feat(proxy): add admin toggle to block requests for models without pricing

[devin-ai-integration]

TLDR

Problem this solves:

• Requests for models with no cost mapping get no pricing from litellm, so their spend reads $0 unless the provider happens to return a cost of its own
• Under-billing hides misconfiguration until someone reconciles spend by hand

How it solves it:

• New admin toggle, off by default, rejects unpriced-model requests with 403
• It covers every LLM API route, so /chat/completions, /responses, /messages, and /embeddings all refuse the same unpriced model; management routes are untouched
• The 403 names every unpriced model the request asked for and says pricing is missing
• A group counts as priced when any deployment prices any billed metric (tokens, characters, seconds, pages, images) or carries tiered pricing, so non-token-billed models stay allowed
• Explicitly configured prices, including zero, still count as priced
• Toggle persists to litellm_settings; peer workers apply it on config reload
• Admins flip "Block Unpriced Models" on the Cost Tracking settings page

User Flow

Before: a request for a model missing from the cost map sails through and litellm prices it at $0, and no setting exists to refuse it

1. A developer sends POST https://litellm-domain/v1/chat/completions with {"model": "team-onprem-llm", ...} where that model has no entry in the cost map, and gets 200 with a normal completion
2. The proxy admin opens https://litellm-domain/ui/?page=logs and sees that request logged at $0.00, since litellm has no price to apply and most providers return no cost of their own
3. The admin opens https://litellm-domain/ui/?page=cost-tracking and finds no way to block unpriced models, so every later request repeats steps 1 and 2

After: once the admin flips the new toggle, the same request is refused with a 403 naming the model

1. The proxy admin opens https://litellm-domain/ui/?page=cost-tracking, expands "Block Unpriced Models", and turns the switch on; a toast confirms unpriced-model requests will now be blocked
2. The developer sends the same POST https://litellm-domain/v1/chat/completions with {"model": "team-onprem-llm", ...} and gets 403 with error type model_cost_map_missing and a message naming team-onprem-llm and explaining its pricing is missing
3. The admin adds pricing for team-onprem-llm, either in the cost map or as an input_cost_per_token/output_cost_per_token override on that deployment, and the developer's next identical request returns 200 again with spend litellm can actually price

Relevant issues

Linear ticket

Resolves LIT-4984

Pre-Submission checklist

• I have added meaningful tests (auth check: toggle off allows, toggle on blocks with 403 naming the model, priced model allowed, non-LLM route ignored, alias resolving to an unpriced group blocked, a request naming several models blocked when any one of them is unpriced and allowed when all of them are priced, an alias over a group priced through its model_info block allowed, models priced per character, page, image, via litellm_params override or tiered pricing allowed; settings endpoints: GET reflects the flag, PATCH persists and updates it, PATCH requires STORE_MODEL_IN_DB, peer worker applies the persisted value on config reload)
• The handful of test files covering my change pass locally, e.g. uv run pytest tests/test_litellm/<your_test_file>.py -v. Leave the suites (make test-unit-*, make test-unit) to CI: it finishes in ~15 minutes where a laptop takes an hour or more
• My PR passes all required CI/CD checks (e.g., lint, schema.d.ts sync check, etc.)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

Screenshots / Proof of Fix

Two live proxies, both booted from a real checkout against a real Postgres with STORE_MODEL_IN_DB=True, both calling real provider APIs and spending real money. No mocks, no stubs, no pytest

team-onprem-llm points at openrouter/moonshotai/kimi-k2-0905, which OpenRouter serves today and which has no entry in model_prices_and_context_window.json, so it is genuinely unpriced and genuinely callable. priced-openai-llm points at openai/gpt-4o-mini, which the cost map prices

Before, at the merge base 996693f1eb (proxy on 127.0.0.1:25070)

The setting does not exist and the unpriced model answers normally, so there is nothing an admin can do about it

$ curl -sS -w '\nHTTP_STATUS:%{http_code}\n' -X GET \
  'http://127.0.0.1:25070/config/block_requests_for_models_without_pricing' \
  -H "Authorization: Bearer $LITELLM_KEY"
{"detail":"Not Found"}
HTTP_STATUS:404

$ curl -sS -w '\nHTTP_STATUS:%{http_code}\n' -X PATCH \
  'http://127.0.0.1:25070/config/block_requests_for_models_without_pricing' \
  -H "Authorization: Bearer $LITELLM_KEY" -H 'Content-Type: application/json' \
  -d '{"enabled":true}'
{"detail":"Not Found"}
HTTP_STATUS:404

