Open-weight models vs API models: which should Australian enterprises use?

Open-weight vs API LLMs: which should you use?

Start with a hosted API model unless you have a specific reason not to; move to open-weight when sovereignty, isolation, volume or control demands it. The decision is less about ideology and more about four practical questions: how capable does the model need to be, where may the data go, how much volume will you run, and who will operate it.

The terms first. An API model (Claude, GPT, Gemini) runs only on its vendor’s or a cloud partner’s infrastructure; you send a request and pay per token. An open-weight model (Llama, Mistral, Qwen, DeepSeek, OpenAI’s gpt-oss) has its trained weights published, so you can download it and run it wherever you like. “Open-weight” isn’t the same as open source: the training data and code usually aren’t released, and some licences restrict use. Buyers searching for an “open source LLM”, a “self-hosted LLM”, a “local LLM” or a “private LLM” usually mean an open-weight model run on infrastructure they control; API models are sometimes called closed or proprietary models.

How do they compare side by side?

API models win on capability and simplicity; open-weight models win on control and, at scale, on cost. Details reflect vendor pages as at 28 September 2026.

	Hosted API model	Open-weight, served by a cloud platform	Open-weight, self-hosted
Examples	Claude, GPT, Gemini via vendor API or Bedrock, Azure, Vertex AI	gpt-oss, Qwen, Mistral on Amazon Bedrock	Any open-weight model on your own GPUs
Capability ceiling	Highest: frontier models are API-only	Good to very good	Good to very good; you pick the exact model
Who sees your prompts	Model vendor or cloud provider	Cloud provider, not the model maker	Only you (and your cloud provider’s infrastructure)
Australian processing	Some models, on some platforms, at the time of writing	Some models listed in Sydney	Yes, in any Australian region with GPU capacity
Pricing	Per token	Per token	Per GPU hour, whether used or not
Operations	None	None	You run serving, scaling, patching, monitoring
Model version control	Vendor retires versions on its schedule	Platform retires versions on its schedule	You keep a version as long as you like
Fine-tuning	Limited, on selected models	Varies by platform	Full control
Offline or air-gapped	No	No	Yes

Where do API models win?

On raw capability, speed to launch, and low or unpredictable volume. The best models for complex reasoning, long documents, coding and agentic work are, at the time of writing, available only through APIs. You pay per token, so a pilot costs almost nothing, and you inherit the vendor’s scaling, safety work and continuous improvements.

Residency, once the main objection, has narrowed. Several API models can now process data inside Australia through Amazon Bedrock, Azure or Google Vertex AI, though availability varies by model and changes often. Our guide to which models can run onshore tracks the detail, and the Bedrock vs Azure vs Vertex comparison covers the platforms.

Choose an API model if:

• The task needs top-tier reasoning, writing or coding quality.
• Volume is low, spiky or unknown.
• You don’t have, and don’t want, people to operate GPU infrastructure.
• An Australian-region API option exists for the model, or offshore processing has been assessed and accepted.

When should you use a self-hosted LLM?

When control matters more than the last increment of capability. Open-weight models win on control: over where data goes, which exact model runs, how it’s tuned, and whether it runs at all without an internet connection. When you self-host, prompts and outputs never reach a model vendor, and your own logs prove where every request went.

Situations where that matters:

1. Data that can’t go to a third-party AI service under contract, client instructions, legal privilege or government classification.
2. Isolated or offline environments, such as secure facilities, remote sites with poor connectivity, or on-premises deployments.
3. Version stability. API vendors retire models; a self-hosted model behaves identically until you choose to change it, which matters for validated workflows.
4. Deep customisation. Full fine-tuning, custom decoding, or distilling a small specialised model.
5. Steady, high volume of simpler tasks, where owned capacity undercuts per-token prices.

Open-weight quality is good enough for much enterprise work. Many tasks, such as extracting fields from invoices, classifying emails or answering questions over retrieved documents, don’t need a frontier model. OpenAI’s gpt-oss-120b, for example, has 117 billion parameters and fits on a single 80 GB GPU; gpt-oss-20b runs within 16 GB of memory.

