Open-source AI in the enterprise: licenses and real savings

On November 3, 2026, Duolingo explained that the cost of an AI video call had dropped from about $0.30 to "well under $0.01", "mainly by moving to open-source models or open-weight models." A few days earlier, Pinterest reported a cost per transaction below 8% of comparable closed models, and Revolut said it had cut its AI coding bill to about a tenth of what Claude Code and Codex cost it. In all three cases, an open-source AI model brought the bill down.

AI budgets, meanwhile, are running over. According to Menlo Ventures, enterprises spent 4.7 times more on AI than in 2025, when they had planned for 1.9 times. At Tinder, according to its CTO, annualized AI spending went from $1 million to $10 million between January and July. And every seat of Microsoft 365 Copilot Business costs $21 a month. An open model can be downloaded, hosted anywhere and paid for at the price of the infrastructure. But what you are allowed to do with it is set by its license, not by the words "open source."

For a CIO, moving to an open model therefore raises two questions: how much will we save, and does the license give us the protection we need, legally and in terms of continuity? Our answer: the savings are real, but they rest on a text the publisher can tighten in the next release. Several publishers already do so for their flagship models, so a model license has to be managed like a contract.

We write "open-source AI" because that is the term everyone uses, but it is inaccurate: these models publish their weights, not their training data or training code. "Open weight" is the right term, as we explain in our article on open source vs open weights.

How much do companies save with open-source AI?

The figures published in 2026 are striking, but they are claims: each company speaks about itself, on its own scope, without detailing its calculation.

Pinterest is the strongest case: its CEO has said the same thing for a year, from an "orders of magnitude reduction in cost" in November 2025 to the 8% of October 2026. Revolut is the most debatable: the comparison sets an API bill against an inference cost on 56 B200 GPUs that the bank runs itself, and the figure says nothing about the price of the hardware or the team that operates it. AT&T, finally, is talking about some tasks, not its whole bill.

At Vercel, whose gateway relays calls for many customers, open models accounted for 71.5% of tokens but 22.6% of spending in September 2026. Menlo Ventures measures 19% of enterprise tokens for 4% of spending. In other words, the open model takes the volume, the closed model keeps the hard cases. Duolingo says so in plain words: "There are some small cases where a frontier model's still needed." The savings come from routing more than from a wholesale replacement.

Does an open model really cost less than a closed API?

For a model under Apache 2.0, no publisher takes a cut: competing hosts serve it and charge for compute. In October 2026, the same Qwen 3.8 27B cost €0.40 per million input tokens and €2.70 per million output tokens at OVHcloud, and €0.60 / €3.30 at Scaleway, both in France.

At the same level, the gap is not what most people assume. In the November 10 Artificial Analysis leaderboard (version 5.0.1 of its index), Qwen 3.8 27B scores 20 points. The closed models at that level are the small models from the big vendors, and they are priced lower:

Per token and at the same level, a small closed model now costs more than five times less for output than Qwen 3.8 27B served in France. The real gap is even wider, because Qwen 3.8 27B thinks at length: Artificial Analysis saw it produce 128 million tokens to complete its index, against a median of 77.7 million.

These prices are recent. OpenAI launched GPT-5.6 Luna on July 9, 2026 at $1 per million input tokens and $6 per million output tokens, then brought it down to $0.20 / $1.20 three weeks later. On September 22, GPT-6 Luna arrived at $0.10 / $0.50: in less than three months, the output price of OpenAI's small model was divided by twelve. Anthropic followed on October 7 with Claude Haiku 5.5, priced 90% lower than Haiku 4.5 for requests under 100,000 tokens, and about 75% lower on average according to the company, because its new tokenizer splits the same text into more tokens. OpenAI attributes its cut to efficiency gains. We also see it as a response to small open models, which were taking the volume of simple tasks at the cost of hosting. Competition from open models first brought down the price of closed models.

The savings reported by Duolingo or Pinterest come from elsewhere. They compare an open model with frontier models, which cost far more, on tasks that did not need them; or, as at Revolut, they rely on GPUs running around the clock. For a company that pays per token, savings come first from using the right level of model for each task, open or closed.

So the case for an open model rests less on list price than on control. An open model can be hosted wherever you want: with a European provider, in your own data center, or on air-gapped servers. You choose who sees the data and which law applies, the core question of AI sovereignty. You can fine-tune it on your business vocabulary, it behaves the same until you change it, and you can replace it without renegotiating a contract. The ways to run it, from buying GPUs to a host's API, are compared in our article on enterprise AI inference, and the full cost of running it on your own servers in our article on on-premise LLMs.