$ curl -sS -w '\nHTTP_STATUS:%{http_code}\n' -X POST \
  'http://127.0.0.1:25070/v1/chat/completions' \
  -H "Authorization: Bearer $LITELLM_KEY" -H 'Content-Type: application/json' \
  -d '{"model":"team-onprem-llm","messages":[{"role":"user","content":"Reply with exactly: pong"}]}'
{"id":"gen-1787266870-Qg636PUs3nhpBbaGrPZf","model":"team-onprem-llm","object":"chat.completion",
 "choices":[{"finish_reason":"stop","index":0,"message":{"content":"pong","role":"assistant"}}],
 "usage":{"completion_tokens":3,"prompt_tokens":14,"total_tokens":17}}
HTTP_STATUS:200

GET /model/info for that group at the same commit reports "input_cost_per_token": 0, "output_cost_per_token": 0, and every other cost field null, which is exactly the state this PR refuses to bill against

After, at the PR tip c73480c653 (proxy on 127.0.0.1:47831)

Step 0, the toggle exists and defaults to off:

$ curl -isS -X GET "$PROXY/config/block_requests_for_models_without_pricing" \
  -H "Authorization: Bearer $LITELLM_KEY"
HTTP/1.1 200 OK
{"enabled":false}

Step 1, with it off the unpriced model still answers, same as before the PR:

$ curl -isS -X POST "$PROXY/v1/chat/completions" \
  -H "Authorization: Bearer $LITELLM_KEY" -H 'Content-Type: application/json' \
  -d '{"model":"team-onprem-llm","messages":[{"role":"user","content":"Reply with exactly: OK"}],"max_tokens":16}'
HTTP/1.1 200 OK
x-litellm-model-name: openrouter/moonshotai/kimi-k2-0905
{"id":"gen-1787269183-btisPNuevjEUYBjxtLyc","model":"team-onprem-llm",
 "choices":[{"finish_reason":"stop","index":0,"message":{"content":"OK","role":"assistant"}}],
 "usage":{"completion_tokens":2,"prompt_tokens":14,"total_tokens":16}}

Step 2 and 3, the admin turns it on through the endpoint the Cost Tracking page calls, and it reads back on:

$ curl -isS -X PATCH "$PROXY/config/block_requests_for_models_without_pricing" \
  -H "Authorization: Bearer $LITELLM_KEY" -H 'Content-Type: application/json' \
  -d '{"enabled": true}'
HTTP/1.1 200 OK
{"enabled":true}

$ curl -isS -X GET "$PROXY/config/block_requests_for_models_without_pricing" \
  -H "Authorization: Bearer $LITELLM_KEY"
HTTP/1.1 200 OK
{"enabled":true}

Step 4, the identical request from step 1 is now refused, and the message names the model:

$ curl -isS -X POST "$PROXY/v1/chat/completions" \
  -H "Authorization: Bearer $LITELLM_KEY" -H 'Content-Type: application/json' \
  -d '{"model":"team-onprem-llm","messages":[{"role":"user","content":"Reply with exactly: OK"}],"max_tokens":16}'
HTTP/1.1 403 Forbidden
{"error":{"message":"Model 'team-onprem-llm' has no pricing in the cost map, so litellm cannot price the request. Requests for unpriced models are blocked because 'block_requests_for_models_without_pricing' is enabled. Add pricing (input_cost_per_token/output_cost_per_token) to allow the request.","type":"model_cost_map_missing","param":"model","code":"403"}}

Step 5, priced groups keep working with the toggle on, so the block is narrow:

priced-openai-llm       openai/gpt-4o-mini, priced by the cost map        HTTP/1.1 200 OK
team-onprem-llm-priced  same unpriced model, explicit per-token prices    HTTP/1.1 200 OK

Step 6, a request naming several models is blocked when any one of them is unpriced, and books no spend:

$ curl -sS "$PROXY/spend/logs" -H "Authorization: Bearer $LITELLM_KEY" | jq ...
spend_log_rows=3 total_spend=1.58e-05

$ curl -isS -X POST "$PROXY/v1/chat/completions" \
  -H "Authorization: Bearer $LITELLM_KEY" -H 'Content-Type: application/json' \
  -d '{"model":"team-onprem-llm,priced-openai-llm","messages":[{"role":"user","content":"Reply with exactly: OK"}],"max_tokens":16}'
HTTP/1.1 403 Forbidden
{"error":{"message":"Model 'team-onprem-llm' has no pricing in the cost map, ...","type":"model_cost_map_missing","param":"model","code":"403"}}

