Open vs Closed Models: How to Choose

An open-weights model is one whose trained weights you can download and run on your own hardware or any host. A closed model is reachable only through its maker's API or apps. Choose open when you need control, privacy or low token cost at volume. Choose closed when you want the strongest model with the least operations work. Many teams use both. Checked 31 October 2026.

The short version

What the words mean

Open weights. The numbers a model learned during training are published. Anyone can download them, run them, and sometimes modify them, subject to a licence. The training data and the training code are often not released, which is why many people say "open weights" rather than "open source".

Closed models. The weights stay with the vendor. You send a prompt to an API or an app and get an answer back. You cannot inspect the model, move it, or run it elsewhere.

Licence. This is where "open" gets slippery. A permissive licence such as MIT or Apache-2.0 lets you use the model commercially with few conditions. A custom licence may restrict some uses or add attribution rules. Always read the licence on the model page, not the headline.

Licences on real models

We read the licence tag on each model's Hugging Face page on 31 October 2026.

Two points. First, the same company can ship both: OpenAI is closed on its flagship GPT line and Apache-2.0 on gpt-oss. Second, "other" means a custom licence you must read. For example, the Llama 4 Community License asks for a "Built with Llama" display and a licence request above 700 million monthly active users. We did not read the Kimi or GLM licences.

Our model pages cover the families: open-source LLMs, DeepSeek, Qwen, Llama and gpt-oss.

Cost, in three ways to pay

1. A closed API

You pay per token. From the pricing pages: Claude Sonnet 5.5 is $2 input and $10 output per million tokens, Opus 5.5 is $4 and $20, and Fable 5.1 is $10 and $50. Haiku 5.5 starts from $0.10 and $0.50. See Anthropic API vs OpenAI API for the two big vendors side by side.

2. An open model on someone else's API

You pay per token to a host that serves open weights. DeepSeek's flash model lists $0.15 input and $0.60 output per million tokens off-peak, and $0.30 and $1.20 at peak. Its V4 Pro lists $0.66 and $1.98 off-peak. Gateways such as OpenRouter route to many hosts; see OpenRouter alternatives.

3. An open model on your own GPUs

You pay for the machine whether or not it is busy. Runpod's list rates on 31 October 2026 put an H100 SXM at $2.69 an hour on Community Cloud and $3.99 on Secure Cloud. Running one for a whole month (about 730 hours) is roughly $1,964 and $2,913, before storage. That is only worth it if the GPU is busy most of the day, or if privacy forces your hand. A rented GPU idle at night still bills. Compare hosts at /compare/hosting/, and read cloud vs self-hosted AI.

Where open models are stronger

Control over data. Self-hosted weights keep prompts inside your network. No third party sees your code or customer data.

Version stability. A closed API model can change or be retired. An open model you host does not change unless you change it.

Fine-tuning and customisation. With weights you can adapt a model to your domain, quantise it to fit smaller hardware, or merge it with others.

Price competition. Many hosts serve the same weights, so you can shop. If one raises prices, you move.

No vendor lock-in on the model. Your prompts and tooling stay portable.

Where closed models are stronger

Operations. No GPUs to buy, patch or monitor. The vendor handles capacity and uptime.

Product around the model. Closed vendors ship agents, apps, safety systems and support alongside the model. Claude Code and Codex are examples; compare them in Claude Code vs Codex.

Top-end capability. The flagship tiers, such as Fable 5.1 or GPT-6 Astra, are closed. Whether open models match them on your task is a question for your own test. We do not quote benchmark scores here, and we explain how to read them in the benchmark guide.

Simplicity. One key, one bill, one set of terms.

The middle path: both

Most teams end up using both, routed by task.

• Use a closed top model for hard reasoning, review and the work where mistakes are costly.
• Use an open or cheap model for bulk work: summaries, test generation, classification, first drafts.
• Keep the code behind one interface. DeepSeek, for one, accepts both OpenAI-style and Anthropic-style requests, so an agent can switch with an environment variable. We show how in DeepSeek vs Claude.

This also hedges. If a vendor changes terms or prices, you have a working second route.

A break-even check

Here is arithmetic you can redo with your own numbers. A month of one H100 SXM on Runpod Community Cloud is about $1,964 at the list rate. Divide that by an API price to see how many tokens the same money buys.

Read it carefully. If your monthly output is under those figures, the API is cheaper than the rented GPU on list price alone. The GPU also has to produce those tokens. How many tokens an H100 generates per month depends on the model, the serving engine and your batch size, and we do not publish a figure. Measure it on your own workload.

The numbers above also ignore the people cost. A server needs an owner, and an hour of that person's time each week is part of the real price. The practical conclusion: self-hosting rarely wins on price for a small team. It wins when you have enough steady volume to keep the card busy, or when a rule keeps data in house. Input tokens are cheaper than output tokens on every vendor above, so a workload heavy on reading and light on writing shifts the numbers further toward APIs.

