vCPU count (e.g. a 1 vCPU workload on a Small box is billed as 2 x util); reality bills only cores actually consumed, so small workloads are slightly overestimated.
- Keep-alive is cheaper than PAYG even for part-time use once a box is active more than ~67 h/month at 30% CPU (Medium: $16 / (4 x 0.3 x $0.20)). The keepalive-fleet pool modes approximate a persistent fleet of C boxes (tiers 1/5/10/25/50/100/250/500/1000); real cost is exactly C x $8/$16/$32. Engine must respect each fleet mode's max_vcpu/max_ram_gib (one session per box).
- Keep-alive proration (create/delete mid-month) and fair-use limits at sustained 100% CPU are unpublished. Whether snapshots taken from a keep-alive box are covered by the flat price is unclear ('no separate storage charges' on that box); modeled as billed at $0.10/GB-month.
- Storage $0.10/GB-month applies to 'all data on disk, including snapshots', presumably actual used GB (not the 5/10/20 GB limit); the pricing card's '5 GB Included Storage' is read as the disk size, not a free allowance (included_disk_gib 0). Paused/frozen boxes keep billing storage; retention limit unpublished.
- No RAM charge on any plan. Idle-but-running boxes cost $0 CPU; the idle timeout (6 h PAYG / 1 h Free) only affects wake latency, not cost.
- Egress/ingress not published (null); docs cite 22.5 Gbps network per host. No public IPv4 (HTTPS preview URLs + SSH gateway only).
- Billing granularity and minimum billed time unpublished (null).
- Single region: AWS us-east-1. Enterprise can request other regions.
- Custom Docker images 'coming soon'; environments are reused via filesystem snapshots (restorable into a different size). EphemeralBox variant (exec+files only, TTL max 3 days, no agent/git/public URL/pause) has no separate price published; assumed same active-CPU rates.
- arm64 set true because the computesdk DAX run reports an aarch64 host (48 logical CPUs / 185 GiB visible inside the container); features research lists x86_64. The in-box CPU/RAM view shows host resources, so the cgroup cap vs billed cores is unverified.
- Keep-Alive pricing moved from a tooltip to its own 'Fixed' plan column on 2026-09-27 (upstash-web PR #662); no price change.
- perf from benchmarks.json (computesdk cold start 445 ms median, burst 5,633 ms median at 100 concurrency, DAX 53.4 s). No hpc CPU benchmark.
- Verified 2026-09-28 against https://upstash.com/pricing/box.md and https://upstash.com/docs/box/overall/pricing: sizes, $0.1/$0.2/$0.4 per active CPU hour, $8/$16/$32 keep-alive, $0.10/GB-month storage incl. snapshots, 1 h / 6 h idle timeouts, 10 / 1,000 concurrency all confirmed. Per-core-by-size reading still only evidenced by Small-rate examples.

## How this provider charges


Upstash Box is a Docker container per box (AWS us-east-1 only) with a built-in coding agent. There are three fixed sizes: Small 2 vCPU/4 GB/5 GB, Medium 4/8/10 and Large 8/16/20. RAM is never billed. The same box can be paid for in two very different ways:
- **Pay as You Go (PAYG):** you pay for CPU core-hours actually consumed, plus $0.10/GB-month storage.
- **Keep-Alive ("Fixed"):** a flat $8/$16/$32 per box per month covers CPU at any utilisation and the box's storage.

The Keep-Alive pricing moved out of a tooltip into its own "Fixed" plan column on 2026-09-27 (upstash-web PR #662). There was no price change.

## Resolving "active CPU hour" (three readings)

The per-size table lists Small $0.10, Medium $0.20 and Large $0.40 "per active CPU hour". There are three ways to read that for a 4 vCPU Medium box at 100% CPU:

| Reading | Meaning | Medium at 100% | Evidence |
|---|---|---|---|
| **A (adopted)** | Per **core-hour** consumed, at a rate set by box size | 4 × $0.20 = **$0.80/h** | FAQ: "measured in core-hours… 100% of 2 cores for one hour costs $0.2, 10% of a single core for one hour costs $0.01". Upstash's comparison blog says "Active core-hours only… $0.10/$0.20/$0.40 per active CPU-hour for small (2 vCPU), medium (4), large (8)" and works an example as "0.167 core-h × $0.10 = ~$0.017" on a 2 vCPU box. The site source (`box.ts`) stores `cpuHourPrice` per size |
| B | Uniform $0.10 per core-hour. The table is just "4 cores × …" marketing | $0.40/h | Consistent with the FAQ numbers but contradicts the table's per-size rates |
| C | Per **box-hour** at full load, so the price scales with cores (= $0.05/core-h) | $0.20/h | Contradicted by the FAQ: 2 cores at 100% would be $0.10, not $0.20 |

Reading A is the only one consistent with both the table and the FAQ/blog arithmetic. The only doubt left is that every official worked example uses the Small rate. The card uses A. B and C appear only in the worked example.

