Ollama Cloud: Pricing, API Keys, Models, Free Use and Privacy

Ollama Cloud is Ollama's hosted inference service. Models with a :cloud tag run on Ollama's servers, not on your computer. On 14 November 2026 the plans were Pay as you go at $0, Pro at $20 a month, Max at $100 and Team at $500, all billed per token. You connect through a signed-in local Ollama or with an API key at https://ollama.com/api.

Ollama made its name as the simplest way to run a model on your own machine, so the cloud service causes a recurring mix-up: people assume anything started with ollama run stays local. It does not, if the model name ends in :cloud. This guide covers what the service is, what it costs, how keys and base URLs work, and what Ollama says about your data.

Every price, limit and policy line below was read on Ollama’s pricing page, documentation, blog, privacy policy and terms on 14 November 2026, when the newest Ollama release was v0.40.16 (9 November). Ollama changed its billing model in August, so older articles about GPU-time limits and five-hour windows describe a scheme that new subscribers no longer get.

What is Ollama Cloud?

Ollama Cloud runs open-weight models on data-centre hardware and returns the answer through the same API shapes as local Ollama. Ollama announced cloud models as a preview on 19 September 2025. Its documentation now simply calls the feature Cloud.

There are two ways in, and both need an Ollama account:

The first route is why the service is easy to use by accident. A cloud model appears in the same app, CLI and local API at http://localhost:11434 as a model on your disk, and tools pointed at the local port keep working. The only visible difference is the tag.

Ollama’s documentation notes a few functional limits: cloud models do not currently support structured outputs, the hosted Responses endpoint is stateless, and cloud models are set to their maximum context length by default, unlike local models, which start at 4K to 256K tokens depending on VRAM.

Ollama Cloud pricing and plans

Ollama’s pricing page listed these tiers on 14 November 2026:

How the billing works, according to the page’s FAQ and the announcement of 31 August 2026:

• Per token. Usage is measured in tokens at each model’s published rate. This replaced GPU-time billing, which Ollama said users found hard to predict.
• Credits first. Included plan credits are used before any purchased balance. Every plan, Pay as you go included, can buy extra usage credits.
• Monthly reset, no rollover. Included usage resets on the day of the month your subscription started, on annual plans too. Unused credits do not carry over.
• No session caps. The announcement says the new plans have no five-hour or weekly limits. When the pool is used up, requests continue at the same per-token rate.
• Queueing. Requests above your concurrency limit are queued, and rejected if the queue is full.
• A 90% warning. Paid plans get an email at 90% of included monthly usage.
• Legacy plans. Subscribers from before 31 August 2026 stay on the old scheme until they change billing cycle or tier, or cancel and re-subscribe. Anyone who had an active subscription at launch and switched can ask support to convert them back.
• One account each. The FAQ says Ollama is one account per person; a team membership creates a separate team account under the same login.

The page does not say which models count as larger pro models, so check a model’s own page before assuming a Pay as you go account can call it.

Is Ollama Cloud free?

There are three different free things here, and they are often confused.

Local Ollama is free. The software is MIT-licensed, and the pricing page says running models on your own hardware is always unlimited. No account is needed.

An Ollama account is free. The Pay as you go tier has no subscription and no service fee. The pricing page says these accounts pay for each request with usage credits.

A free cloud allowance is less clear. The 31 August 2026 announcement said the free plan now includes a small amount of monthly usage for a set of starter models, resetting monthly from the sign-up date. It gave no figure and did not name the models. The pricing page on 14 November 2026 called the $0 tier Pay as you go and listed no allowance. Treat any free cloud usage as a trial quantity and check the usage page in your account for what you actually have.

Two exceptions are documented. Ollama’s blog said on 16 October 2026 that Reflection’s Beam model was freely available for a limited time, although by 14 November the pricing page listed Beam at $0.125 input and $0.40 output per million tokens, so check whether the offer still applies. Web search through Ollama’s API needs only a free account and an API key. Ollama also notes that its Clef decision model in the cloud requires a paid plan or purchased credits.

If the aim is zero cost with no data leaving the machine, a local model is the answer. Our VRAM calculator shows what fits on a given GPU.

Ollama Cloud models and per-token prices

On 14 November 2026, https://ollama.com/api/tags returned 24 cloud models. The pricing page publishes input, cached-input and output prices per million tokens for each. A selection:

The two DeepSeek models have off-peak rates at half price; the page says off-peak applies outside 12:00 to 18:00 UTC on weekdays and all day at weekends.

To put the rates in context, take a coding session that sends 2 million input tokens and receives 200,000 output tokens, with no cache hits. On glm-5.3-flash that is $0.30 plus $0.10, or $0.40. On kimi-k3.1-pro it is $4.00 plus $2.00, or $6.00. Agent loops resend context on every step, so input tokens dominate and cached-input pricing matters more than the headline output price.

Other points from Ollama’s pages:

• Weights and precision. The FAQ says cloud models use native weights as released by the model provider, and may use accelerated formats such as NVFP4 on current NVIDIA hardware.
• Tool calling. Ollama says cloud models trained for tools are tested with real agent workflows before going live.
• Cloud-only models. Several of the largest models appear in Ollama’s library only with cloud tags. Kimi K3.1 Pro has 2.81 trillion parameters, which is why our guide to running Kimi K3.1 Pro yourself starts at eight data-centre GPUs.
• Retirements. Ollama retires older cloud models and lists replacements on a lifecycle page. The most recent batch, on 25 September 2026, removed qwen3.5:397b, glm-5.1 and deepseek-v4-flash:0731. Anything pinned to a retired name stops working, which does not happen to a model file on your own disk.

