Skip to content
senn-techsenn-tech
Back to blog
AI & Development
AI in production2026-11-16· By Franz Senn

LiteLLM Customer Management is only in a release candidate, in no stable version

LiteLLM Customer Management is contained only in the release candidate v1.111.0-rc.2 as of November 19, 2026. The latest stable version v1.110.1 of the same day has none of the new routes and no migration, measured on the source tree of the release tags on GitHub. If you need a budget per customer, the existing end-user tracking works on the stable line, and the docs continue to support it.

LiteLLM announced Customer Management on November 13, 2026: usage and budgets per customer, their teams and individual members, mapped onto a single gateway key. On November 16 the docs contained the sentence "A published release containing the complete integration has not been verified." That day I measured the source tree of three version states and our running gateway, and on November 19 the source tree again. The sentence has since disappeared from the docs. On November 16, main held 152 paths with customer in the name, the then-latest stable version v1.108.9 held six, and our lane does not know the documented routes.

LiteLLM Customer Management is a set of administration routes under /management/v1 in the LiteLLM gateway, plus two headers on inference: x-litellm-customer-org-alias and x-litellm-customer-member-id. From both, the gateway derives a customer hierarchy of organization, team and member, checks the budgets of these three levels before every request and writes usage into counters that you read back through usage endpoints. This customer hierarchy is separate from LiteLLM's internal organizations, teams and users.

At a glance

  • Announced on November 13, 2026; docs retrieved on November 16: the availability table of the docs points to main and to seven individual pull requests. As of November 19, the same docs carry "Since v1.111.0" instead.
  • Measured distribution on November 16: 152 customer paths on main, 6 each in release v1.106.4 and in the then-latest stable release v1.108.9. Re-measured on November 19: 7 in stable v1.110.1 (one more test, no migration), 154 in release candidate v1.111.0-rc.2, 184 on main.
  • Seven SQL migrations with customer in the name sat on main on November 16, none of them in the stable release. Now there are 12 on main and 8 in the release candidate. Moving up therefore brings database schema changes along.
  • Two of the linked pull requests are closed without being merged, yet they cover functions the docs still described.
  • Our gateway (v1.106.4) answers the documented routes with 404 and the control routes with 200.
LiteLLM wordmark
LiteLLM announced Customer Management on November 13, 2026 and published the docs for it. (Quelle: LiteLLM / Wikimedia Commons, public domain (PD-textlogo), trademark: LiteLLM)

What LiteLLM promises

The announcement targets companies that build agents on the gateway and pass the costs on to their own customers. Each customer level has a budget object with max_budget, soft_budget, rpm_limit, tpm_limit, allowed_models and budget_duration, plus additional windows in budget_limits. soft_budget blocks no request, it only triggers a notification. Members inherit a template through budget_id and can deviate with budget_override; cap_policy decides whether someone at their cap may request a temporary increase.

As a worked example, the announcement shows a customer with 100 dollars of purchased credits and 140 dollars of actual usage, which gives a margin of minus 40 percent (vendor figure, marked as illustrative in the blog itself, without other side costs).

Two notes sit in the docs themselves and matter more for operations than the feature list:

  • Free tier: Without an Enterprise licence, limits_apply: true holds with 5 organizations, 5 teams per organization and 5 members per team; the automatically created default team counts. Creating the sixth object on a level returns a 403 of type urn:litellm:error:customer-hierarchy-free-tier-limit. For a consulting offer with more than five customers this is the real price threshold.
  • Budget webhook: It is not bookkeeping. Delivery is best effort, with a 10-second timeout, no durable retry queue and no replay. The docs say, in essence, not to use the callback as the only basis for invoices. Usage is also booked in the background, so a successful request can still show no usage entry.

The availability table reads like a building site

Of nine pull requests named in the docs, seven were merged on November 16 and two were closed without a merge.

LiteLLM assigns each capability to a pull request number and distinguishes main, preview and "under review". I measured this mapping against the GitHub API on November 16, because the state of a pull request is independent of the announcement.

Pull requestDocs sayMerged?Measured time (UTC)
#51546 Create members on first requeston mainyesNov 13, 05:15
#51568 Budgets across several workerson mainyesNov 13, 15:33
#51579 Budget events and webhookon mainyesNov 13, 17:16
#51596 Customer identity in logson mainyesNov 13, 17:39
#51645 Alerts after schema changeon mainyesNov 13, 18:06
#51607 Safe creation of organizationsunder reviewyesNov 13, 18:21
#51913 Move members between teamspreviewyesNov 15, 04:22
#51604 Cold start of hierarchy requestsunder reviewno, closednone
#51617 Identify members by emailpreviewno, closednone

The last two rows are the actual finding. On November 16 the docs described the email field on the member and the header x-litellm-customer-member-email as a feature and pointed to #51617 for it. That pull request is closed without anything merged, just like #51604. The docs of November 19 corrected this: according to the API reference, v1.111 has neither an email field nor the header, and an email may only serve as an ordinary member ID. The PR table itself is no longer there.