## Regime table

| # | Regime | When it applies | How billed | Numbers | Source |
|---|---|---|---|---|---|
| 1 | PAYG Small | Default box, size `small` | Core-hours of actual CPU use. RAM free | $0.10/core-h. Full load $0.20/h | https://upstash.com/pricing/box, /pricing/box.md |
| 2 | PAYG Medium | `size: "medium"` | Same, at a higher per-core rate | $0.20/core-h. Full load **$0.80/h**. At 30%: $0.24/h | same |
| 3 | PAYG Large | `size: "large"` | Same | $0.40/core-h. Full load **$3.20/h** (the per-core rate is 4× Small's) | same |
| 4 | Running but idle | Box up, no CPU use | **$0 compute**. Storage still accrues | $0 | FAQ "When a standard box is idle with no CPU usage, no charges apply" |
| 5 | Frozen / paused (auto) | After 6 h idle on PAYG (1 h on Free). Processes stopped, filesystem kept, auto-wakes on the next request | Storage only | $0.10/GB-month. Retention limit unpublished | /pricing/box, https://upstash.com/docs/box/overall/how-it-works |
| 6 | Keep-Alive Small / Medium / Large | `keepAlive: true`. Never pauses. `initCommand` on start | **Flat per box per month.** "This is the only charge for that Keep Alive box": CPU (any utilisation) and storage included | **$8 / $16 / $32** per month (= $0.011 / $0.022 / $0.044 per hour at 730 h) | /pricing/box, https://upstash.com/docs/box/overall/keep-alive |
| 7 | Keep-Alive fleet used part-time | Keep C persistent keep-alive boxes and reuse them for your sessions | C × flat price, whatever the hours used | Medium beats PAYG above 20 h/month at 100% CPU, or 67 h at 30% | derived from #2 and #6 |
| 8 | Storage (PAYG) | "All data on disk, including snapshots", for live and paused boxes | Per GB-month (presumably on data used, not the size cap) | $0.10/GB-month. Disk caps 5/10/20 GB | /pricing/box FAQ |
| 9 | Snapshots | Filesystem + agent config. They survive box deletion and can be restored into a different size | Counted in storage | $0.10/GB-month. No retention limit published | /pricing/box, https://upstash.com/docs/box/overall/snapshots |
| 10 | EphemeralBox | `EphemeralBox.create()`: exec + files only, TTL ≤ 3 days then auto-deleted. No agent, git, public URLs or pause | No separate price published. Presumably the same active-CPU rates | null | https://upstash.com/docs/box/overall/ephemeral-box |
| 11 | Free plan | No card | **Hard cap** of 5 active CPU-hours/month (≈ $0.50 at the Small rate). Small only, 10 concurrent, 1 h idle freeze, storage free. $1 LLM budget, then the API returns 400 | $0 | /pricing/box |
| 12 | Pay as You Go plan | Card on file | No fee. 1,000 concurrent boxes (soft limit, raised on request). $100/month LLM budget | $0 fee | /pricing/box |
| 13 | Enterprise | Contact sales | Custom sizes, limits, regions, prices. Dedicated support | null | /pricing/box |
| 14 | Built-in agent LLM tokens | Agent runs using Upstash's key | Included up to $1 (Free) or $100 (PAYG/Fixed) per month, then BYOK or contact sales | — | /pricing/box FAQ |
| 15 | Egress / ingress / IPv4 | Any | **Not published.** No public IPv4 (HTTPS preview URLs + SSH gateway). 22.5 Gbps per host | null | /docs/box/overall/how-it-works |
| 16 | Regions | us-east-1 only | No multipliers | 1× | /pricing/box FAQ |