What does each open-source AI license really allow?

A model license is the contract that comes with the published weights: it says who may use them, for what, and on what terms. You read it in the repository's LICENSE file, not in the page summary: at Mistral, Timestral's revenue cap sits in that file, not in the model description. Open model licenses fall into four families, from the most permissive to the most restrictive.

Apache 2.0 and MIT: almost everything is allowed

The Apache 2.0 and MIT licenses come from free software. They allow commercial use, modification, fine-tuning and redistribution, provided you keep the license notices. Apache 2.0 adds a patent license and calls itself "perpetual" and "irrevocable": a version downloaded under it stays under it.

This is the regime of Qwen 3.8 27B, Google's Gemma 4, OpenAI's gpt-oss, Mistral Small 4 and Mistral Large 3. DeepSeek releases its recent models under MIT. For a company, this is the simple case: no threshold to watch and no contract to negotiate.

One variant adds a display obligation. The Kimi K2 license, a modified MIT, requires you to display "Kimi K2" in the product above 100 million monthly active users or $20 million in monthly revenue. The constraint concerns branding; use remains free.

Llama: a community license that excludes Europe

Meta releases Llama under its own license. The Llama 4 license requires you to display "Built with Llama," to start the name of any distributed derivative model with "Llama," and to request a license from Meta above 700 million monthly active users. It is governed by California law, and Meta may terminate it for breach, in which case you must delete the models.

For a European reader, the decisive point sits in the acceptable use policy the license refers to. For the multimodal models in Llama 4, the rights are not granted to a company "with a principal place of business in the European Union". And all Llama 4 models are multimodal. A company with its principal place of business in the EU cannot deploy Llama 4 itself; it can only use a third-party product that embeds it.

Mistral Large 4, Qwen, Meta Muse: a contract to resell the model as a service

A more recent family leaves internal use free but reserves resale of the model as an inference service, known as "Model as a Service." Giving a third party direct access to the model, through an API for example, then requires a contract with the publisher.

• The Mistral Large 4 license, released on October 29, 2026, requires a separate commercial license from any company whose group runs a Model as a Service business. Internal use is exempt, as are products that embed the model in their own features.
• The Qwen Community License of Qwen 3.8 Flash-Next, which Alibaba presents as a preview of the Qwen 4 architecture, goes further: the separate license also applies to an "AI Work Assistant," a product designed mainly for AI-assisted coding or office work.
• The Muse Model License of Meta's large open model, released on October 15, 2026, requires a separate agreement for an inference service above $20 million in revenue over twelve months.

For a company deploying an assistant for its employees, these licenses change nothing. For a software vendor, a host or an IT services firm that resells access to models, they change the business model.

Mistral Medium 3.5: a revenue cap

The most restrictive family withdraws all rights above a certain company size. The Mistral Medium 3.5 license, a "modified MIT," does not apply if your group's consolidated monthly revenue exceeds $20 million, roughly $240 million a year. The restriction also covers derivative models. Above that, you must ask Mistral for a commercial license, which it grants "at its sole discretion." Devstral 2 and Timestral carry the same clause. A small business can use it freely; a large group cannot, not even internally.

The operational table

Here is our reading of the texts, which is not a substitute for legal advice:

That leaves distillation: training a small model on the outputs of a large one is allowed by open licenses, within their naming or threshold conditions. Closed APIs, by contrast, forbid it: Anthropic's terms bar using its services "to train competing AI models".

Will open model licenses tighten?

Our view: yes, among the publishers who can afford it, starting at the top end. The signals of 2026 all point to the same pattern: the small model stays under Apache 2.0 to feed the ecosystem, the flagship moves to an in-house license.

• Mistral wrote in January 2025 that it was renewing its commitment to Apache 2.0 for its general-purpose models. Since then, Devstral 2, Medium 3.5 and Timestral have shipped with a revenue cap, and Mistral Large 4 under an in-house license, while Large 3 was Apache 2.0.
• Alibaba keeps Qwen 3.8 27B under Apache 2.0, but released its large model under a "Qwen3.8-Max License", and the Qwen 4 preview under the Qwen Community License. We had flagged this risk for Qwen 4 in our article on Qwen 3.8 27B.
• Meta announced as early as July 2025 that it would be "careful about what we choose to open source". Its Muse Spark model came out in April 2026 through an API for selected partners, its open version in October under an in-house license.

The texts already provide the means to tighten the rules. The Llama use policy points to "the most recent copy of this policy", a document Meta can change. Falcon's "may be updated from time to time". NVIDIA's open model license states that "NVIDIA may update this Agreement … at any time", leaving users to comply or stop.

