A self-hosted chat front end on raw API keys is cheaper than $20 seats for most of your staff, and more expensive for the few who live in it. On a typical usage mix, 500 employees on LibreChat or Open WebUI cost about $1,860 a month in Claude Sonnet 5.5 tokens against $10,000 for seats. That $8,140 monthly gap is real, but it has to pay for the SSO, provisioning, retention and patching a seat already includes, and it disappears once somewhere between one user in seven and one in five runs a frontier model all day.
Our recommendation: if you have a platform team and a gateway, self-host LibreChat behind per-user budgets. If you do not, buy seats, and look hard at Claude Enterprise, which already bills usage at API rates. Prices below were checked on each vendor's live page on October 22, 2026.
| Option | Price (checked Oct 22, 2026) | Licence | SCIM | Who should NOT pick it |
|---|---|---|---|---|
| ChatGPT Business | $20/user/mo annual, $25 monthly; Premium $100/$125 | Proprietary SaaS | No (Enterprise only) | Anyone who needs automated deprovisioning |
| ChatGPT Enterprise | Contact sales | Proprietary SaaS | Yes | Buyers who need a published price to budget |
| Claude Team | $20 annual, $25 monthly; Premium $100/$125 | Proprietary SaaS | No | Regulated teams needing audit logs or custom retention |
| Claude Enterprise | $20/seat annual + usage at API rates, 20-seat minimum | Proprietary SaaS | Yes | Teams that want a fixed, flat monthly bill |
| Gemini in Workspace | Bundled: Business Standard $14, Plus $22 | Proprietary SaaS | Via Workspace | Microsoft 365 shops |
| LibreChat (self-hosted) | $0 licence + tokens + ops | MIT | Not documented | Teams with no one to own patching |
| Open WebUI (self-hosted) | $0 licence + tokens + ops; enterprise licence by quote | Custom (not OSI) since v0.6.6 | Yes | Anyone planning to rebrand it above 50 users |
| LobeChat (self-hosted) | $0 licence + tokens + ops | LobeHub Community License | Not documented | Anyone planning to modify and redistribute the code |
Which Option Wins at 500 Users?
For a 500-person company with no platform team, Claude Enterprise or ChatGPT Business wins, and the self-hosted route only wins once you already run a gateway.
The comparison uses one defined workload so the vendors' preferred metrics do not muddy it. Assume 500 users, priced at Anthropic's list rates for Claude Sonnet 5.5 ($2 per million input tokens, $10 per million output) and Opus 5.5 ($4 and $20), uncached, global routing. Each user falls into one of three monthly profiles:
- Light (70% of staff): 200,000 input and 20,000 output tokens.
- Median (25%): 1.5 million input and 100,000 output tokens.
- Heavy (5%): 20 million input and 600,000 output tokens, the person who pastes contracts and keeps one thread open for a week.
No vendor publishes a per-user distribution for chat usage, so these profiles are our assumption. Replace them with two weeks of pilot logs before you sign anything. The arithmetic that follows from them:
| Profile | Sonnet 5.5 per user/month | Opus 5.5 per user/month | vs $20 seat |
|---|---|---|---|
| Light | $0.60 | $1.20 | Far cheaper |
| Median | $4.00 | $8.00 | Cheaper |
| Heavy | $46.00 | $92.00 | 2.3x to 4.6x the seat |
| Fleet of 500 (70/25/5) | $1,860 total ($3.72 avg) | $3,720 total ($7.44 avg) | Seats: $10,000 |
The 25 heavy users are about 62% of the Sonnet bill. That is the number the "self-hosting is far cheaper" pitch leaves out. A flat seat absorbs the heavy user up to the vendor's usage limit, while an API key passes every token straight to you.
Run the break-even the other way. If your users split between the median and heavy profiles, API spend matches a $20 seat once about 38% are heavy on Sonnet, and once about 14% are heavy on Opus. If you default everyone to the frontier model because it is "the good one", one heavy user in seven wipes out the saving before you have paid an engineer.
What Does a $20 Seat Include That an API Key Does Not?
A seat includes identity, admin and compliance plumbing that you would otherwise build and maintain, and the plan tier decides how much of it you get.
On OpenAI's side, ChatGPT Business costs $20 per user per month billed annually or $25 monthly, with a two-seat minimum, and SSO over SAML or OIDC is included at no extra cost. SCIM is not. OpenAI's SCIM guide requires an eligible Enterprise or Edu workspace. Without SCIM, a leaver keeps their seat until someone removes it by hand. OpenAI also added a Premium seat for Business, listed on its pricing page at $100 annual or $125 monthly for heavier usage. ChatGPT Enterprise has no public price.