Measured on the source tree: main against the releases

On November 16, main held 152 customer paths and the latest stable version v1.108.9 held 6.

The release list alone does not help. In the twelve newest release descriptions, the word customer appears only in the three pre-releases v1.111.0-dev.1, -dev.2 and -dev.4, in no stable version and not in v1.110.0-rc.3 either. The latest stable version was v1.108.9, published on November 14 at 22:00 UTC.

The file tree is more telling. The GitHub API returns the complete tree of a tag or branch. I queried it for three states and counted paths with customer in the name, outside docs/ and ui/.

Version statePaths totalPaths with customerMigrations with customer
v1.106.4 (runs here)17,11560
v1.108.9 (latest stable, Nov 14)17,16460
main (state of Nov 16)19,2051527
customer paths in the LiteLLM source tree by version statev1.106.4 (our lane)6 · 6 pathsv1.108.9 (stable, Nov 16)6 · 6 pathsmain (Nov 16)152 · 152 paths0167
Own count over the complete tree listing of the GitHub API, recursive per state and not truncated. Counted were path names with customer, excluding docs/ and ui/. Raw data: GitHub API, tree of main. (source: Own measurement, Senn IT, Nov 16, 2026)

The 152 paths on main split into 87 entries in the litellm/ package outside the proxy folder, 26 in the proxy itself, 27 under tests/ and 12 evidence files under .github/pr-assets. Folders and files are counted together. The numbers say nothing about the quality of the feature. They only show the gap between the announced state and what can be shipped.

Then come the migrations (state of November 16). main holds seven SQL migrations with customer in the name, dated between November 12 and 15, among them 20261112000000_customer_hierarchy, 20261113130000_customer_hierarchy_schema_cleanup and 20261115000000_customer_hierarchy_indexes. The stable release has none. The schema thus moved on three days, and the last step of November 15 only sets indexes. The setup instructions explicitly require applying the customer migrations.

Measured on November 19: release candidate against stable

On November 19, 2026, release candidate v1.111.0-rc.2 has 154 customer paths and 8 migrations, while stable v1.110.1 has 7 paths and no migration.

The state moved after November 16. v1.111.0-rc.2 appeared at 05:40 UTC on November 19, and the first release candidate v1.111.0-rc.1 came on November 18. Its release notes list the customer hierarchy, among them #51459 (organization, team, member with budget templates), #51489 (check on the request path) and #51687 (free tier limit). The latest stable version is v1.110.1 from November 19 at 05:05 UTC. Its source tree holds the same paths as v1.108.9 plus one test file.

Version state (GitHub, queried Nov 19, 2026)Paths totalPaths with customerMigrations with customer
v1.110.1 (latest stable)18,75670
v1.111.0-rc.2 (release candidate)19,2791548
main19,90018412

Counted as on November 16: complete tree listing of the GitHub API, path names with customer, excluding docs/ and ui/. Measured with the same means, by Senn IT; the results are a snapshot of a branch that changes hourly.

When a release candidate becomes a stable version is set by the release cycle in the docs: nightly builds (dev) only run through the pipeline, an rc adds manual review and a window of about one week for early testers, and on the following Tuesday a passed rc is promoted to stable. If v1.111.0-rc.1 passes review, that would be Tuesday, November 24 (derived from the cycle, not announced by the vendor). An rc that fails is not promoted.

From nightly to stable version at LiteLLM1.x.x-dev.NNightly: CI/CD only, nomanual review1.x.x-rc.NRelease candidate: reviewplus about 7 days testwindow1.x.xStable: on the followingTuesday if the rc passes
Stages according to the release cycle in the LiteLLM docs. On November 19, Customer Management sits on the middle stage. (source: LiteLLM docs, Release Cycle)

Two constraints from the docs affect the choice of line. The version support provides patches for the four most recent stable lines, but since the week of November 2 it is suspended for a one-month LTS pilot, whose first LTS line is 1.108. For our line 1.106, v1.106.11 still appeared on November 19, a backport of a prompt-caching fix for OpenAI models.

Budget per customer on the stable line

On a stable version the existing end-user tracking stays. The LiteLLM docs on customer management (retrieved on November 19, 2026) say this tracking keeps working, but that both systems have separate records and separate usage interfaces.

