From 9d64fa4201b372e3bf679dac4fcb51536dcefecd Mon Sep 17 00:00:00 2001 From: Yuneng Jiang Date: Mon, 2 Nov 2026 16:57:17 -0800 Subject: [PATCH] test(e2e): refresh retired provider fixtures --- tests/e2e/guardrails/guardrails_client.py | 2 +- .../test_chat_completions_regression_e2e.py | 4 ++-- .../e2e/llm_translation/test_custom_pricing_e2e.py | 10 ++++------ tests/e2e/llm_translation/test_google_native_e2e.py | 2 +- tests/e2e/llm_translation/test_image_edits_e2e.py | 2 +- tests/e2e/llm_translation/test_passthrough_e2e.py | 6 +++--- tests/e2e/llm_translation/test_responses_e2e.py | 2 +- .../spend_tracking/test_key_attribution_e2e.py | 2 +- .../spend_tracking/test_spend_tracking_e2e.py | 6 +++--- tests/e2e/router/reliability_support.py | 13 +++++-------- 10 files changed, 22 insertions(+), 27 deletions(-) diff --git a/tests/e2e/guardrails/guardrails_client.py b/tests/e2e/guardrails/guardrails_client.py index cf517c7a4fe..9db95822734 100644 --- a/tests/e2e/guardrails/guardrails_client.py +++ b/tests/e2e/guardrails/guardrails_client.py @@ -38,7 +38,7 @@ from models import ( from proxy_client import ProxyClient from pydantic import BaseModel, Field -GUARDRAIL_BACKEND: Final = "gemini/gemini-2.5-flash" +GUARDRAIL_BACKEND: Final = "gemini/gemini-3.8-flash" GuardrailMode = Literal["pre_call", "post_call", "during_call", "logging_only"] BlockedWordAction = Literal["BLOCK", "MASK"] diff --git a/tests/e2e/llm_translation/test_chat_completions_regression_e2e.py b/tests/e2e/llm_translation/test_chat_completions_regression_e2e.py index 1e0a8cbb412..4967f18818a 100644 --- a/tests/e2e/llm_translation/test_chat_completions_regression_e2e.py +++ b/tests/e2e/llm_translation/test_chat_completions_regression_e2e.py @@ -45,7 +45,7 @@ from passthrough_client import PassthroughClient pytestmark = pytest.mark.e2e COHERE_BACKEND = "cohere/command-r-08-2024" -GEMINI_BACKEND = "gemini/gemini-2.5-flash" +GEMINI_BACKEND = "gemini/gemini-3.8-flash" VERTEX_BACKEND: Final = "vertex_ai/gemini-2.5-flash" AZURE_OPENAI_BACKEND: Final = "azure/gpt-5.4-nano" AZURE_OPENAI_API_VERSION: Final = "v1" @@ -243,7 +243,7 @@ _PERSON_SCHEMA: dict[str, object] = { OPENAI_CHAT_MODEL: Final = "gpt-5.5" ANTHROPIC_CHAT_MODEL: Final = "claude-haiku-4-5" -GEMINI_FLASH_MODEL: Final = "gemini-2.5-flash" +GEMINI_FLASH_MODEL: Final = "gemini-3.8-flash" CHAT_MODELS: Final = ( pytest.param( diff --git a/tests/e2e/llm_translation/test_custom_pricing_e2e.py b/tests/e2e/llm_translation/test_custom_pricing_e2e.py index 6d8cd24648e..63b8b60867a 100644 --- a/tests/e2e/llm_translation/test_custom_pricing_e2e.py +++ b/tests/e2e/llm_translation/test_custom_pricing_e2e.py @@ -5,7 +5,7 @@ teardown) instead of relying on a statically configured model, so the check is self-contained and never inherits pricing another suite or a stale config left on the shared proxy. custom-priced-flash sets input/output rates deliberately far above the canonical gemini price; the isolation sibling shares the same -gemini/gemini-2.5-flash backend but sets no override. Three behaviors are checked +gemini/gemini-3.8-flash backend but sets no override. Three behaviors are checked independently: - billing: a real call's logged cost breakdown charges input and output tokens at @@ -37,10 +37,8 @@ from models import ( pytestmark = pytest.mark.e2e -BACKEND_MODEL = "gemini/gemini-2.5-flash" +BACKEND_MODEL = "gemini/gemini-3.8-flash" GEMINI_API_KEY = "os.environ/GEMINI_API_KEY" -# Deliberately ~100x above canonical gemini-2.5-flash (input 3e-7 / output 2.5e-6) -# so an override that is ignored or under-applied bills at the base rate and fails. CUSTOM_INPUT_RATE = 5e-05 CUSTOM_OUTPUT_RATE = 1e-04 @@ -78,7 +76,7 @@ def _provision( input_cost_per_token: float | None, output_cost_per_token: float | None, ) -> str: - """Register a fresh gemini/gemini-2.5-flash deployment (deleted on teardown) and + """Register a fresh gemini/gemini-3.8-flash deployment (deleted on teardown) and return its model name. With the cost fields set the deployment carries a custom pricing override; with them None it is a plain sibling on