Commercial AI APIs or Open Weights

The commercial-versus-open question comes up in almost every deployment project, usually halfway through a workshop and usually framed as a matter of principle. The answer is rarely a matter of principle. It depends on what the model has to do, how often, whose data it touches and who carries the consequences when the output is wrong.

Three operating models to compare

In practice there are three distinct operating models, even though the discussion tends to assume two. The first is a commercial model available only through the vendor’s API (application programming interface): the GPT, Claude and Gemini families. The second is an open-weights model running at an external hosting provider, where the weights are public but the infrastructure still belongs to somebody else. The third is an open-weights model running on hardware you own or in your own private cloud.

Dimension	Commercial API	Open weights, hosted	Open weights, self-run
Data confidentiality	Zero retention and an EU region in the contract	The infrastructure is still somebody else’s	Data never leaves the network
Cost at low volume	Lowest: pay for what you use	Medium	Highest: the hardware idles
Cost at high steady volume	Rises linearly	Depends on the contract	Lowest at full utilisation
Control over model versions	The vendor can retire a variant	Limited	Full
Team effort	Close to zero	Small	Monitoring, maintenance, on-call

It is also worth separating “open source” from “open weights”. Most open models publish weights and architecture, while the training corpus and training code stay closed. For business use that distinction matters mainly when auditing data provenance.

Data confidentiality

This is the most common reason open models enter the conversation, and also the argument made least precisely. Commercial vendors now offer zero-retention agreements, exclusion of customer data from training, and a choice of processing region inside the European Union, including through cloud platforms. For most corporate data that is sufficient protection and a much shorter path than building your own infrastructure.

A locally hosted model wins where the constraint is hard: data covered by professional secrecy, medical records, material from legal proceedings, information that security policy forbids from leaving your own network even to a trusted processor. In those cases the benchmark discussion becomes secondary, because security policy fixes the deciding variable before anyone looks at the scores.

Cost at scale

This is where we most often find an error in the arithmetic. Open models carry no licence fee, but they do carry running costs. Renting a single H100-class accelerator costs roughly two to three dollars an hour at specialist GPU (graphics processing unit) clouds and noticeably more at large public cloud providers. Running one continuously therefore costs well over a thousand dollars a month in hardware alone, before engineering time, monitoring and maintenance.

The hardware footprint can be smaller than intuition suggests. OpenAI states that gpt-oss-120b fits on a single 80 GB GPU and that gpt-oss-20b runs within 16 GB of memory, which puts it inside a well-specified workstation. Both are published under Apache 2.0.

The consequence is straightforward. At low and irregular volume the API wins almost every time, because you pay only for what you use. The break-even point appears at high, predictable volume on repetitive work: ticket classification, field extraction from documents, first-pass summarisation. That is when a dedicated instance runs near full utilisation and the arithmetic changes. Before anyone signs off on a migration plan, count actual request volume from the last three months.

Quality outside English

This is where commercial models held their largest lead for years, and the lead is still visible on work that needs stylistic range: drafting correspondence, handling register and tone, working with legal text. The gap has narrowed considerably. The Qwen3 family is published under Apache 2.0 and claims support for 119 languages and dialects. Mistral, a French company, publishes both its flagship Mistral Large 3 and the smaller Ministral 3 series in 14B, 8B and 3B variants under Apache 2.0, which adds a European option on the jurisdiction side as well.

National ecosystems are worth knowing too. In Poland, the Bielik model from the SpeakLeash foundation, developed with support from the Cyfronet AGH computing centre and released under Apache 2.0, is trained on Polish data and splits Polish text into billing units differently, so the same text costs less than it would on models trained mainly on English. Its 11B version arrived at the start of 2026. For comparisons there is SpeakLeash’s Open PL LLM Leaderboard on Hugging Face, built on Polish NLP (natural language processing) tasks. It measures classification and reading comprehension, so we treat it as a signal with limited reach; it says little about editorial quality. Our recommendation is a private evaluation set: 50 to 100 cases taken from the business, scored by the people who do that work daily.

Version control and continuity

This dimension is underrated and can decide the whole question. An API vendor can retire a model version, change its behaviour or tighten content filters, and a process that passed validation starts returning different results. An open-weights model downloaded once behaves the same way a year later, as long as you maintain the environment. For systems under validation, audit or compliance documentation, that reproducibility is valuable in its own right.

Licensing and liability

Read the licence of an open model before anyone builds a product on it, because the families are not equally open. The Llama 4 licence contains this clause: “With respect to any multimodal models included in Llama 4, the rights granted under Section 1(a) of the Llama 4 Community License Agreement are not being granted to you if you are an individual domiciled in, or a company with a principal place of business in, the European Union.” Since the entire Llama 4 family is natively multimodal, a company headquartered in the EU faces a legal barrier that has nothing to do with model quality. Apache 2.0, the licence behind Qwen3, Mistral Large 3 and gpt-oss, carries no such restriction.

The other side of that dimension is liability. Buying an API also buys a contract, a service level, a support channel and an entity answerable for the service. Running an open model means taking all of that in house: availability, security updates, output quality and the consequences when quality slips.

One regulatory note. The EU AI Act grants providers of general-purpose models released under free and open licences a limited exemption from some documentation obligations, conditional on weights and architecture being publicly available and the model not being monetised. The exemption does not extend to models with systemic risk, and the obligations covering copyright policy and the training-data summary remain in force. Deployer obligations apply regardless of whether the model is open.

What we usually recommend

In most organisations we work with, the answer settles into a mixed architecture. The commercial model handles work that needs the strongest language quality and the widest context. The open model, running locally, handles repetitive volume and data that cannot leave the network. Draw the line once, write it into policy, and review it quarterly, because both sides of this market move faster than an annual budget cycle.