For most mid-sized companies, ChatGPT Business or Enterprise is the cheaper and faster way to give everyone a capable AI assistant, and OpenAI states that it does not train on business data from those plans by default. A private LLM, a model running on infrastructure dedicated to you or on your own hardware, is worth it when data must not leave your control at all, when you need to fix the model version and behaviour yourself, or when heavy, steady use of one workflow makes a fixed monthly cost cheaper than paying per seat or per token. Plenty of companies should use both.
This comparison uses OpenAI's published terms as of September 2026 and our own published pricing. Both change, so check the linked sources before you decide.
What do ChatGPT Business and Enterprise actually promise?
It's worth being precise here, because "ChatGPT isn't private" is out of date for the business plans.
According to OpenAI's enterprise privacy commitments:
- Data from ChatGPT Business, Enterprise and the API Platform isn't used to train OpenAI's models by default, unless you explicitly opt in.
- You own your inputs and outputs, where allowed by law.
- Workspace admins control retention. Deleted conversations are removed from OpenAI's systems within 30 days, unless the law requires longer retention. For Business, OpenAI also reserves longer retention where it's reasonably necessary to protect its services or others from harm.
- ChatGPT Business, Enterprise and the API Platform have each completed a SOC 2 Type 2 audit.
OpenAI's business pricing page lists what each plan adds:
- Business: standard seats at $20 per user per month billed annually ($25 monthly) and premium seats at $100 ($125 monthly), for teams of 2 to 200 employees, with SAML SSO, MFA, centralised billing and connectors to tools such as Google Workspace, Slack and Microsoft 365.
- Enterprise: custom pricing, with SCIM, enterprise key management, role-based access controls, custom data retention policies, data residency in ten regions, and SLAs.
For a lot of companies, that covers it. If your concern is "we don't want our data training someone's model" and your contracts allow a reputable processor, the business plans may well be enough.
So what does a private LLM add?
What a private model changes is where the data goes and who controls the model, not the training policy.
- The data doesn't have to leave. With an on-premise install, prompts and documents never cross your network boundary. With a hosted private model, they go to infrastructure dedicated to you rather than to a shared multi-tenant service.
- You pick the model version, and it stays put. Hosted services retire and replace models on their own schedule. A private deployment runs the version you tested until you decide to change it.
- Cost is fixed, not per seat or per token. You pay for capacity, however many people use it and however much.
- No vendor-imposed rate limits or feature changes on the workflows you've built around it.
What you give up is the convenience of a finished product. A ChatGPT workspace comes with a polished app, mobile clients, file handling, connectors and a steady stream of new features. A private model gives you a model, and the value comes from what you connect it to.
How do they compare side by side?
| ChatGPT Business | ChatGPT Enterprise | Private LLM (hosted or on-premise) | |
|---|---|---|---|
| Pricing shape | Per seat: $20/user/mo annual, $25 monthly (standard) | Custom, per contract | Fixed: hosted from $890/mo (Nexos); on-premise quoted per install |
| Scales with | Headcount | Headcount and contract | Workload (hardware capacity) |
| Trains on your data by default? | No | No | No; nobody else is in the loop |
| Where data is processed | OpenAI's infrastructure | OpenAI's infrastructure, with data residency options | Dedicated infrastructure, or inside your network |
| Model choice | OpenAI's current models | OpenAI's current models | The open-weight model you choose and test |
| Model changes | On OpenAI's schedule | On OpenAI's schedule | On your schedule |
| Out-of-the-box app | Yes | Yes | No; it's integrated into your tools instead |
| Best for | Everyday AI for a whole team | Large rollouts with compliance requirements | Specific workflows on data that must stay in your control |
Sources: OpenAI business pricing, OpenAI enterprise privacy, Nexos pricing.
What does the cost comparison really look like?
Seat pricing is easy to work out. At the published Business rate, 100 standard seats billed annually come to $2,000 a month, and the cost rises with headcount.
Our hosted private model starts at $890 a month. That doesn't make the private model cheaper, though, because the two aren't the same product. $890 buys a dedicated model and its upkeep. It doesn't include a polished chat app for 100 people or connectors to every SaaS tool you use. Any integration work is a separate, scoped build.
The fair comparison is by workload:
- "Everyone should have an AI assistant." Seats win. A private model would mean rebuilding a product that already exists.
- "This one workflow processes thousands of sensitive documents a month." A private model can win, because the cost is fixed however much the workflow does, and the data never leaves.
- "We need both." Many mid-sized companies will end up here: a ChatGPT workspace for general use, and a private model behind the one or two workflows that handle restricted data.
Do you need a research team to run a private model?
No. Capable open-weight models now run on hardware a mid-sized company can own. OpenAI's own open-weight models are released under the Apache 2.0 licence, and OpenAI says gpt-oss-120b runs on a single 80 GB GPU while gpt-oss-20b can run with 16 GB of memory. Other providers publish open-weight models too, and which one to use should be decided by testing on your own examples, not by a leaderboard.
What you do need:
- Honest hardware sizing. That depends on model size, how many requests arrive at once, and how fast each answer has to come back.
- Integration with your documents and tools, with the same approval points as any other automation.
- Someone accountable for it. That means monitoring, updates and security patches, the same as any production system.
We go through each of those in Private AI without a research team.
How do you decide?
Go through these questions in order:
- Does a contract, regulation or internal policy forbid sending this data to a third-party processor? If yes, you need a private model for that data, whatever else you use.
- Is the need general (everyone drafting, summarising, researching) or one specific workflow? General needs point to seats. A specific high-volume workflow can point to private.
- Do you need to fix the model version for consistency, validation or audit reasons? That points to private.
- Who will own it? A ChatGPT workspace needs an admin and a usage policy. A private model needs an operator. If you have neither, budget for ongoing support.
If you get to the end and the answer is still "it depends", that's normal. The deciding details are usually specific: which documents, how many, and under which contract. A free audit sorts that out workflow by workflow. If private turns out to be right, our hosted private LLM and on-premise installation services are built so you can start hosted and move on-premise later without a rebuild.
Weighing a ChatGPT rollout against a private model? We'll size both against your actual workloads.
Discuss a project with Abrams & Luchanski →