# Together Code Sandbox

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

## Pricing regimes (raw)

- **Code Sandbox VMs on a Together custom plan (sales only; self-serve = codesandbox-sdk card)** (resource): $0.0446/vCPU-h, $0.0149/GiB-h [sales]
- **Together Code Interpreter ($0.03 per 60-minute session)** (sizes): session-60min ? vCPU/? GiB $0.03/h

## Features

Yes: snapshots, memory snapshots, fork/clone, pause/resume, persistent disk, 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, code interpreter, HTTPS preview URLs, custom domains, open internet, SOC 2, SSO, Python, Node.js, Go, Rust, wake on request, live resize, live fork (no pause), memory fork, EU data residency, spend limits, MCP server, automatic snapshots, snapshots on demand, gVisor or VM (no shared kernel), inbound access rules, firewall inside (nftables)

No: volumes, nested virtualization, browser, desktop GUI, computer-use API, browser + desktop control, anti-bot stealth, CAPTCHA solving, residential IPs, preinstalled agents, SSH, public IPv4, raw TCP inbound, egress allowlist, static egress IP, self-hosting, BYOC, open source, HIPAA, GPU, arm64, Windows, macOS, GPU desktop, credential injection, HTTP method/path egress rules, webhooks, agent harness API, hosted agent API (their own agent), secret proxy, secret proxy for any API, extra volumes, shared volumes

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

## Caveats

- The VM mode is sales-only (Together custom plans) and restricted to that plan, so it is not priced here; the self-serve route to the same VMs is the codesandbox-sdk card (Build free / Scale $170). The former bulk-credit mode (assumed 50% off, from CodeSandbox Enterprise) was removed: it duplicated codesandbox-sdk pricing on a guess.
- Two price expressions for the same VMs: per-resource list rate ($0.0446/vCPU-h, $0.0149/GiB-h) vs CodeSandbox credit tiers ($0.01486/credit); they agree within ~0.1% on 2 GB/core tiers. Only fixed tiers actually exist.
- Code Interpreter: $0.03 per 60-min session, resource spec unpublished (sizes vcpu/ram null). Self-serve on the pay-as-you-go plan; not flagged alt. Session state persists within the hour; billed per session not per call.
- Together is fully prepaid ($5 minimum, no free trial credit); API suspended at zero balance.
- Storage (hibernation snapshots), egress, IPv4 pricing not published (null). Hibernated VMs consume no credits in CodeSandbox's model.
- Regions not published (null).
- perf null: benchmarks.json has no together-code-sandbox entry; same infrastructure as providers.codesandbox (cold start 2,329 ms, burst 7,658 ms, dax 81.2 s).
- Hibernated sandboxes are kept on disk up to 7 days, then may be archived to long-term storage (resume 20-60 s instead of 0.5-2 s). No storage fee for hibernated/archived state is published (snapshot_gib_month stays null).
- Official sources disagree on the smallest tier and credit price: Together docs and codesandbox.io/docs list Pico = 2 cores / 1 GB at $0.01486/credit ($0.0743/h), while codesandbox.io/pricing lists Pico = 1 core / 2 GB and credits at $0.015 (Nano $0.15/h, 'Without subscription savings'). The Together per-resource list rate prices Pico (2c/1GB) at $0.1041/h, 40% above the credit price; it matches the credit price only on 2 GB-per-core tiers.

## How this provider charges


Together AI owns CodeSandbox and sells two sandbox products.

1. **Together Code Sandbox** is the CodeSandbox SDK VM product (same `@codesandbox/sdk`, same Firecracker infrastructure, API key from codesandbox.io). It is "currently available on **Together's custom plans**" only.
   - Together's pricing page lists per-resource rates ($0.0446/vCPU-h, $0.0149/GiB-h).
   - Together's docs bill it in CodeSandbox VM credits ($0.01486/credit, fixed sizes, per started minute).
   - The only self-serve path is a codesandbox.io workspace. That path is priced in `codesandbox-sdk`, not duplicated here.
2. **Together Code Interpreter (TCI)** is a hosted Python execution API at **$0.03 per session**, where a session lives 60 minutes. It is self-serve on prepaid Together credits.

