Open source vs closed source AI: how enterprises choose

On IBM's third-quarter earnings call, chief executive Arvind Krishna said open-weight AI models, the kind any company can download and run itself, now run at anywhere from 25% to 10% of the token cost of the leading closed models, while being 90% as effective; IBM sells the hardware and software to run them in-house. That turns open source vs closed source AI from a developer debate into a board decision about operating cost, control and risk.

The short answer for a large organisation is to use both, on purpose. Put steady, high-volume work on open-weight models that run where you control the data. Keep the hardest reasoning on closed models that you rent from a vendor. Build every system so you can swap the model without a rebuild.

Three things decide the split: how much volume a task has, how sensitive its data is, and whether you have people who can run models well. This guide explains how each option works, what it costs, where the risks sit, and how to test the choice before you commit budget.

What is the difference between open source and closed source AI models?

The difference is who holds the model, and so who controls where it runs. An AI model is a large file of numbers, called weights, that the model learned in training. Whoever holds the weights can run the model.

A closed model keeps its weights private. Your systems send text to the vendor's servers through an API, a connection one program uses to call another. The vendor runs the model and charges you per token. A token is a small piece of text, roughly one word. MIT Sloan names OpenAI, Anthropic and Google as the main closed providers.

An open model publishes its weights. You download them and run the model on servers you choose: your own data centre, your cloud account, or a hosting partner. The same MIT Sloan article describes open models as ones that make public one or more details, such as the weights, source code, training data or design. Models from Meta, Mistral and DeepSeek are the common examples.

Open weight vs open source: why the label matters

Most models sold as open are open weight, not open source in the full sense. Open source software usually means you get the code, may change it, and may use it for any purpose. Many open-weight models publish the weights but not the training data, and they come with a licence that sets terms.

Meta's Llama licence is the clearest example. A company whose products had more than 700 million monthly active users in the month before release must request a licence from Meta, which Meta may grant or refuse. Few firms reach that size. The lesson still holds for everyone: read a model's licence before you build on it, as you would read any software contract.

Regulators treat the difference as real. Under Article 53(2) of the EU AI Act, some documentation duties do not apply to a model released under a free and open-source licence with its weights, design and usage information made public. The exemption does not apply to general-purpose models with systemic risk, the EU's term for the most advanced models, which could cause harm at large scale. That exemption helps the model's maker. It does not remove your duties as the company that uses the model.

What changed: open models caught up, and clouds now sell both

The quality gap that once justified paying more for closed models has narrowed to weeks. Krishna, whose company sells on-premises AI infrastructure, said that in 2024 a broad benchmark, a standard test set used to score models, might score the best closed model at 70 and an open model at 20 or 30. By late 2025, he said, it was 70 against 50.

Independent research agrees. A study by Frank Nagle of the MIT Initiative on the Digital Economy and Daniel Yue of Georgia Tech found that open models averaged 89.6% of closed-model performance and usually closed the gap within 13 weeks of a closed model's release. A year earlier, the same catch-up took 27 weeks.

Price moved the other way. In the same study, closed models cost US$1.86 per million tokens on average, against 23 cents for open models. Yet closed models still carried about 80% of model usage. The researchers point to the real cost of switching, and to a mistaken belief that open models perform worse.

Buyers also no longer need two suppliers to get both kinds. On 21 October 2026, Google Cloud began offering OpenAI's models, in private preview for selected enterprise customers, alongside more than 200 first-party, third-party and open models. Google said customers can swap models by cost and performance without being locked into one design.

Large firms already mix. In Andreessen Horowitz's 2025 survey of enterprise technology chiefs, 37% of respondents used five or more models, up from 29% a year earlier. The same survey found open model use was higher at larger enterprises, driven by data security, compliance and the wish to tune models to their own work.

How each option works, step by step

Both options answer the same request. The difference is whose machines do the work, and who carries the running cost.

A closed model, step by step:

1. Your system sends the prompt, and any company data the task needs, to the vendor's service.
2. The vendor's model reads the text and writes an answer on the vendor's chips.
3. You pay per token, in and out. You own no hardware.
4. The vendor decides when the model changes. Anthropic gives at least 60 days' notice before it retires a public model, and after that, requests to it fail. Google lists the Gemini models it keeps available for at least 12 months after release.