Software has been through this before. MongoDB left its open-source license in 2018 because it was "too easy for cloud vendors … to capture all of the value", Elastic left Apache 2.0 in 2021, HashiCorp changed Terraform's license in 2023. Each time, the target was resale as a service by third parties, exactly what the new model licenses aim at. The last open version, though, remained usable: OpenTofu was born from the last open version of Terraform.

The counterpoints are serious. We found no case of a license withdrawn from an already released version: changes apply to later versions. The opposite movement exists too: Google moved to Apache 2.0 with Gemma 4, DeepSeek and Z.ai moved to MIT with DeepSeek V3 and GLM-4.5, and NVIDIA will adopt the OpenMDW open license for its next models. The US government has encouraged open models since July 2025, and China's draft export controls exclude what is already published as open source. As long as competing labs release near state-of-the-art models under Apache 2.0, a publisher that closes too much loses its audience. The tightening therefore targets flagship models and resale as a service; mid-size models, the ones most companies deploy, should stay open longer.

What CIOs should take away

Unless you breach the license, the version you use today will remain usable. It is the next one, the one you will need in a year, that may arrive on different terms. The law firm Cooley advises planning for changes in licensing terms and notes that a self-hosted model may come without any indemnification from the provider. In practice:

1. Read the license file and the use policy, not the model card, with your legal team. Check thresholds at group level (consolidated revenue, affiliates) and against your business model: an IT services firm or a software vendor does not face the same constraints as a bank equipping its staff.
2. Archive the weights, the license and the use policy with their dates, along with the exact identifier of the revision you downloaded. If the repository changes, you keep proof of what you were granted.
3. Freeze the version in production. Any new version goes through a license review before deployment.
4. Keep an Apache 2.0 or MIT alternative, already tested on your own questions. If a publisher tightens its terms, you switch.
5. Make the model replaceable: an OpenAI-compatible API, a gateway between the application and the models, a reference set of questions. Switching models should take a day, not a project.
6. Choose the host as carefully as the license. An open model served by a host whose parent company is outside the European Economic Area remains exposed to that country's law. For hosting in Europe, OVHcloud; for European inference, Scaleway.

One last reminder: the cheapest model is the one you never call. AI is only worth using where the task is not deterministic; elsewhere, an algorithm is more reliable and more frugal, as we explain in our article on automation without unnecessary AI.

At ATG, we use open models a lot. Qwen 3.8 27B writes the answers to our users' messages and is the model behind our FinanceBench benchmark, and we compute our embeddings with Qwen 3 Embedding 8B. A company that does not want to deal with licenses can also rely on us: our platform selects and runs the models, with flat-rate billing per message or per user.

Frequently asked questions about open-source AI licenses

Can a European company use Llama 4?

Not directly. The Llama 4 acceptable use policy, which is part of its license, does not grant rights over its multimodal models to companies whose principal place of business is in the European Union, and all Llama 4 models are multimodal. The company can, however, use a third-party product that embeds Llama 4: the restriction does not apply to end users. Models under Apache 2.0, such as Qwen 3.8 27B, Gemma 4 or Mistral Small 4, have no such limit.

Can a publisher revoke the Apache 2.0 license of a model already released?

No. The Apache 2.0 text grants a perpetual and irrevocable license: a version downloaded under it remains usable, modifiable and redistributable, even if the publisher releases the next one on other terms. The only exception is the patent license, which ends if you yourself sue the publisher for patent infringement. Hence the value of archiving the weights and the license file of the deployed version.

What is a "Model as a Service" clause in a model license?

It is a clause that requires a contract with the publisher to give third parties direct access to the model, for example by reselling it through an API as an inference service. It appears in the licenses of Mistral Large 4, the Qwen Community License and Meta's large open model. Internal use stays free, and products that embed the model in their own features are generally outside the clause.

Does an open-source model cost less than the OpenAI API?

Not necessarily. On November 10, 2026, Qwen 3.8 27B hosted in France by OVHcloud cost €0.40 per million input tokens and €2.70 per million output tokens, while OpenAI's GPT-6 Luna, at a comparable level on the Artificial Analysis index, cost $0.10 and $0.50. The savings companies report come mainly from replacing frontier models on simple tasks and from self-hosting at scale. The open model keeps the advantage of control: choice of host, a frozen version, and reversibility.

---

The companies saving the most often keep a closed model for the hard cases; their strength is being able to switch models painlessly. That same skill protects against a license being tightened. Before choosing a model for 2027, we would measure just one thing: how long it takes you to change it.