## Gotchas

1. **The per-core rate rises with box size.** A Large box costs 4× Small per core-hour, so run a workload on the smallest box that fits. The same 1 core-hour of work costs $0.10 on Small and $0.40 on Large.
2. **Keep-Alive is massively cheaper than PAYG for anything busy.** A Medium at 100% for a month costs $584 on PAYG against $16 on Keep-Alive (36×). Even part-time fleets usually win: Keep-Alive beats PAYG once a Medium box is active more than about 67 h/month at 30% CPU. The catch is that the box never resets, so state and processes carry over between sessions.
3. **RAM is free on every plan,** and an idle-but-running PAYG box costs $0. The 6 h idle freeze is a latency matter, not a cost matter. Paused boxes pay storage only.
4. **Storage is billed on everything,** including snapshots that outlive their box and frozen boxes you forgot about. It is included only in Keep-Alive. Whether snapshots *taken from* a Keep-Alive box are also included is unclear.
5. **The Free plan is a hard wall.** It is 5 CPU-h, Small boxes only. After the $1 LLM budget the agent API returns 400. It is not a credit against paid usage.
6. **Unpublished:** billing granularity, minimum billed time, Keep-Alive proration for mid-month create/delete, fair-use limits on 100%-CPU Keep-Alive boxes, and egress.
7. **Custom Docker images are "coming soon".** Environments are templated through filesystem snapshots instead.
8. **The benchmark host is aarch64.** Inside the box it reports 48 logical CPUs and 185 GiB, so the container sees host resources. Research notes say x86_64.

## Worked example

Workload: 4 vCPU / 8 GiB → **Medium**. 50 concurrent × 8 h/day × 22 days = **8,800 box-hours**. At 30% CPU that is 8,800 × 4 × 0.3 = **10,560 core-hours**. Also: 50 GiB of retained snapshots and 100 GiB egress.

Assumptions:
- 50 persistent boxes, each with its full 10 GB disk: 50 × 10 × $0.10 = $50/month (an upper bound, since storage is on data used).
- Snapshots: 50 × $0.10 = $5.
- Egress unpublished ($0 assumed).
- PAYG plan: no fee, and 50 concurrent is within the 1,000 limit.

| Regime | Compute | Box storage | Snapshots | Egress | Fee | **Monthly total** |
|---|---|---|---|---|---|---|
| PAYG Medium, reading A ($0.20/core-h), **adopted** | 10,560 × 0.20 = $2,112.00 | $50.00 | $5.00 | null ($0) | $0 | **$2,167.00** |
| PAYG Medium, reading B ($0.10/core-h) | $1,056.00 | $50.00 | $5.00 | null | $0 | **$1,111.00** |
| PAYG Medium, reading C ($0.05/core-h) | $528.00 | $50.00 | $5.00 | null | $0 | **$583.00** |
| PAYG, boxes deleted daily (no idle disk), reading A | $2,112.00 | 10 GB × 176/730 × 50 × $0.10 = $12.05 | $5.00 | null | $0 | **$2,129.05** |
| **Keep-Alive fleet: 50 Medium boxes reused daily** | included | included | $5.00 (if billed) | null | 50 × $16 = $800 | **$805.00** |
| Keep-Alive, always-on 730 h (same 50 boxes) | included | included | $5.00 | null | $800 | **$805.00** |
| Free plan | n/a (Small only, 10 boxes, 5 CPU-h cap) | | | | | not viable |

Sensitivity:
- At 100% CPU, PAYG reading A would be 35,200 core-h × $0.20 = $7,040 + $55, while Keep-Alive stays at $805.
- Under reading A, the Keep-Alive fleet is cheapest for this workload at any CPU utilisation above about 11%: $800 / (8,800 box-h × 4 × $0.20).

Sources: https://upstash.com/pricing/box · https://upstash.com/pricing/box.md · https://upstash.com/docs/box/overall/how-it-works · https://upstash.com/docs/box/overall/keep-alive · https://upstash.com/docs/box/overall/ephemeral-box · https://upstash.com/docs/box/overall/snapshots · https://upstash.com/blog/upstash-box · https://upstash.com/blog/ai-agent-sandbox-providers-compared-2026 · https://upstash.com/blog/running-claude-code-in-a-remote-sandbox-with-upstash-box · https://github.com/upstash/upstash-web/pull/662 · https://github.com/upstash/upstash-web/blob/master/src/data/pricing/box.ts