test(e2e): source the cheap Anthropic model from e2e_config across suites

[mateo-berri]

Relevant issues

Linear ticket

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• 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)

Screenshots / Proof of Fix

This PR is stacked on #32914, which introduces the CHEAP_ANTHROPIC_MODEL constant in tests/e2e/e2e_config.py; merge that first

This is a pure test-code refactor with no proxy behavior change, so the proof is the touched suites still passing against a live proxy plus the type and registry gates, all captured at commit e44aa07 against the tests/e2e docker compose stack (ghcr.io/berriai/litellm:main-latest with the suite's inline config, mapped to host port 4010 here because 4000 was taken by another local proxy) with real Anthropic and OpenAI traffic:

$ cd tests/e2e && LITELLM_PROXY_URL=http://localhost:4010 uv run pytest \
    llm_translation/test_chat_completions_regression_e2e.py llm_translation/test_passthrough_e2e.py \
    quota_management/budgets/test_soft_budget_e2e.py quota_management/spend_tracking/test_spend_tracking_e2e.py \
    -k "claude or anthropic_passthrough or soft_budget_does_not_block or spend_calculate_returns_nonzero" -v
llm_translation/test_chat_completions_regression_e2e.py::TestChatCompletionsRegression::test_chat_returns_real_completion[claude-haiku-4-5-anthropic] PASSED
llm_translation/test_passthrough_e2e.py::test_anthropic_passthrough_nonstreaming_logs_cost PASSED
llm_translation/test_passthrough_e2e.py::test_anthropic_passthrough_streaming_logs_cost PASSED
llm_translation/test_passthrough_e2e.py::test_anthropic_passthrough_tool_call_logs_cost PASSED
quota_management/budgets/test_soft_budget_e2e.py::test_soft_budget_does_not_block PASSED
quota_management/spend_tracking/test_spend_tracking_e2e.py::test_spend_calculate_returns_nonzero_cost PASSED
====================== 6 passed, 18 deselected in 40.04s =======================

$ uv run --no-sync basedpyright tests/e2e
0 errors, 0 warnings, 0 notes

$ cd tests/e2e && uv run python -m coverage_registry.collector --strict
Headline coverage: 53/313  (16.9%)

The --strict collector run performs a full pytest collect-only pass over tests/e2e, so it proves every touched module still imports and every covers marker still resolves to a registry cell

Type

🧹 Refactoring

Changes

Follow-up to #32914, which added CHEAP_ANTHROPIC_MODEL to tests/e2e/e2e_config.py (env E2E_CHEAP_ANTHROPIC_MODEL, default claude-haiku-4-5) after review feedback that a hardcoded model name is hard to maintain, but only used it in the new ratelimit suite; the older suites still hardcode the alias, so retiring or swapping the cheap Anthropic model would mean touching dozens of call sites

This replaces the 22 remaining hardcoded "claude-haiku-4-5" proxy-alias literals across 16 files with the shared constant, each with the matching e2e_config import: 14 occurrences across 12 files in quota_management/budgets/, the DRIVER_MODELS proxy alias in quota_management/spend_tracking/conftest.py plus a call and two row-matching assertions in test_spend_tracking_e2e.py, the 3 Anthropic passthrough calls in llm_translation/test_passthrough_e2e.py, and the claude row of CHAT_MODELS in llm_translation/test_chat_completions_regression_e2e.py. Every occurrence was a standalone literal, so no f-strings were needed and no behavior changes

Following review feedback (commit 65e90c2, gates rerun clean: basedpyright 0 errors, collector --strict passes, the spend_calculate test passes live), the spend_tracking DRIVER_MODELS row now derives its provider route as f"anthropic/{CHEAP_ANTHROPIC_MODEL}", so overriding the env var on a proxy where the alias is not pre-registered can no longer silently register the new alias against the old upstream model, and the passthrough module docstring references CHEAP_ANTHROPIC_MODEL instead of naming the model

Deliberately left alone: the upstream provider route id "anthropic/claude-haiku-4-5" in llm_translation/test_messages_e2e.py (that test registers its own e2e-messages-* alias, so its alias and route stay consistent regardless of the env var), dated bedrock model ids, and the docker-compose.yml inline config, which defines the alias the stack actually serves and matches the constant's default

[greptile-apps]

Greptile Summary

Pure test-code refactor that replaces 22 hardcoded "claude-haiku-4-5" proxy-alias literals across 16 e2e test files with the CHEAP_ANTHROPIC_MODEL constant imported from e2e_config.py (env-backed via E2E_CHEAP_ANTHROPIC_MODEL, default "claude-haiku-4-5"). No proxy behaviour changes; all assertions and model registration logic remain functionally identical to the previous state.

• All 16 changed files receive a matching from e2e_config import CHEAP_ANTHROPIC_MODEL import and substitute every standalone literal occurrence with the constant.
• quota_management/spend_tracking/conftest.py derives the provider route dynamically as f"anthropic/{CHEAP_ANTHROPIC_MODEL}", so overriding the env var no longer silently registers the new alias against a mismatched upstream model.
• Row-matching predicates in test_spend_tracking_e2e.py use CHEAP_ANTHROPIC_MODEL in m, which is correct because the logged proxy alias will equal the constant value at runtime.

Confidence Score: 5/5

Safe to merge — test-only refactor with no proxy code touched and verified clean against a live stack.

Every change is a mechanical substitution of a string literal with a shared constant; no test logic, assertions, or fixtures were weakened. The provider-route derivation in conftest.py was tightened to use the same constant, which is a strict improvement. All touched modules were verified with basedpyright (0 errors) and the coverage registry collector, and the affected test functions were run against a real Anthropic-backed proxy.

No files require special attention.

Important Files Changed

Filename	Overview
tests/e2e/e2e_config.py	Defines CHEAP_ANTHROPIC_MODEL constant (env-backed, default "claude-haiku-4-5") that all changed files now import — the source of truth for this refactor
tests/e2e/quota_management/spend_tracking/conftest.py	DRIVER_MODELS now derives the provider route dynamically as f"anthropic/{CHEAP_ANTHROPIC_MODEL}", addressing the previous review concern about alias/route mismatch
tests/e2e/quota_management/spend_tracking/test_spend_tracking_e2e.py	Row-matching assertions updated to use CHEAP_ANTHROPIC_MODEL; substring matching is correct since the logged proxy alias will equal the constant value
tests/e2e/llm_translation/test_passthrough_e2e.py	Three call sites updated to CHEAP_ANTHROPIC_MODEL; docstring updated from the literal model name to the constant name
tests/e2e/llm_translation/test_chat_completions_regression_e2e.py	CHAT_MODELS tuple anthropic entry updated to CHEAP_ANTHROPIC_MODEL — mechanical substitution, no logic change

_{Reviews (2): Last reviewed commit: "fix(e2e): derive the anthropic driver ro..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[mateo-berri]

@greptileai

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_e2e_haiku_model_refs (318c5c0) with litellm_internal_staging (582907d)}