feat(proxy): serve Anthropic-native /v1/models for Claude Code gateway discovery

[Ar-maan05]

Re-lands #30273 (reverted during staging). No functional changes to the original; rebased onto current main and refactored so the diff is purely additive.

The problem

Claude Code 2.1.126+ added gateway model discovery: when ANTHROPIC_BASE_URL points at a gateway, it queries {base_url}/v1/models at startup and populates the /model picker with the discovered models. That discovery only parses the Anthropic-native Models API shape, so against litellm - which returns OpenAI's {id, object, created, owned_by} list - Claude Code finds nothing and the picker stays empty, even though /v1/messages already works.

The fix

This serves the Anthropic-native shape from the same /v1/models route via content negotiation on the anthropic-version header. Claude Code already sends that header for /v1/messages, so when it is present the endpoint returns the Anthropic Models envelope (type / display_name / created_at per entry, plus top-level has_more / first_id / last_id); otherwise the response is byte-for-byte the existing OpenAI shape, so aider and other OpenAI-compatible clients are unaffected.

Entries also carry the model's token limits when litellm knows them: max_input_tokens as-is, and the output budget as max_tokens, the name the Messages API uses for it. Both come from create_model_info_response, the same helper the OpenAI listing goes through, so cost-map values and router-configured overrides resolve identically in either shape

A separate endpoint was not used because Claude Code discovers at the gateway root's /v1/models, and a global config flag would break the OpenAI clients that share the route.

Design notes

• Full list returned, client filters. Claude Code applies its own claude/anthropic id-prefix filter client-side, so no server-side filtering is imposed - a model aliased to claude-* pointing at any backend still shows up, which is the point for gateway users.
• display_name falls back to the model id, the stable label a gateway can offer for arbitrary upstream models.
• created_at is the ISO 8601 (Z-suffixed) form of the same timestamp the OpenAI shape already returns.
• Hidden/unhealthy models are filtered before formatting in both the scope=expand and normal branches, exactly as for the OpenAI shape.
• Both branches now build the OpenAI listing first and reshape it for Anthropic callers, so team name translation, hidden-model filtering and token limits can never drift between the two shapes. A model with no known limits simply omits the keys.
• The formatter lives in litellm/llms/anthropic/common_utils.py per the repo rule against provider-specific code outside llms/; the request parameter is optional so existing direct (non-HTTP) callers of model_list are unaffected.

Tests

• tests/test_litellm/llms/anthropic/test_anthropic_common_utils.py: full envelope shape (per-entry type/display_name/created_at with a Z-suffixed ISO timestamp, top-level has_more/first_id/last_id, no object) and the empty-list case (first_id/last_id null).
• tests/test_litellm/llms/anthropic/test_anthropic_common_utils.py also pins the token limits: max_input_tokens passes through, max_output_tokens is renamed to max_tokens, and a model with only one known limit or none omits the rest.
• tests/test_litellm/proxy/proxy_server/test_routes_models.py: drives GET /v1/models (and /models) with the anthropic-version header and asserts the negotiated Anthropic shape, including the limits for a model that has them and their absence for one that does not; the existing happy-path test pins that the default response stays OpenAI.

All three affected test files pass locally (215 tests). No server-side state or config changes.

[greptile-apps]

Greptile Summary

Adds Anthropic-native model discovery responses to the existing authenticated model-list routes while preserving the default OpenAI response.

• Selects the Anthropic envelope when anthropic-version is present and formats model identifiers with ISO 8601 creation timestamps.
• Adds route and formatter coverage for populated and empty model lists.
• Extends the E2E transport with per-request GET timeouts and strengthens model-readiness polling for multi-worker propagation.

Confidence Score: 5/5

The PR appears safe to merge with no concrete changed-code failure identified.

The negotiated response is selected only for requests carrying the Anthropic-specific header, after the existing authorization and model-visibility filtering, while direct callers and default OpenAI clients retain their prior behavior.

Important Files Changed

Filename	Overview
litellm/proxy/proxy_server.py	Negotiates the model-list response from the request header after existing authorization and visibility filtering, with no accepted correctness issue.
litellm/llms/anthropic/common_utils.py	Adds a focused formatter for the Anthropic Models API envelope, including safe empty-list handling.
tests/e2e/proxy_client.py	Adds deadline-aware continuous model-visibility polling before E2E callers use newly created deployments.
tests/e2e/transport.py	Adds an optional per-call GET timeout and propagates it through both HTTP and split transports.
tests/test_litellm/llms/anthropic/test_anthropic_common_utils.py	Covers the populated and empty Anthropic model-list envelopes.
tests/test_litellm/proxy/proxy_server/test_routes_models.py	Verifies header-based negotiation on both model-list route aliases while existing tests retain the default OpenAI contract.

_{Reviews (1): Last reviewed commit: "fix(proxy): make model_list request para..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing Ar-maan05:feat-anthropic-native-v1-models-2 (b3248d7) with litellm_internal_staging (423b791)}

[veria-ai]

PR overview

All previously flagged issues have been addressed. No open security concerns remain on this pull request.

Security review

No open security issues remain on this pull request.

Fixed/addressed: 1 · PR risk: 0/10

[connorgh14789]

@Ar-maan05 any update on this? We are running into the same issue in our org trying to implement LiteLLM

[yassin-berriai]

This is green and merges cleanly, please review for merge. Worth adding max_input_tokens and max_tokens to the Anthropic entries, create_model_info_response already computes both

[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.
2 out of 3 committers have signed the CLA.

✅ yuneng-berri
✅ Ar-maan05
❌ yassin-berriai
_{You have signed the CLA already but the status is still pending? Let us recheck it.}