An open-weight model, step by step:

1. Your team picks a model and checks its licence.
2. You download the weights and load them onto GPUs, the specialised chips that run AI, in your data centre or cloud account.
3. Inference software, the program that serves answers, takes requests from your systems and returns results. The data stays in the environment you chose.
4. You pay for the GPUs by the hour, or buy them, whether they are busy or idle.
5. You decide when to upgrade. You also test, patch and secure the whole stack yourself.

There is a middle route. Cloud providers and specialist hosts run open-weight models for you and charge per token, much as a closed vendor does. You keep the option to move the same model in-house later. But your data goes to the host, so the data question returns.

What the choice means for cost, people, risk and scale

Closed models are a variable cost. Self-hosted open models are mostly a fixed cost. So volume decides which is cheaper.

The MIT study estimated that moving demand to the best-value open models could cut average spending by more than 70% and save the AI economy about US$25 billion a year. That figure leaves out switching costs, which the authors say vary widely. Treat it as the size of the prize, not as your saving.

Fixed capacity only pays when it is busy. DataRobot, which sells AI software, argues that at sustained inference volumes, owned or dedicated GPUs often cost less than on-demand cloud capacity. The key word is sustained. A GPU that sits idle at night costs the same as one that works.

Worked example, continued. Now the group puts invoices, contracts, customer emails and call notes through AI, and volume rises to 200 billion tokens a year. The closed bill becomes US$600,000. Eight busy GPUs at US$25,000 each a year cost US$200,000, plus US$80,000 for the people who run them. Self-hosting now costs US$280,000 and saves about US$320,000 against the closed model. Same models, opposite answer. But the hosted open model costs US$100,000, still less. Self-hosting beats that route only when the data must stay in your environment, or when volume is high enough to keep every GPU fully busy.

People are the cost that rarely appears in the business case. Krishna thinks most businesses can build useful AI infrastructure in two or three racks of hardware. He then named the real constraint: finding engineers who can tune the chips, the network and the caching. Closed models leave that work with the vendor. Open models bring it back to you or to a partner.

Risk moves, it does not disappear. A closed vendor can change the price, the terms or the model, and you re-test on its timetable. DataRobot calls this a concentration risk for any enterprise built on a single provider's platform. With an open model you keep a copy of the weights, so no one can retire it under you. But every patch and every control is now yours.

At scale, the cheapest pattern is a mix with a gateway in front. A gateway is a thin layer of software between your systems and the models. It sends each task to the model that is good enough at the lowest cost, and it logs the cost of each call. Changing model becomes a setting, not a rebuild. Our guide to AI cost management covers the other levers on the bill.

What people get wrong about open source AI

Four beliefs push firms to the wrong choice. Each fails against the evidence.

Myth 1: open source AI is free

The weights are free to download. Running them is not. You pay for hardware, power, engineers and security, and you take on the licence terms. A free model on idle GPUs can cost more than a paid model used well, as the worked example shows.

Myth 2: open models put company data at risk

Where the model runs decides where the data goes, not the licence. Nagle told MIT Sloan that open models can be built and run within your own infrastructure, so the data never leaves your servers. The reverse also holds. An open model used through a third-party host sends data to that host. Check the contract and the hosting region in both cases.

Myth 3: closed models are always better

At release, open models are close behind, and the gap now closes within about a quarter. For routine work such as reading a contract, answering a customer call or drafting an email, Krishna argued the cheaper model does the job as well. Some tasks still need the best reasoning on the market. Find out which by testing, not by assuming.

Myth 4: closed models are safe and open models are not

Safety differs by model, not by label. What open release changes is who carries the safety work. The European Commission notes that once a model is open-sourced, risk mitigations are more easily circumvented or removed. Whoever holds the weights can retrain away the built-in refusals.

Model files themselves can carry hostile code. Protect AI, which scans models on Hugging Face, the main public site for sharing AI models, had by April 2025 scanned 4.47 million model versions and flagged 352,000 unsafe or suspicious issues across 51,700 models.

Built-in guardrails, the filters that block harmful requests and answers, also vary. In January 2025, Cisco researchers jailbroke DeepSeek R1, an open-weight model, with a 100% attack success rate. A jailbreak is a prompt written to make a model ignore its safety rules. Other leading models resisted at least part of the attack, but labels did not predict the result: the same test broke OpenAI's closed GPT-4o on most prompts, and Meta's open-weight Llama 3.1 405B on almost all of them. If you run an open model, you add the guardrails.