the same backend. The marker keeps the name unique so concurrent runs on the shared proxy never @@ -251,7 +249,7 @@ class TestCustomPricing: assert sibling_entry is not None, f"{sibling} absent from /model/info" # custom-priced-flash overrides pricing; the sibling shares the same - # gemini/gemini-2.5-flash backend but sets no override, so it must keep its + # gemini/gemini-3.8-flash backend but sets no override, so it must keep its # own price. Equal rates mean the override leaked into the shared cost map. assert ( sibling_entry.model_info.input_cost_per_token diff --git a/tests/e2e/llm_translation/test_google_native_e2e.py b/tests/e2e/llm_translation/test_google_native_e2e.py index b18683bd4ad..65a341dbc2a 100644 --- a/tests/e2e/llm_translation/test_google_native_e2e.py +++ b/tests/e2e/llm_translation/test_google_native_e2e.py @@ -21,7 +21,7 @@ from pydantic import BaseModel pytestmark = pytest.mark.e2e -UPSTREAM_MODEL = "gemini/gemini-2.5-flash" +UPSTREAM_MODEL = "gemini/gemini-3.8-flash" class _GenerateContentPart(BaseModel): diff --git a/tests/e2e/llm_translation/test_image_edits_e2e.py b/tests/e2e/llm_translation/test_image_edits_e2e.py index 66b0ebfc538..917662b9e52 100644 --- a/tests/e2e/llm_translation/test_image_edits_e2e.py +++ b/tests/e2e/llm_translation/test_image_edits_e2e.py @@ -24,7 +24,7 @@ from sdk_clients import SdkClients pytestmark = pytest.mark.e2e -IMAGE_EDIT_BACKEND: Final = "openai/gpt-image-1" +IMAGE_EDIT_BACKEND: Final = "openai/gpt-image-2.5-sunburst" _TEST_PNG = base64.b64decode( "iVBORw0KGgoAAAANSUhEUgAAAEAAAABACAIAAAAlC+aJAAAAS0lEQVR42u3PMQ0AAAwDoPo3" diff --git a/tests/e2e/llm_translation/test_passthrough_e2e.py b/tests/e2e/llm_translation/test_passthrough_e2e.py index 52d1c10d829..3eaeef6bddd 100644 --- a/tests/e2e/llm_translation/test_passthrough_e2e.py +++ b/tests/e2e/llm_translation/test_passthrough_e2e.py @@ -5,7 +5,7 @@ and verifies the proxy still logged a costed SpendLogs row (call_type="pass_through_endpoint"), correlated by the id the caller was served: the x-litellm-call-id header on gemini, the `msg_...` message id on anthropic. -Covered: gemini ("gemini-2.5-flash") + anthropic ("claude-haiku-4-5"), streaming + +Covered: gemini ("gemini-3.8-flash") + anthropic ("claude-haiku-4-5"), streaming + non-streaming, plus native tool calls. See LLM_TRANSLATION_COVERAGE_MATRIX.md. A passthrough call returning non-2xx fails hard (never a skip); once it returns @@ -32,7 +32,7 @@ from passthrough_client import ( completed_responses_object, ) -GEMINI_MODEL: Final = "gemini-2.5-flash" +GEMINI_MODEL: Final = "gemini-3.8-flash" ANTHROPIC_PASSTHROUGH_MODEL: Final = "claude-haiku-4-5" EMBEDDING_MODEL = "text-embedding-3-small" REALTIME_MODEL = "gpt-realtime-2" @@ -291,7 +291,7 @@ class TestPassthroughModelAllowlist: result = client.anthropic_message(key, ANTHROPIC_PASSTHROUGH_MODEL, f"say hi {unique_marker()}") assert result.status_code == 403, ( - "a key restricted to gemini-2.5-flash must be denied a claude passthrough call, " + "a key restricted to gemini-3.8-flash must be denied a claude passthrough call, " f"got {result.status_code}: {result.body[:300]}" ) diff --git a/tests/e2e/llm_translation/test_responses_e2e.py b/tests/e2e/llm_translation/test_responses_e2e.py index a0f185c4077..6b299fad18c 100644 --- a/tests/e2e/llm_translation/test_responses_e2e.py +++ b/tests/e2e/llm_translation/test_responses_e2e.py @@ -57,7 +57,7 @@ OPENAI_VISION_BACKEND: Final = "openai/gpt-4o" ANTHROPIC_BACKEND: Final = "anthropic/claude-haiku-4-5" BEDROCK_CONVERSE_BACKEND: Final = "bedrock/us.anthropic.claude-haiku-4-5-20251001-v1:0" VERTEX_BACKEND: Final = "vertex_ai/gemini-2.5-flash" -GEMINI_BACKEND: Final = "gemini/gemini-2.5-flash" +GEMINI_BACKEND: Final = "gemini/gemini-3.8-flash" OPENAI_RESPONSES_BACKEND: Final = "openai/gpt-5.5" INSTRUCTIONS = "You are a helpful assistant" CAT_IMAGE_URL = "https://upload.wikimedia.org/wikipedia/commons/3/3a/Cat03.jpg" diff --git a/tests/e2e/quota_management/spend_tracking/test_key_attribution_e2e.py b/tests/e2e/quota_management/spend_tracking/test_key_attribution_e2e.py