Summary

Introduces a unified cross-provider service_tier on ModelSettings (values: auto / default / flex / priority) and wires it through OpenAI, Bedrock, and the Google Gemini API. Supersedes / rebuilds #4926 on top of current main, picking up the design direction from @DouweM's review there.

• New top-level field: ModelSettings.service_tier with a ServiceTier TypeAlias. 'auto' is an explicit "omit" value (useful to override an inherited setting).
• OpenAI: pass-through. openai_service_tier still wins when set.
• Bedrock: pass-through for default / flex / priority; auto omits. bedrock_service_tier remains the only way to request reserved.
• Google (Gemini API / GLA): maps default → 'standard', flex → 'flex', priority → 'priority', auto → omit. Response's x-gemini-service-tier header is surfaced as provider_details['service_tier'] on both non-streaming and streaming responses.
• Google (Vertex AI): intentionally ignores top-level service_tier. Vertex's PT / Flex PayGo routing has different failure modes (e.g. pt_only → 429 without PT quota) and no direct priority equivalent, so there is no silent cross-map. Use google_vertex_service_tier explicitly on Vertex.

Google settings refactor

• Split GoogleServiceTier into GoogleGLAServiceTier ('standard' | 'flex' | 'priority') and GoogleVertexServiceTier (existing PT / Flex values).
• Add per-provider fields: google_gla_service_tier and google_vertex_service_tier on GoogleModelSettings.
• google_service_tier is kept as a deprecated alias for google_vertex_service_tier; setting it emits a DeprecationWarning when consulted on Vertex. GoogleServiceTier (the type alias) is kept so existing imports continue to type-check.
• Previously, Vertex PT / Flex headers were sent on every Google request (including GLA, where they are silently ignored server-side). They are now only sent when self.system == 'google-vertex'.

Side fixes

• Cerebras: added openai_service_tier to openai_unsupported_model_settings so it is correctly filtered alongside the pre-existing service_tier entry. (Latent bug discovered while unifying the setting.)

Dependency

• google-genai bumped >=1.66.0 → >=1.70.0 (required for the ServiceTier enum used in GenerateContentConfigDict). The file_search_store additions in the test_google_model_file_search_tool* snapshots are a cosmetic downstream effect of the bump, not a behaviour change.

Test plan

• uv run ruff format pydantic_ai_slim tests — clean
• uv run ruff check pydantic_ai_slim tests — clean
• uv run pyright on all touched source + test files — 0 errors
• uv run pytest tests/models/test_google.py tests/models/test_openai.py tests/models/test_bedrock.py tests/providers/test_cerebras.py tests/test_settings.py — 488 passed, 21 skipped (API-key gated)
• Wider sweep: uv run pytest tests/models tests/providers tests/test_agent.py tests/test_streaming.py — 2271 passed, 339 skipped, 3 xfailed (all pre-existing)
• Live smoke test against Gemini API (gemini-2.5-flash):
  • service_tier='auto' → server returns 'standard' (identical to "nothing set")
  • service_tier='flex' → server returns 'flex'
  • service_tier='flex' + google_gla_service_tier='priority' → server returns 'priority' (per-provider wins)
  • google_service_tier='pt_only' on GLA → silently ignored, no crash, server returns 'standard'

Notes / follow-ups

• The gateway provider for 'gemini' currently routes through GoogleProvider(vertexai=True), whose .name is 'google-vertex'. With the Vertex refusal, the gateway's Gemini route will ignore top-level service_tier. This is pre-existing behaviour of the gateway mapping and not addressed here; flagging for follow-up if desired.

Replaces the work in #4926 — thanks @markmcd for the initial implementation and design exploration; the Flex PayGo test surface in tests/models/test_google.py is based on tests from that PR.