Anthropic's pricing page splits the same way. Claude Team runs $20 a seat annual or $25 monthly (2 to 150 seats) with SSO and usage analytics, but without SCIM, audit logs, custom retention or the Compliance API. Claude Enterprise adds all four and is priced at $20 a seat plus usage at API rates, with admin spend limits per user and per organization.
That last detail changes the self-hosting maths, because Claude Enterprise already charges you the API token bill. Against it, self-hosting saves only the $20 seat, $10,000 a month at 500 users, and you give up SCIM, audit logs, retention controls and a vendor-run Compliance API to get it. If your reason for self-hosting is "pay for tokens, not seats", Anthropic now sells that, with the governance attached. Our guide to Claude Enterprise spend controls covers what to switch on first.
Google takes the third route: Gemini is not a separate line item. Per Google Workspace pricing, Business Standard at $14 a user per month and Business Plus at $22 include Gemini in Gmail and Docs, and Business Starter at $7 includes Gemini in Gmail and the Gemini app but not Gemini in Docs. If you already pay for Workspace Standard or Plus, the marginal cost of company-wide chat is close to zero, and no self-hosted stack beats zero. Who should not pick it: a Microsoft 365 company, which would be buying a second productivity suite to get a chat window.
Which Self-Hosted Front End Should You Run?
LibreChat, because its licence is MIT and its identity options cover what an enterprise directory needs. Open WebUI is the better admin product with the more awkward licence. LobeChat is the loser for this job.
LibreChat: The Default If You Self-Host
The LibreChat repository carries a plain MIT licence, so you can modify, rebrand and embed it without asking anyone. ClickHouse acquired LibreChat on November 4, 2025, bringing over founder Danny Avila and his team, and said existing deployments "continue unchanged with ongoing investment in the open-source project." That announcement does not restate the licence, so the MIT file is your only guarantee. Watch for a change of terms the way you would on any acquired dependency.
On identity, LibreChat's authentication docs list OpenID Connect, SAML, LDAP and passkeys. They do not document SCIM, and user creation and deletion are command-line scripts (npm run create-user, npm run delete-user), not an admin console. Plan to deprovision through your IdP's session control and a scheduled job.
Its balance feature is the budget tool you need. It is off by default. When enabled, it gives each user a token credit balance (20,000 tokens on registration by default) with optional auto-refill on an interval. Turn it on before launch, not after the first invoice.
Who should not pick LibreChat: a company with no engineer who will own upgrades and the deprovisioning job.
Open WebUI: Better Admin, Read the Licence First
Open WebUI's feature list includes SSO over OIDC and LDAP, SCIM 2.0 provisioning, RBAC with groups, and usage dashboards with token and cost tracking. On features alone it is the stronger enterprise product of the two.
The catch is the licence. Since v0.6.6, released April 19, 2025, you may not alter or remove "Open WebUI" branding unless you had 50 or fewer users in a 30-day period, hold an enterprise licence that permits it, or fall under narrow contributor or research exceptions. Open WebUI says itself that the licence is not OSI-approved. Pricing for the enterprise licence is by quote.
For an internal tool, keeping the Open WebUI logo is usually fine. The licence becomes a problem when the business wants the chat to carry your company's name, or when legal applies an "OSI-approved licences only" policy.
Do not fork v0.6.5 to escape the clause, even though code up to that version stays BSD-3-Clause. A frozen fork misses every security fix since. CVE-2025-64496, a Direct Connections flaw that let a malicious model server run JavaScript in the user's browser and steal the session token, affected v0.6.34 and earlier and was fixed in v0.6.35. Researchers showed it could escalate to code execution on the server for users with tool permissions. A team that pins an old version to avoid a logo carries that exposure.
Who should not pick Open WebUI: anyone who needs to rebrand it for more than 50 users and will not pay for the enterprise licence.
LobeChat: The Loser for This Job
LobeChat ships under the LobeHub Community License, Apache 2.0 plus extra conditions. Commercial use is allowed "without modifying the source code"; developing and distributing a modified version needs a commercial licence from LobeHub LLC, and the producer reserves the right to change the terms. An enterprise deployment nearly always ends up patching an auth flow or a connector, which is exactly what the licence restricts. It is a polished consumer-grade interface, and the wrong base for an internal platform you expect to customise for years.
What Do You Gain by Self-Hosting?
You gain model choice, routing through your own gateway, logs you own, and residency you choose per request. Those are the legitimate reasons to self-host, and cost is a weaker one.
A self-hosted front end can point at any provider, so you can route summarisation to a small model and reserve a frontier model for the work that needs it. Every prompt and response lands in your own store under your own retention policy, rather than behind a vendor's export API.