A simple decision flow

1. Can the data leave your network? If no, self-host an open model, or use a host under a contract your lawyers accept. Stop here.
2. Is the task hard and the cost of a wrong answer high? If yes, start with a top closed model and test an open one against it.
3. Is the volume high and the task simple? If yes, price an open model on a hosted API first.
4. Do you need a pinned version or fine-tuning? If yes, you need weights.
5. Do you have someone to run servers? If no, stay on APIs.
6. Does the licence fit? Read it before you build on any model.

What you need to self-host

Before you commit, count the work:

1. Hardware. Enough GPU memory for the weights plus context. A large mixture-of-experts model may need several GPUs. Check the model card for the memory it needs.
2. A serving engine. Ollama for single users, vLLM for many; see Ollama vs vLLM and llama.cpp vs Ollama.
3. Monitoring and updates. Someone watches the server and applies fixes.
4. A security plan. An open model server on a network needs authentication.
5. A test set. Your own tasks, so you can tell when a model update helps or hurts.

If that list sounds like a lot, a hosted open model gives you most of the benefit with none of it.

Privacy and data handling

With a closed API, your prompts go to the vendor. Their terms decide retention and training. Read them for your plan. Each comparison page says where we did not read a vendor's data terms, such as Anthropic API vs OpenAI API.

With an open model on a hosted API, your prompts go to the host, not the lab. The host may be a different company with different terms. With self-hosting, nothing leaves, unless you send logs or telemetry somewhere. For fully local work, see private AI.

Limits and gotchas

• "Open" varies. MIT and Apache-2.0 are permissive. "Other" needs reading.
• Big open models are not small. A model that is open is not a model that fits on your laptop.
• Peak pricing. Some hosts charge double at busy hours; DeepSeek's peak window is 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays.
• Quality gaps are task-specific. An open model may match a closed one on your tests or fail on them. Do not assume either.
• Support. A closed vendor has a support line. An open model has a community.
• Provenance. Download weights only from the maker's official page.

Who should pick which

A solo developer

Closed through a subscription for the main agent, plus an open model through a cheap API for bulk jobs.

A startup

Closed API at first, for speed. Add an open model when your volume makes the bill matter.

A company with strict data rules

Open weights, self-hosted or on a host with a contract you trust, after a pilot.

A researcher or hobbyist

Open weights on your own hardware, to learn and tinker.

A team building an AI product

Design for both. Keep model names in config and test every release against your own evaluation set.

What routing saves, worked through

The middle path is easy to describe and easy to price. Take a team that generates 100 million output tokens a month. Use the output rates from the break-even table: Claude Sonnet 5.5 at $10 per million, DeepSeek flash off-peak at $0.60 per million. Output is the larger share of most agent bills, so we leave input out to keep the arithmetic readable.

Sending only the hard fifth to the closed model cuts this output bill by three quarters. The saving depends entirely on whether the cheap model passes on the work you give it, so the useful number is the share of tasks it handles at an acceptable pass rate. Measure that share on twenty real tasks before you build a router. A router built on a guess can cost more than it saves, once you count the retries and the review time for work the cheap model got wrong.

Reading a licence tagged "other"

Several rows in the licence table say "other." That is a prompt to read the text, not a verdict. Before you build a product on such a model, read for these points and write the answers down.

1. Commercial use. Does the licence allow it, and under what size or revenue limits?
2. Attribution. Must you name the model or its maker in your product?
3. Naming rules. Some licences require derived models to carry the original name.
4. Acceptable use. Is there a list of uses the licence forbids, and does your product touch it?
5. Redistribution. Can you ship the weights to a customer, or only run them yourself?
6. Change of terms. Does the licence apply to the version you downloaded, and can the maker change it for later versions?

Keep the licence file with your deployment record, since the page you read today can change. A model under MIT or Apache-2.0 needs a shorter read, and still deserves one.

Questions people ask

Does open mean free? The weights are free to download under the licence. Running them costs hardware or a host's per-token fee.

Can I switch from closed to open later? Yes, if you keep your code behind one interface and your prompts and tests in files you own. Expect to retune prompts, since models respond differently.

Is an open model more private? Only when you run it yourself. On a host's API, your prompts go to the host.

Our take

The right question is not which camp wins. It is which tasks need which model. Closed models still lead the vendor-support and product-bundle story, and open models lead on control and price at volume. The licence table above also shows that "open" covers a range, so check each model's licence.

We did not run a head-to-head of any open model against any closed one for this page, so we make no quality claims. If you take one action, run your own five hardest real prompts through one closed and one open model, and compare the outputs and the cost. That tells you more than any table.

Related pages

• Compare models
• Cloud vs self-hosted AI
• Qwen vs Llama
• Best open-source LLM for coding