Did Anyscale shut down?

No. Anyscale still operates the Ray platform and now sells GPU compute for Ray workloads, and in July 2026 it signed a definitive agreement to join Nscale. What retired was Anyscale Endpoints, the per-token hosted LLM API.

Can I still call Anyscale Endpoints with per-token pricing?

No. The Endpoints product was folded into the platform, and current pricing is sold as GPU hours for Ray workloads rather than per million tokens. Traffic that depended on per-token billing has to move to another inference provider.

What if I need Ray itself, not inference?

These five platforms replace the inference endpoint. If you need managed Ray clusters, that is a separate purchase — look at GPU cloud providers that run Ray, and treat the list above as the answer for model serving only.

5 Anyscale Endpoints Alternatives in 2026 (The Per-Token API Retired)

Pricing data as of Nov 7, 2026 - verify current rates on the official site before purchase

Anyscale Endpoints was the per-token hosted LLM API — you sent tokens, you paid per million, and Anyscale ran the model. That product has been retired and folded into the Anyscale platform, which now sells GPU compute for Ray workloads. Anyscale itself is still operating and in July 2026 signed a definitive agreement to join Nscale, so the company has not gone away; the per-token endpoint has. The five platforms below replace the endpoint, not Ray.

1. Baseten

Baseten is the closest match if what you actually needed was your own weights behind an API. You bring a model or a container, Baseten handles GPUs, autoscaling and the HTTP surface. It is priced as infrastructure rather than as a trial, so it fits production traffic better than experimentation.

2. Modal

Modal takes the "write code, get GPUs" approach: you define a Python function, it runs on a GPU when called, and nothing stays warm between requests. That suits the workloads people built around Anyscale's programmable layer — batch scoring, custom preprocessing, model pipelines — better than a fixed model catalogue does.

3. Replicate

Replicate is the fastest way to get a working endpoint again. You pick a public model, call it over HTTP, and pay for compute time rather than tokens. The catalogue is community-driven, so versioning and cold starts are things you have to plan around, but for a stopgap migration it is the least setup.

4. Hugging Face Inference

Hugging Face Inference gives you the widest model catalogue of the five, plus inference credits on a free account, which is useful while you work out which model replaces the one you were calling. Steady traffic moves to dedicated endpoints billed by the hour — the same shape as buying compute instead of tokens.

5. Together AI

Together AI is the volume option for open models: a large hosted catalogue, per-token billing, and fine-tuning on the same platform. If your Anyscale usage was mostly chat and completion traffic against open models, this is the most direct replacement of that line item.

*This post contains affiliate links, we may earn a commission at no extra cost to you.

📬 Get weekly AI coding tools in your inbox

We send 5 best new AI dev tools every week, no spam.

Browse LLM APIs