$ curl -sS "$PROXY/spend/logs" -H "Authorization: Bearer $LITELLM_KEY" | jq ...
spend_log_rows=3 total_spend=1.58e-05

Steps 7 to 10, aliases and groups priced only through their model_info block are still allowed with the toggle on:

priced-alias                      -> priced-openai-llm                HTTP/1.1 200 OK
zero-priced-alias                 -> explicitly zero-priced group     HTTP/1.1 200 OK
team-onprem-llm-modelinfo-priced  priced via model_info only          HTTP/1.1 200 OK
modelinfo-alias                   -> that model_info-priced group     HTTP/1.1 200 OK

The same run against the pre-fix commit, to show the last two fixes were real

Same proxy setup, same database, same config, only the commit differs:

Request, toggle on	at df00c334	at c73480c653
team-onprem-llm,priced-openai-llm	200, request billed	403, spend unchanged
modelinfo-alias	403, wrongly blocked	200
team-onprem-llm-modelinfo-priced	200	200

The pre-fix multi-model call reached the provider and billed cost: 1.34e-05, so it was spending money straight through the gate the toggle is supposed to close. The pre-fix modelinfo-alias 403 is the false block, and the group it aliases returned 200 in that same run, which pins the failure to alias resolution rather than to pricing detection

Type

🆕 New Feature

Caveats (if any)

• UI toggle persistence needs STORE_MODEL_IN_DB=True
• Peer workers pick the change up on their next config reload
• A model group with an explicitly configured zero price stays allowed
• A model the built-in cost map prices at exactly zero, such as codestral/codestral-2405, still counts as unpriced, since a deployment litellm cannot price and one it prices at zero look the same by the time the router has registered it. Declaring input_cost_per_token: 0 in that model's litellm_params keeps it allowed
• The periodic DB sync now applies every LITELLM_SETTINGS_SAFE_DB_OVERRIDES key the database holds, not only the new toggle. That matches what the startup config load has always done at proxy_server.py:6552, so a worker's live settings no longer drift from a peer's between restarts

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

• c73480c passes /live-pr-risk

Link to Devin session: https://app.devin.ai/sessions/630f4ee5757e4211b2287850b12c09fa
Requested by: @mateo-berri

---

Note

Medium Risk
Touches proxy auth and can 403 production LLM traffic when enabled. Also applies DB-backed litellm_settings onto live workers, so a bad or unexpected override could change request behavior cluster-wide.

Overview
Adds an opt-in block_requests_for_models_without_pricing flag (default off) so admins can reject LLM requests whose resolved model has no cost mapping instead of logging them as $0 spend.

When enabled, common_checks returns 403 with model_cost_map_missing. A group is treated as priced if any deployment declares a cost_per* field (including explicit zero), tiered_pricing, or a positive billed metric (tokens, seconds, pages, images, etc.). Non-LLM routes are skipped.

GET/PATCH /config/block_requests_for_models_without_pricing persist the flag in litellm_settings. Peer workers pick it up via a new safe-overrides apply on the periodic DB sync. Cost Tracking UI adds a proxy-admin switch for the same setting.

^{Reviewed by Cursor Bugbot for commit df00c33. Bugbot is set up for automated code reviews on this repo. Configure here.}

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[greptile-apps]

Greptile Summary

Adds an opt-in, DB-persisted setting that blocks requests for model groups without recognized pricing.

• Detects token, non-token, tiered, explicitly configured zero, alias-resolved, and model-info pricing
• Returns a structured 403 response when blocking is enabled and an unpriced model is requested
• Synchronizes the persisted setting across workers during periodic config reconciliation
• Adds an admin dashboard toggle, generated API types, and regression coverage