Residency becomes a per-request setting. Anthropic's API, for example, keeps US-only inference behind inference_geo: "us", which applies a 1.1x multiplier on all token prices for Claude 4.6 and later models. Budget that in: the Opus fleet above goes from $3,720 to about $4,092 a month with US-only routing.
The gateway is where budgets actually get enforced. LiteLLM's proxy attaches a max_budget and budget_duration to virtual keys and teams, and requests fail once a key passes its cap; budgets need its database connected. Issue one LiteLLM key per user or per team behind the front end and the heavy tail is capped at the gateway, whichever UI you pick. Our LLM gateway comparison covers the alternatives, and the LiteLLM key-theft incident covers what to lock down once one gateway holds every provider key.
Where Does the Self-Hosted Saving Go?
Most of it goes to staff time. A small VM runs LibreChat or Open WebUI for hundreds of users; the expensive parts are upgrades, security patching, connector maintenance and the deprovisioning glue the seat vendor handles for you.
At the 500-user Sonnet mix above, the saving is $8,140 a month, or about $97,700 a year. That figure has to cover the share of an engineer's time spent tracking releases (Open WebUI shipped a fix for a session-stealing flaw in a point release), rebuilding connectors the SaaS vendors maintain for you, and answering "why can't I see my chat history" tickets. If that work fits in a fraction of one engineer already running your gateway, the saving is real. If it means hiring for it, the saving is gone.
The one public per-user spend figure from a vendor shows how fast the tail grows once usage turns agentic. Anthropic's Claude Code cost guide reports an average of about $13 per developer per active day, $150 to $250 per developer per month, with 90% of users under $30 per active day. That is coding, not chat. But a chat front end with tools, file uploads and MCP connectors pushes your heaviest users toward that pattern, at many times the $20 seat. Anthropic's own Enterprise consumption guide tells admins to define light, standard and power tiers by role before rollout and to start conservatively. Do the same with your gateway budgets.
How Should You Decide?
Decide on four things: whether you already run a gateway, whether you need SCIM and retention controls on day one, what share of your users will be heavy, and which productivity suite you already pay for.
- If you already pay for Workspace Business Standard or Plus, use the Gemini you have paid for.
- With no platform team and a day-one need for SCIM, audit logs and retention, buy Claude Enterprise, or ChatGPT Enterprise if your users are already on ChatGPT. Claude Enterprise's seat-plus-API-rate model gives you token pricing with the compliance features included.
- A team under 150 people with no SCIM requirement should buy ChatGPT Business or Claude Team at $20 a seat and build nothing.
- With a gateway already in production and a need for multi-model routing or your own logs, run LibreChat behind LiteLLM with per-user budgets, or Open WebUI if you will keep its branding and want built-in SCIM.
What changes the answer: a pilot showing more than about 14% heavy users on a frontier model makes seats cheaper, or about 20% if most of the rest are light users. A legal ruling that chat logs must stay in your own store makes self-hosting mandatory regardless of price. A move to Microsoft 365 Copilot puts you in a different seat model entirely; see what a Copilot seat really costs.
What to Do Before You Sign
This Week:
- Pull 30 days of usage from whatever chat tool staff use today and sort users by tokens. Count how many fall in the heavy profile above.
- Get a written Claude Enterprise quote and a ChatGPT Enterprise quote at your seat count, so the self-hosted plan is compared against real numbers, not list prices.
This Month:
- If you self-host, stand up LibreChat or Open WebUI behind LiteLLM with one budget per user, and enable LibreChat's balance feature or the gateway cap before you invite anyone.
- Write the deprovisioning job and test it on a leaver before launch.
- Set the default model to the mid-tier one and make the frontier model an opt-in.
Before Next Budget Cycle:
- Re-run the break-even with real logs: tokens per user, heavy share, and engineering hours spent on upgrades. If the saving no longer covers the hours, move to seats.
The Bottom Line
Raw tokens are cheaper than seats for the median employee, by a wide margin. A seat is flat-rate insurance against the heavy user, plus identity and compliance work you would otherwise staff. Claude Enterprise now sells token pricing with that work included, which narrows the case for self-hosting to model choice, owned logs and residency. If those are your reasons, run LibreChat behind a gateway with per-user caps set on the first day.
Continue Reading
- Claude Enterprise Spend Controls: 5 Things to Configure Now
- Best LLM Gateways for Cost Control: Self-Host First
- True Cost of a Copilot Seat: The Licence Is the Floor
- Best FinOps Tools for AI Spend: No Dashboard Fixes an Untagged Key
- Google Gemini Enterprise vs. ChatGPT Work: Pick Wisely
- AI Gateway vs API Management: No Managed Gateway Caps Your Spend