## Regime table

| # | Regime | When it applies | How billed | Numbers | Source |
|---|---|---|---|---|---|
| 1 | Code Sandbox, Together custom plan (per-resource list rate) | Together sales contract | Per vCPU-h + per GiB-h on allocated resources while running. Per-minute granularity per the docs | **$0.0446/vCPU-h, $0.0149/GiB-h**. 4 vCPU/8 GiB = **$0.2976/h** | https://www.together.ai/pricing |
| 2 | Code Sandbox, credit accounting (docs) | Same product as described in Together docs | Fixed sizes in VM credits at $0.01486/credit. Smallest unit 1 minute, rounded up | Pico 5 cr/h $0.0743 … Micro (4c/8GB) 20 cr/h **$0.2972**, XLarge 320 cr/h $4.7552. Differs from row 1 by ≤$0.0004/h (row 1 lets RAM/vCPU vary, but only fixed tiers exist) | https://docs.together.ai/docs/together-code-sandbox |
| 3 | Enterprise credit discount | Custom plan | CodeSandbox Enterprise: "up to 50% off bulk VM credit packs" | Best case Micro $0.1486/h. Actual terms negotiated | https://codesandbox.io/pricing, https://codesandbox.io/docs/sdk/pricing |
| 4 | Concurrency by plan | — | Build 10, Scale 250, Enterprise/custom = custom (the docs repeat CodeSandbox's plan table) | — | docs.together.ai/docs/together-code-sandbox |
| 5 | Self-serve fallback | No Together contract | Open a CodeSandbox workspace (Build free / Scale $170/mo). See `regimes/codesandbox-sdk.md` | — | docs.together.ai/docs/together-code-sandbox |
| 6 | Hibernated / shut down | Same lifecycle as CodeSandbox SDK | No compute. No storage fee published | $0 (unpublished) | codesandbox.io/docs/sdk |
| 7 | **Code Interpreter session** | `/tci/execute` API | **$0.03 per session**. A session has a 60-minute lifespan and can be reused for many executions (state, packages and variables persist). Billed per session, not per call. Resource spec **unpublished** | $0.03/60-min session = $0.03/h equivalent if chained | https://www.together.ai/pricing, https://docs.together.ai/docs/together-code-interpreter |
| 8 | Together prepaid credits | Any Together self-serve use (TCI) | Prepaid. **$5 minimum purchase**. API access suspended at a zero balance. Auto-recharge optional. Credits never expire. **No free trial credit** | $5 minimum | https://docs.together.ai/docs/billing |
| 9 | Storage / egress / IPv4 | — | Not published for either product. No public IPv4 | null | together.ai/pricing |

## Gotchas

1. **There is no self-serve Together path for VMs.** "Together Code Sandbox" at the listed per-vCPU rate needs a sales-negotiated custom plan. Self-serve means CodeSandbox's own plans: a $170 Scale fee for more than 10 concurrent VMs.
2. **Two price expressions exist for the same VM.** They are per-resource ($0.0446/$0.0149) on together.ai and per-credit fixed tiers in the docs. They agree to within 0.1% for the 2 GB/core tiers. The per-resource form overstates flexibility, since only fixed sizes exist (Pico 2c/1GB, Nano 2c/4GB, then 2 GB per core up to 64c/128GB).
3. **Code Interpreter is session-priced, not resource-priced.** $0.03 buys a 60-minute session regardless of use. That is cheap for bursty notebook-style LLM code execution, but there is no published CPU/RAM, no Docker and no custom image. It is not a general VM.
4. **Together is fully prepaid.** At a zero balance the API is suspended, which kills running sessions. The minimum top-up is $5.
5. **Per-started-minute rounding** is inherited from CodeSandbox for VMs.

## Worked example

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

| Regime | Compute | Snapshots 50 GiB | Egress 100 GiB | Plan fee | **Monthly total** |
|---|---|---|---|---|---|
| Together custom plan at list per-resource rate | 8,800 × $0.2976 = $2,618.88 | $0 (hibernation unmetered, unpublished) | null | unpublished (custom) | **$2,618.88 + contract fee?** |
| Together custom plan, credit accounting (Micro $0.2972) | $2,615.36 | $0 | null | unpublished | **$2,615.36 + ?** |
| Custom plan with best-case 50% bulk credit discount | $1,307.68 | $0 | null | unpublished | **$1,307.68 + ?** |
| Self-serve via CodeSandbox Scale | see codesandbox-sdk.md | | | $170 | **$2,761.58** |
| Code Interpreter (not the same machine) | 50 × 8 sessions/day × 22 = 8,800 sessions × $0.03 = $264 | n/a | null | $0 (prepaid) | **$264.00** (illustrative only: spec unknown, not 4/8, no Docker) |

- 30% utilisation saves nothing on the VM regimes, which are billed on allocation.
- TCI is billed per session, so utilisation doesn't matter there either.

Sources: https://www.together.ai/pricing · https://docs.together.ai/docs/together-code-sandbox · https://docs.together.ai/docs/together-code-interpreter · https://docs.together.ai/docs/billing · https://docs.together.ai/reference/tci-execute · https://codesandbox.io/pricing · https://codesandbox.io/docs/sdk/pricing