Existing end-user trackingCustomer Management
Available fromstable versions before v1.111v1.111.0 (only as release candidate on Nov 19, 2026)
Identity on the requestone flat end-user value in user, x-litellm-customer-id or x-litellm-end-user-idorganization alias and member ID in x-litellm-customer-org-alias and x-litellm-customer-member-id
Management/customer/new, /customer/update, /customer/info/management/v1/customer_organizations, customer_teams, customer_members
Budgetone value per end usercaps per organization and team, templates and overrides per member

Source: LiteLLM docs "Customer Management and end-user tracking", retrieved on November 19, 2026, vendor statement. We did not re-measure the existing routes on our gateway. When moving over, LiteLLM according to the docs does not copy old usage or budget counters into the new records, so switching a budget mid-period can give a customer a fresh allowance.

Measured on our own gateway

Our gateway on LiteLLM v1.106.4 answered the two documented Customer Management routes with 404 on November 16, 2026, and the control routes with 200.

Our lane runs on ghcr.io/berriai/litellm:v1.106.4, container litellm-router, started on November 15 at 22:33 UTC, with no restart counter. For checking your own build the docs name a simple way: look at /openapi.json. That is what we did, plus a route sample.

Measurement on Nov 16, 2026, Senn ITResult
GET /health/liveliness (control)200
GET /v1/models (control)200
GET /management/v1/customer_portal/limits404
GET /management/v1/customer_organizations404
GET /openapi.json200, 3,953 bytes, no hit on customer_organizations, customer_portal or /management/v1

The two control routes are the important part of the measurement. Without them a 404 would just be a 404. This establishes that the routes are missing in our state while the API is reachable. The openapi field on our gateway is small at 3,953 bytes and may be served shortened. The route probe stays authoritative, the JSON is the second piece of evidence.

What this means for an operation

First, the free tier of 5 organizations is a business limit, not a technical one. An offer that bills per customer hits a 403 response at the sixth customer (derived from the measured meaning of organization_limit and the documented error type). For smaller operations that may fit; for a company that models customer groups, it is the point where open source turns into a licence case.

Second, the state of main is not a delivery state. A build from main hangs on no update line, and its migrations (twelve on November 19) land in the same database as keys and usage records. If you want to test it, start an empty second container with its own database and check GET /management/v1/customer_portal/limits there before anything changes on a production gateway. If you expect ready-made images: according to LiteLLM's incident report, the Docker tags main-latest and main-stable were among the tags that pointed to a tampered image on October 21, and were afterwards set to safe sources. Since v1.111.0-dev.3, releases are signed keyless with cosign, and the release notes carry the verification command. A release candidate can therefore be checked against its signature, a nightly tag without a version number cannot sensibly be. How we checked on October 24, 2026 whether our containers ran in the malware window is in the post on LiteLLM 1.107.0; that was a different question.

Third, customer costs are a billing matter, and billing needs traceability. A webhook without retry and a counter that is posted in the background are enough as a trigger. For an invoice you need the usage read with your own time window. Until then we keep attributing by key and our own tags, as it has run in our own AI stack for months, plus evaluation of the usage records. If you need that for several company operations, the building blocks are in our AI gateway for companies.

Watching is still worthwhile: the feature solves a real problem for many who resell AI requests to third parties. The catch in operation lies in delivery, and the best way to measure that is yourself. More posts on running AI in a company are collected in the hub AI in production.

What did not happen

Nothing changed on our gateway: no build of main, no applied migration. The configuration file stayed as measured, and the pre-releases v1.111.0-dev.* were left alone, since they carry no promise of stability. The measured version v1.106.4 stays in use until a stable release contains the routes and a test in a separate environment has run through. Open is whether the lane should first move to the patch state v1.106.11 of its own line; that is a separate decision.

Further Reading

Questions?
Which LiteLLM version contains Customer Management?+

On November 19, 2026 only the release candidate v1.111.0-rc.2. The latest stable version v1.110.1 contains none of the new routes and no migration, measured on the source tree of the release tag. The release candidate carries eight migrations and, per the release cycle in the docs, is not yet promoted. To check whether a tag contains the feature, list its tree: it must show customer_hierarchy_auth and migrations with customer in the name.

How do I set a budget per customer in LiteLLM while no stable version has Customer Management?+

With the existing end-user tracking: the customer ID goes into user, x-litellm-customer-id or x-litellm-end-user-id and is managed through /customer/new, /customer/update and /customer/info. According to the LiteLLM docs of November 19, 2026, this tracking keeps working. It knows one flat ID per end user, no hierarchy of organization, team and member.

What does LiteLLM Customer Management cost?+

The feature is part of the open-source version, but without an Enterprise licence it is capped at 5 organizations, 5 teams per organization and 5 members per team. GET /management/v1/customer_portal/limits shows the cap and the current count of your own build.

Is the LiteLLM budget webhook enough as a billing basis for customers?+

No, and LiteLLM says so itself. Delivery is best effort with a 10-second timeout, there is no durable retry queue and no replay. For billing, the usage reads under /management/v1 are authoritative, with background accounting and a settling delay.