The three options compared

The right column depends on the task, not on the company. One group can sensibly run all three at once.

How to choose: a checklist for leadership teams

Decide per workload, with six questions. Ask them for each AI system in production or in the budget.

1. Volume. Is the work steady every day, or small and uneven? Steady volume favours self-hosting. Uneven volume favours paying per token.
2. Difficulty. Does the task need the best reasoning on the market, or is it routine reading, sorting and drafting? Test before you decide.
3. Data. May this data leave your environment under a contract, or must it stay in one country or one building?
4. Exit. If the model is retired or repriced tomorrow, what breaks? If the answer is a rebuild, you need a gateway, or an open model you hold.
5. Skills. Who will run the GPUs, patch the stack and watch quality at night? If no one, use a hosted option or a partner.
6. Regulation. Will you fine-tune the model, meaning train it further on your own examples? The European Commission says firms that fine-tune or otherwise modify a general-purpose model may become providers of a new model, with duties limited to the change they made. Its indicative test is whether the amount of computing used for the change is more than a third of the computing used to train the original model. Light tuning on your own data uses far less, so it usually does not make you a provider. The open-source exemption in Article 53(2) does not cover models with systemic risk.

Then put four questions to every AI vendor, closed or open:

• If you retire or reprice this model, how much notice do we get, and what must we re-test?
• Does our data train your models, and where is it processed and stored?
• Can we export our prompts, test sets and logs in a form another model can use?
• A test DataRobot suggests: if we needed to run the platform completely cut off from your network, could we, and what would we lose?

How to run a pilot that compares open and closed models

Test one real workload on both, with the same data and the same scorecard, before you buy any hardware.

1. Pick one high-volume task with a clear right answer, such as coding invoices or pulling clauses from contracts.
2. Build a test set from a few hundred past cases where you already know the correct result.
3. Run your current closed model and two open-weight candidates on the same set, through the same gateway.
4. Score each one on accuracy, errors that would reach a customer, response time, and cost per completed task. For the open options, include hardware and staff time.
5. Set the pass mark before you see results. For example: the open model must match the closed one on accuracy and cut cost per task by a margin your finance team agrees.
6. Run the winner in shadow mode for two to four weeks. It does the work, people check it, and nothing goes out unreviewed.
7. Decide, and keep the test set so you can rerun it when a new model ships.

Rerun the test every quarter. With the gap closing in weeks, a choice made in January can be wrong by June.

Open source vs closed source AI risks, and the control for each

Each option carries its own risks. Each has a plain control.

• Hostile model files. Download only from the publisher's official account, scan every file, and prefer formats that hold only weights, not runnable code.
• Missing or removed guardrails. Put your own input and output filters around every model. Test it with harmful prompts before launch.
• Licence breach. Have legal review each model's licence. Keep a register of which model runs where, under which terms.
• Vendor retirement or repricing. Keep a test set and a gateway, and write an exit plan with a date for every production model.
• Data leaving your control. Match each workload's data class, meaning how sensitive that type of data is under your policy, to where its model may run. Self-host or use a regional host for the most sensitive data.
• Unclear origin. Know who built a model and where. Some groups have legal or reputational reasons to avoid models from some jurisdictions.
• No owner. Name one person accountable for each model in production: its cost, its quality and its updates.

Where Thriiv fits

Thriiv builds and runs custom AI systems, workflow automation and secure data pipelines for international enterprises. We start where the cost sits: we measure the work and put a number on the saving before budget is committed.

We are private by default: every system runs in an environment we host and manage. Our security page explains how that works. If you are weighing open against closed models for a costly workflow, tell us where the cost sits. We come back with what we would build, and what it saves.

What to do this quarter

Start with a list and one test, not a platform decision.

• List every AI workload, the model it uses and its monthly bill.
• Mark the three with the highest steady volume. They are your open-weight candidates.
• Put a gateway in front of every new AI system, so the model can be swapped.
• Build a test set for each important workload, from real past cases.
• Ask each AI vendor the notice, data and export questions above, and file the answers.
• Run one side-by-side pilot, and report cost per completed task, not cost per token, to the board.