# Together Code Sandbox

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

## Pricing regimes (raw)

- **Code Sandbox (together-sandbox SDK; enabled per organization through Together)** (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, ≥24 h sessions, custom image (Docker or snapshot), start from your own snapshot, your own Docker/OCI image, full VM (own kernel), root, code interpreter, HTTPS preview URLs, open internet, SOC 2, SSO, Python, Node.js, Go, Rust, memory fork, spend limits, MCP server, usage per sandbox (API), snapshots on demand, gVisor or VM (no shared kernel), inbound access rules, firewall inside (nftables), FUSE mounts, kernel features (eBPF, mounts, cgroups)

No: volumes, idle auto-stop, 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, wake on request, live resize, live fork (no pause), webhooks, automatic snapshots, 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

## Company

- Ownership: independent
- Stage: series-c [source](https://www.together.ai/blog/announcing-our-series-c)
- Accelerators: none found
- Total raised: $1.3B (confirmed: Sum of Together AI's own announced rounds: $20M seed + $102.5M Series A + $106M extension + $305M Series B + $800M Series C, all confirmed via together.ai blog posts.) [source](https://www.together.ai/blog/announcing-our-series-c)
- Last round: series-c, $800M, 2026-07 (confirmed) [source](https://www.together.ai/blog/announcing-our-series-c)
- Investors: General Catalyst, Prosperity7 Ventures, Kleiner Perkins, NVIDIA, Salesforce Ventures, Lux Capital, Emergence Capital, Coatue, Aramco Ventures, Vista Equity Partners, SE Ventures, Pegatron, March Capital, DTCP Growth, Geodesic, PSP Partners, DAMAC Capital, Greycroft, NEA, Definition Capital, Long Journey Ventures, Cadenza Ventures, Brave Capital, SK Telecom, Scott Banister, SV Angel, Factory, John Chambers
- Revenue: unknown

This card is for Together's Code Sandbox (sales-only, custom-plan Firecracker VMs) and Code Interpreter ($0.03/60-min session) products, built on CodeSandbox, which Together AI acquired in December 2024 (https://www.together.ai/blog/codesandbox-acquisition-together-code-interpreter). All ownership/funding/investor data here is for the parent company, Together AI, not a standalone entity for the sandbox product. Together AI's funding history: $20M seed (Lux Capital, undated on blog), $102.5M Series A (Nov 2023, led by Kleiner Perkins, with NVIDIA and Emergence Capital), $106M extension (Mar 2024, led by Salesforce Ventures), $305M Series B (Feb 2025, led by General Catalyst, co-led by Prosperity7), $800M Series C (Jul 2026). No lead investor was explicitly named for the Series C in Together's own post; no ARR/revenue or employee count was found.

## Caveats

- The VM mode is gated (together-sandbox "must be enabled for your organization" through Together), so it is not priced here. The deprecated CodeSandbox-based Code Sandbox (legacy page: Together custom plans, or self-serve via CodeSandbox) is priced on the codesandbox-sdk card.
- 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 (snapshots), egress, IPv4 pricing not published (null).
- Regions not published (null).
- perf null: benchmarks.json has no together-code-sandbox entry. The CodeSandbox benchmark row no longer applies: Code Sandbox is now the together-sandbox SDK.
- Code Sandbox moved from the CodeSandbox microVMs to the together-sandbox SDK (2026-10-14): no live fork, no wake on request, no idle hibernation (ttl counts from creation), no live resize, max 16 vCPU / 32 GB, all ports public unless an experimental inbound allowlist is set, preview URLs "coming soon", Dev Containers/Docker Compose not interpreted (https://raw.githubusercontent.com/togethercomputer/together-sandbox/main/docs/migrating-from-codesandbox.md).
- Sizing sources disagree: python-sdk.md allows 0.1–16 cores at 1–8 GB per CPU (up to 128 GB), the migration doc says "Max 16 vCPU / 32 GB"; the card uses the 32 GB cap. Exec default user is 1000:1000; root is per-exec opt-in.
- together.ai/pricing does not say whether the $0.0446/vCPU-h + $0.0149/GiB-h Code Sandbox rates apply to together-sandbox or the legacy CodeSandbox offering.
- Code Interpreter: its docs page is gone (the URL serves the Code sandbox page); only together.ai/pricing "Session (60 minutes) $0.03" still states the 60-minute session.

## 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