What about licences?

Most popular open-weight models allow commercial use, but the terms differ, so check before you build a product on one.

Model family	Licence at the time of writing	Things to note
OpenAI gpt-oss	Apache 2.0	Permissive, commercial use allowed
Qwen3 (for example Qwen3-32B)	Apache 2.0	Check each model card; licences vary across releases
Mistral (Small 4, Large 3, Ministral 3)	Apache 2.0	Mistral Medium 3.5 uses a modified MIT licence
Meta Llama 4	Llama 4 Community License	Attribution (“Built with Llama”), acceptable use policy, separate licence above 700 million monthly active users

Also consider provenance. Some organisations, particularly in government and critical infrastructure, have policies on the country of origin of AI models. That’s an organisational risk decision, but it needs making explicitly before you pick a model family.

When does self-hosting actually save money?

Only when a GPU is kept busy. Per-token pricing charges nothing when idle; a GPU instance charges every hour it runs. The break-even is simple arithmetic.

Here is an illustration using round, clearly hypothetical numbers. Use the live Bedrock, model vendor and EC2 pricing pages in the sources for real figures in your region.

Step	Illustrative assumption	Result
GPU instance able to serve your chosen model	$10 an hour (AUD)
Running 24/7	$10 × 730 hours	$7,300 a month
Add operations effort	0.2 of an engineer at $200,000 a year loaded	about $3,300 a month
Total monthly cost of self-hosting		about $10,600
API price for a comparable model	$5 per million tokens, blended input and output
Break-even volume	$10,600 ÷ $5 per million	about 2.1 billion tokens a month

Below the break-even, the API is cheaper; above it, owning capacity wins, provided the instance can actually serve that throughput. Real deployments shift the numbers: you may need redundancy (two instances), savings plans reduce GPU cost, prompt caching reduces API cost, and a smaller open-weight model may do the job on a cheaper GPU. Our LLM running costs guide works through these levers in detail.

The practical conclusion: for most organisations starting out, the API is cheaper. Self-hosting is justified first by control, and only later, if ever, by cost.

Is there a middle path?

Yes: run an open-weight model through a managed cloud platform in an Australian region. At the time of writing, AWS’s Bedrock regional table lists Sydney for several open-weight models, including gpt-oss-120b and models from Qwen and Mistral, while Llama 4 models are listed only in US regions. You pay per token, AWS operates the infrastructure, and the model maker never sees your data. You’re still on a foreign-owned cloud, which matters for some sovereignty definitions but not most commercial risk assessments.

The other common pattern is routing by sensitivity: a single internal gateway sends restricted data to a self-hosted or onshore model and everything else to the most capable API model. That gives most users frontier quality while keeping the sensitive lane under your control.

A decision guide

If this is true	Lean towards
You need the best available reasoning or coding quality	API model
Volume is low or unpredictable	API model
An Australian-region option exists for the model you need	API model via that platform
Prompts must never reach a model vendor	Self-hosted open-weight
You must run offline or in an isolated network	Self-hosted open-weight
Steady, very high volume of simpler tasks	Open-weight, self-hosted or managed
You need a frozen model version for a validated process	Self-hosted open-weight
You want onshore processing without running GPUs	Open-weight on a managed platform in Sydney
Mixed data sensitivity	Both, behind one routing gateway

If fine-tuning is the reason you’re considering open-weight models, note that some API providers also offer fine-tuning for selected models, though you don’t get the weights and the options are narrower. Read RAG vs fine-tuning first: retrieval often solves the problem more cheaply than either.

How All Webbed Labs approaches model choice

We test candidate models, API and open-weight, against your own examples before recommending one, and we write down the model, platform, region and date in the architecture decision. Where sovereignty requires it, we deploy open-weight models in your own cloud account in an Australian region, with logging you control. Where it doesn’t, we’ll usually recommend an API model and say why. See our private LLM deployment service and LLM integration work.