Confidence Score: 5/5

The PR appears safe to merge

No blocking failure remains

Important Files Changed

Filename	Overview
litellm/proxy/auth/auth_checks.py	Adds alias-aware pricing detection and opt-in enforcement in centralized request checks; the previously reported pricing cases are handled
litellm/proxy/proxy_server.py	Applies allowlisted persisted LiteLLM settings during every periodic reconciliation, resolving the reported peer-worker inconsistency
litellm/proxy/management_endpoints/cost_tracking_settings.py	Adds authenticated GET and PATCH endpoints for reading and persisting the global toggle
tests/test_litellm/proxy/auth/test_auth_checks.py	Covers priced and unpriced groups, aliases, non-token metrics, tiered pricing, explicit zero pricing, and route enforcement while retaining project-alias coverage
tests/test_litellm/proxy/management_endpoints/test_cost_tracking_settings.py	Covers endpoint persistence, in-memory updates, configuration requirements, and peer-worker reconciliation under both override-gate states
ui/litellm-dashboard/src/app/(dashboard)/cost-tracking/_components/cost_tracking_settings.tsx	Adds the proxy-admin-only control for the new setting to Cost Tracking
ui/litellm-dashboard/src/app/(dashboard)/cost-tracking/_components/use_block_unpriced_config.ts	Adds dashboard state, API integration, loading behavior, and success or failure notifications for the toggle

_{Reviews (8): Last reviewed commit: "fix(proxy): block every unpriced model a..." | Re-trigger Greptile}

[codecov]

Codecov Report

❌ Patch coverage is 70.83333% with 35 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...oxy/management_endpoints/cost_tracking_settings.py	56.41%	17 Missing ⚠️
litellm/proxy/auth/auth_checks.py	79.71%	14 Missing ⚠️
litellm/proxy/proxy_server.py	50.00%	4 Missing ⚠️

📢 Thoughts on this report? Let us know!

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_block_unpriced_models (c73480c) with litellm_internal_staging (e07a712)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (b2aff8b) during the generation of this report, so e07a712 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[devin-ai-integration]

QA: live proxy, real provider APIs, no mocks

Tested against a live LiteLLM proxy on localhost:4000 with LITELLM_LOCAL_MODEL_COST_MAP=True and STORE_MODEL_IN_DB=True, hitting real Fireworks and OpenAI. All scenarios pass

• Flag OFF (default): request to unpriced unpriced-fw returns 200 with a real completion, no behavior change
• PATCH /config/block_requests_for_models_without_pricing {"enabled": true} returns {"enabled": true}; the same request now returns 403 model_cost_map_missing naming the model
• Flag ON: priced gpt-4o-mini still returns 200, so priced traffic is not over-blocked
• Flag ON: alias unpriced-alias returns 403, confirming resolution through get_model_group_info
• Admin UI toggle: defaults OFF, flips ON with a success notification, persists across reload (GET returns true), and gates live Playground traffic (403 while ON, 200 while OFF)

Raw 403 body:

{"error":{"message":"Model 'unpriced-fw' has no pricing in the cost map, so its spend would be tracked as $0. Requests for unpriced models are blocked because 'block_requests_for_models_without_pricing' is enabled. Add pricing for this model (input_cost_per_token/output_cost_per_token) to allow it.","type":"model_cost_map_missing","param":"model","code":"403"}}

Admin UI toggle gating live traffic end to end:

Playground request blocked with 403 while the toggle is ON:

Same request succeeds once the toggle is OFF:

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ mateo-berri
❌ web-flow
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[mateo-berri]

@greptileai

[mateo-berri]

@greptileai

[mateo-berri]

@greptileai

[mateo-berri]

@greptileai

[mateo-berri]

@greptileai

[mateo-berri]

@greptileai

[greptile-apps]

Disabled config overrides still apply

When config_overrides is excluded, this unconditional reload applies every allowlisted database value, overriding unrelated YAML settings despite configuration override loading being disabled

Knowledge Base Used: Proxy Server Request Flow

[mateo-berri]

config_overrides gates the vault table, not litellm_settings. Startup already applies these DB values ungated, so the periodic reload just matches it

[greptile-apps]

You're right — the original comment conflated two distinct loading mechanisms. Looking at the diff:

if self._should_load_db_object(object_type="config_overrides"):
    await self._init_hashicorp_vault_config_override(prisma_client=prisma_client)

await self._apply_safe_litellm_settings_overrides_from_db(prisma_client=prisma_client)

The config_overrides gate is specifically for the Hashicorp Vault table (_init_hashicorp_vault_config_override), not for litellm_settings. Since startup already applies litellm_settings values from the DB unconditionally (outside that gate), having the periodic reload do the same is correct and consistent — not a bypass of any intentional guard. I withdraw the comment.

[mateo-berri]

bugbot run

[mateo-berri]

@greptileai

[mateo-berri]

@greptileai

[mateo-berri]

bugbot run

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit c73480c. Configure here.}

[mateo-berri]

LGTM