Anyscale API 2026: Ray & Endpoints After the $1.65B Nscale Deal

Anyscale is the company behind Ray, the open-source distributed compute engine that scales Python AI workloads from a laptop to a multi-node GPU cluster, and Anyscale Endpoints, a serverless OpenAI-compatible inference API for open-weights models. On July 30, 2026 the NVIDIA-backed UK neocloud Nscale agreed to acquire Anyscale for a reported $1.65 billion ahead of its planned IPO — a deal that folds Ray and Endpoints into Nscale's full-stack AI cloud. This review covers the API, models, August 2026 pricing, the acquisition's practical impact, regional access, competitors and limits.

🧠 Model list

💰 Pricing (verified 2026-08-08)

Anyscale uses pay-as-you-go usage-based billing with no monthly minimum on the hosted tier, plus committed contracts that unlock volume discounts. Inference is billed per million tokens; Ray compute is billed per accelerator-hour.

🔌 OpenAI-compatible Endpoints

from openai import OpenAI

client = OpenAI(
    base_url="https://api.endpoints.anyscale.com/v1",
    api_key="YOUR_ANYSCALE_API_KEY"
)
response = client.chat.completions.create(
    model="meta-llama/Llama-3.3-70B-Instruct",
    messages=[{"role": "user", "content": "Explain Ray in one sentence."}]
)
print(response.choices[0].message.content)

Because the surface is OpenAI-compatible, any framework that speaks the OpenAI SDK (LangChain, LlamaIndex, AutoGen, CrewAI, Vercel AI SDK) works by swapping base_url. The same Anyscale account also runs full Ray jobs through the Python SDK — scale one endpoint, then scale out a distributed training or batch-inference job on the same platform without changing stacks.

🧠 Ray under the Nscale acquisition

Anyscale's defining asset is Ray, the de-facto standard distributed compute engine for AI, deployed across nearly every major cloud. On July 30, 2026 the NVIDIA-backed UK neocloud Nscale agreed to acquire Anyscale for a reported $1.65 billion (confirmed by Bloomberg, SiliconANGLE, Reuters and TechCrunch), folding Endpoints and the Ray platform into Nscale's full-stack AI cloud across data centers in Norway, the UK, Texas and Portugal. This followed Anyscale's June 2, 2026 native integration on Microsoft Azure for enterprises building sovereign AI and controlling variable API costs.

The stated goal is multi-cloud neutrality: Ray workloads can run on Nscale's own NVIDIA infrastructure and on AWS, GCP and Azure simultaneously, giving teams one Ray-consistent control plane across providers. For developers the near-term impact is continuity — the Anyscale API, SDK and pricing remain unchanged — plus a deeper pathway into Nscale's managed Slurm, Kubernetes Service and Inference Endpoints once a Ray job outgrows the hosted tier.

🆚 vs Together AI / Fireworks / Anyscale

On raw token price, Together AI and Fireworks are in the same open-weights band as Anyscale — all three hover around $0.30-$0.39/M for Llama 3.3 70B. Anyscale's edge is the Ray platform: no other open-weights host lets you go from a serverless chat endpoint to a multi-node distributed-training job on one account and one Python API. DeepSeek stays the clear price leader and offers direct mainland-China access, but ships text-only and has no first-party distributed compute offering. If you need only cheap hosted inference, Together or Fireworks are simpler; if you need inference and scalable Ray compute, Anyscale is the differentiated choice.

⚠️ Limits to know

• No mainland China endpoint — production access requires a proxy or an aggregator.
• Model catalog is open-weights focused; no first-party frontier proprietary models (compare GPT-5, Claude, Gemini).
• UI/UX and docs lean technical; the Ray learning curve is real for non-distributed users.
• Per-accelerator compute pricing (especially A100 at $4.96/hr) adds up for long-running jobs without committed contracts.
• Acquisition integration is ongoing — Nscale's roadmap for unifying Endpoints and its Inference Endpoints is not yet fully documented.
• No public batch discount beyond committed contracts; verify the enterprise SLA terms contractually.

🎯 Verdict

Choose Anyscale when you already live in the Ray ecosystem, need one account that spans serverless inference and distributed training/batch inference, or value multi-cloud neutrality through the Nscale deal. For pure cheap hosted open-weights inference, Together AI and Fireworks are simpler and fast. For direct mainland-China access, DeepSeek, Alibaba Bailian (Qwen) and Tencent Hunyuan remain the pragmatic picks. Anyscale is the strongest choice when your roadmap includes Ray compute and you want one API to grow with it.

FAQ

Is Anyscale OpenAI compatible?
Yes. Point the OpenAI client base_url at https://api.endpoints.anyscale.com/v1.

How much does the Anyscale API cost?
From $0.39/M for Llama 3.3 70B; DeepSeek-R1 $1.00/M; compute from $0.0135/hr CPU to $4.96/hr A100.

Is there a free tier?
No permanent free tier, but $100 one-time credits on signup plus small project starter credits.

Was Anyscale acquired?
Yes — NVIDIA-backed Nscale agreed to buy Anyscale for $1.65B on July 30, 2026.

Can it be used from China?
Not directly — it needs a proxy or aggregator; no mainland regional endpoint exists.

What is Ray used for?
Scaling Python AI workloads (distributed training, batch inference, online serving) from laptop to multi-node GPU cluster.