index e0c19ea1b6b..1d0878d073a 100644 --- a/tests/e2e/quota_management/spend_tracking/test_key_attribution_e2e.py +++ b/tests/e2e/quota_management/spend_tracking/test_key_attribution_e2e.py @@ -55,7 +55,7 @@ from spend_e2e_client import ( pytestmark = pytest.mark.e2e -CHAT_MODEL: Final = "gemini-2.5-flash" +CHAT_MODEL: Final = "gemini-3.8-flash" MESSAGES_MODEL: Final = "claude-haiku-4-5" RESPONSES_MODEL: Final = "openai-responses-codex" EMBED_MODEL: Final = "openai-text-embedding-3-small" diff --git a/tests/e2e/quota_management/spend_tracking/test_spend_tracking_e2e.py b/tests/e2e/quota_management/spend_tracking/test_spend_tracking_e2e.py index c7ef81f826a..d452525ec91 100644 --- a/tests/e2e/quota_management/spend_tracking/test_spend_tracking_e2e.py +++ b/tests/e2e/quota_management/spend_tracking/test_spend_tracking_e2e.py @@ -3,7 +3,7 @@ Run against a proxy started with the gateway config. Coverage rationale: SPEND_TRACKING_COVERAGE_MATRIX.md. -Model names are literals from that config: chat tests hit "gemini-2.5-flash", +Model names are literals from that config: chat tests hit "gemini-3.8-flash", embedding tests hit "openai-text-embedding-3-small". Every test: fresh scoped key (isolation) -> real provider call -> unwrap (hard @@ -37,7 +37,7 @@ from spend_reconciliation import BACKEND as TRAFFIC_BACKEND pytestmark = pytest.mark.e2e -GEMINI_MODEL = "gemini-2.5-flash" +GEMINI_MODEL = "gemini-3.8-flash" CLAUDE_MODEL = "claude-haiku-4-5" CODEX_MODEL = "openai-responses-codex" EMBEDDING_MODEL = "openai-text-embedding-3-small" @@ -780,7 +780,7 @@ def test_spend_calculate_returns_nonzero_cost(client: SpendClient) -> None: GEMINI_MODEL, "estimate the cost of this request" ) assert cost > 0, ( - "/spend/calculate returned 0 for gemini-2.5-flash; " + "/spend/calculate returned 0 for gemini-3.8-flash; " "cost map may be missing this model" ) diff --git a/tests/e2e/router/reliability_support.py b/tests/e2e/router/reliability_support.py index 03e585273c0..193a94cb581 100644 --- a/tests/e2e/router/reliability_support.py +++ b/tests/e2e/router/reliability_support.py @@ -57,11 +57,9 @@ CONTENT_POLICY_PROMPT = ( COOLDOWN_SECONDS = 30.0 REPLICA_PROPAGATION_SECONDS = 15.0 -# The smallest-context chat model OpenAI still serves (16385 tokens). A prompt -# past that limit comes back as a real `context_length_exceeded` 400, which is -# what litellm maps to ContextWindowExceededError. -SMALL_CONTEXT_MODEL = "openai/gpt-3.5-turbo" -SMALL_CONTEXT_LIMIT_TOKENS = 16385 +SMALL_CONTEXT_MODEL: Final = "openai/gpt-5.4-nano" +# model_prices_and_context_window.json, gpt-5.4-nano, verified 2026-11-02 +SMALL_CONTEXT_LIMIT_TOKENS: Final = 272000 def oversized_prompt(marker: str) -> str: @@ -107,7 +105,7 @@ def create_timeout_deployment(proxy: ProxyClient, name: str) -> str: @step(f"Add a deployment named {{name}} that calls the small-context model {SMALL_CONTEXT_MODEL}") def create_small_context_deployment(proxy: ProxyClient, name: str) -> str: - """Register a deployment on the smallest-context model OpenAI still serves, so an + """Register a deployment with a smaller context window than the backup, so an oversized prompt earns a real context-window refusal from the provider.""" return proxy.create_model(name, LiteLLMParamsBody(model=SMALL_CONTEXT_MODEL, api_key=REAL_KEY)) @@ -263,8 +261,7 @@ def spend_only_request_of(proxy: ProxyClient, spent_key: str) -> None: @step(f"Add a deployment named {{name}} that calls {SMALL_CONTEXT_MODEL} and takes all of its group's traffic") def create_always_picked_small_context_deployment(proxy: ProxyClient, name: str) -> str: - """The always-picked half of a retry pair on the smallest-context model OpenAI - still serves: it holds all of the model group's shuffle weight, so an oversized + """The always-picked half of a retry pair on the smaller-context model: it holds all of the model group's shuffle weight, so an oversized prompt opens on it and earns a real context-window refusal, which never benches a deployment, so only the retry itself can steer the request off it.""" return proxy.register_model(