How to get an Ollama Cloud API key

1. Sign in at ollama.com.

2. Open the API keys page in your account settings.

3. Create a key and copy it.

4. Export it in your shell:

   export OLLAMA_API_KEY="your_api_key"

5. Send a request with the key as a bearer token:

   curl https://ollama.com/api/chat \
     -H "Authorization: Bearer $OLLAMA_API_KEY" \
     -H "Content-Type: application/json" \
     -d '{
       "model": "gemma4:31b",
       "messages": [{"role": "user", "content": "Say hello in one sentence."}],
       "stream": false
     }'

Ollama’s documentation says API keys do not expire. Revoke the ones you no longer use, and keep keys out of browser code and repositories.

Base URLs and endpoints

With the OpenAI Python client:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://ollama.com/v1",
    api_key=os.environ["OLLAMA_API_KEY"],
)

API key or public key?

Ollama has two credentials. An API key authenticates a program that calls ollama.com directly. The Ollama Public Key, stored as id_ed25519.pub in the .ollama folder, identifies a local install to ollama.com and is what ollama signin registers. You need the second only for cloud models through local Ollama, pushing models or pulling private ones.

To check spending from a script:

curl 'https://ollama.com/api/usage?range=24h' \
  -H "Authorization: Bearer $OLLAMA_API_KEY"
curl https://ollama.com/api/balance \
  -H "Authorization: Bearer $OLLAMA_API_KEY"

Using Ollama Cloud from coding agents

The quickest route is ollama launch, which configures a coding tool for an Ollama model, local or cloud, without editing config files:

ollama signin
ollama launch claude --model kimi-k2.7-code:cloud
ollama launch opencode
ollama launch codex

Claude Code can also skip the local install and call the cloud directly through the Anthropic-compatible endpoint. This is the form in Ollama’s documentation:

ANTHROPIC_BASE_URL=https://ollama.com \
ANTHROPIC_AUTH_TOKEN="$OLLAMA_API_KEY" \
ANTHROPIC_API_KEY="" \
claude --model glm-5.3-flash

Ollama notes that Claude settings files can override shell variables, and that hosted web search and advanced tool controls are not fully supported on this path. For Codex, the manual form is codex --oss -m gpt-oss:120b-cloud.

Two practical cautions. First, the same agent command works with a local tag, so a team can believe it is running a local set-up while one person’s default model is a cloud tag. Second, an agent can spend credits quickly, because every tool call resends the conversation. Check /api/usage after the first long session.

For local versions of these set-ups, see our guides to Claude Code on a local model and OpenCode with Ollama, LM Studio and llama.cpp. If you want one endpoint in front of several subscriptions and local servers, the OmniRoute guide covers a self-hosted gateway that can sit in front of Ollama.

Is Ollama Cloud private? What Ollama says it stores

Ollama’s statements are consistent across its documentation, pricing page, privacy policy (last updated March 2026) and terms (last updated August 2026):

That is a strong position for a hosted service. Three limits remain.

• It is a promise, not a property of the architecture. With a local model, the prompt never leaves your machine. With a cloud tag, it does, and you rely on the policy. Our zero data retention explainer covers what such commitments usually include and exclude.
• Location is not selectable. The pages we read describe no region choice. An EU organisation should treat cloud tags as a transfer to the United States and assess them accordingly.
• Contract terms are thin on the self-service plans. The pricing page mentions custom security questionnaires only for Enterprise. If you need a data processing agreement or sector-specific terms, ask Ollama before sending regulated data.

Our Ollama safety review looks at the local side: the unauthenticated port, published vulnerabilities and a hardening checklist.

Ollama Cloud vs local Ollama

How to keep Ollama local only

If policy says prompts must stay on the machine, switch the cloud features off. Ollama documents two ways. Add this to ~/.ollama/server.json:

{
  "disable_ollama_cloud": true
}

Or set an environment variable for the server:

OLLAMA_NO_CLOUD=1

Restart Ollama afterwards. The logs then show Ollama cloud disabled: true, and both cloud models and Ollama’s web search stop working. On shared machines, set this centrally and do not rely on each user avoiding cloud tags.

Which should you use?

Use local Ollama when the data is confidential, the model fits your hardware and you can live with its speed. Use Ollama Cloud for public or low-sensitivity work that needs a model too large to run locally, or to test a model before buying hardware for it. Past one person, the choice changes again: a shared local server needs authentication and capacity planning, which our guide to serving a team from one Ollama server covers, and heavier loads usually move to a different engine, as the inference server comparison explains.

How VDF AI fits

Ollama Cloud is one vendor’s hosted models with a retention promise and per-token billing. An organisation usually needs the opposite starting point: a list of which models may see which data, enforced before a request is sent, and a record afterwards. VDF AI does not host, resell or integrate with Ollama Cloud.

VDF AI Router is the model-routing layer of the VDF AI platform. It runs on your own VMs, Kubernetes or bare metal, registers a local Ollama server and other on-premises deployments next to the cloud models your policy approves, applies allow and deny lists before each call, and has an air-gap mode that keeps routing on local models only. Each decision is returned with its reason for the audit trail. The wider platform adds governed agents, private RAG and role-based access control on every plan.