Anyscale

What it is for

Anyscale is a hosted Ray workspace: open a cluster, run distributed training or serving, and watch CPU / GPU / memory in the same console. Use it when the job is Ray at production scale, not when you need a single-GPU notebook.

It is not fully closed. Ray itself is open (ray-project/ray). Anyscale is the paid control plane. Billing is pay-as-you-go — no monthly fixed fee. Hosted CPU is AC 0.0135/hr; A100 is AC 4.9591/hr.

What makes it unique: the cover is Anyscale’s official Workspace still — Metrics for anyscale-saved-marriage, node CPU / GPU / memory charts, 2 nodes with 56 CPU / 4 GPU, Grafana and Ray Dashboard in the chrome.

Key features

• Workspaces — a managed Ray cluster you develop on like a laptop: VS Code / Jupyter / Git, live CPU·GPU·memory charts, workers that scale from one node to many GPUs without a rewrite (cover; Workspaces).
• Jobs — production batch (training, batch inference, Ray Data / ETL) with max_retries, job queues, and cron schedules; the cluster starts for the run and stops when it finishes (Jobs).
• Services (Ray Serve) — deploy continuously running HTTP endpoints with Anyscale rollouts; docs separate this from jobs (Ray Serve on Anyscale; Deploy services).
• Hosted or Bring Your Own Cloud — Anyscale-managed VMs, or Ray inside your VPC / on-prem / Kubernetes (Pricing; Workspaces).
• Structured logs — JSON logs with Ray ids, a Logs tab on workspaces / jobs / services, and export into your own stack (Structured logging).

Who it's for

Best fit: ML teams already writing Ray who want a workspace instead of a pile of raw VMs.

Poor fit: app builders who just need a serverless GPU function (Modal).

Pricing

Verified 2026-10-22 (Asia/Shanghai) from anyscale.com/pricing. Official line: only pay for compute, no monthly fixed fees. “AC” is Anyscale’s published compute unit on that table. Confirm what AC converts to in your invoice currency.

Hosted example rates (per hour)

• CPU only — AC 0.0135 /hr
• NVIDIA T4 — AC 0.5682 /hr
• NVIDIA L4 — AC 0.9542 /hr
• NVIDIA A10G — AC 1.3635 /hr
• NVIDIA A100 — AC 4.9591 /hr
• H / B / GB families — listed on the same table (open the page for the current AC).

Hosted vs BYOC

• Hosted — limited regions, Anyscale-managed VMs, monthly card invoice, business-hours support (5 cases).
• BYOC — any cloud or on-prem, VMs or K8s, your VPC, use existing GPU reservations, invoice or marketplace, 24×7 enterprise SLA.

Committed contracts unlock volume discounts. There is no public “Free $0 workspace” card on the page we fetched — Get started with $100 credit appears in the marketing chrome; treat that as a promo, confirm in-account.

Limitations / costs

• No published monthly seat. An idle workspace still burns instance AC until you terminate it (the cover has a Terminate control for a reason).
• AC is not USD on the table. Do not invent a dollar-per-hour we did not read.
• Hosted is limited regions. Multi-cloud / on-prem is BYOC.
• This is Ray ops, not a chat app. If you do not have a Ray job, you are on the wrong page.

Alternatives

• Modal — serverless GPUs if you want functions, not a Ray workspace.
• Together AI — hosted inference if you only need tokens, not training clusters.
• Fireworks AI — inference API if you are serving, not distributing Ray actors.

Official tutorial

Official Anyscale channel video: Anyscale Workspaces: A Scalable Interactive ML Development Environment with Zero Setup (oEmbed author: Anyscale).

Written path: Pricing plus Workspaces overview.

Website

https://www.anyscale.com

Source

Catalog row curated for AI Tools Nav. Cover is Anyscale’s official Workspace metrics still. Pay-as-you-go; Hosted CPU AC 0.0135/hr; A100 AC 4.9591/hr. Tutorial: Workspaces on @anyscale (oEmbed).

Last verified

2026-10-22 (Asia/Shanghai) editorial check.