Prefer MiniMax Flash defaults and preserve reasoning transports

README.md

@@ -256,6 +256,53 @@ uncommon-route provider add google     AIza...
 uncommon-route serve
 ```
 
+MiniMax uses the global OpenAI-compatible endpoint by default. Pass the China
+endpoint explicitly when needed:
+
+```bash
+uncommon-route provider add minimax eyJ-...
+uncommon-route provider add minimax eyJ-... --url https://api.minimaxi.com/v1
+```
+
+| Region | OpenAI-compatible base URL | Anthropic-compatible base URL |
+|---|---|---|
+| Global | `https://api.minimax.io/v1` | `https://api.minimax.io/anthropic` |
+| China | `https://api.minimaxi.com/v1` | `https://api.minimaxi.com/anthropic` |
+
+Provider configuration stores the OpenAI-compatible base URL. Anthropic
+requests use the corresponding `/anthropic` base and append `/v1/messages`
+internally.
+
+New MiniMax connections offer **MiniMax-M3.1-Flash**, then **MiniMax-M3**.
+Saved custom and legacy model lists remain usable; older MiniMax text models
+are omitted from ordinary discovery suggestions. Image, speech, and video
+model entries are unaffected.
+
+Direct Flash requests use adaptive thinking with `max` effort by default;
+`low`, `medium`, `high`, `xhigh`, and `max` are accepted. The proxy removes
+unsupported `service_tier` and `reasoning_split=false` controls. Native
+Anthropic requests carry effort in `output_config.effort`; OpenAI-compatible
+requests use `reasoning_effort`. M3 supports adaptive/disabled thinking without
+an effort control. Tools, reasoning history, and image/video content are kept.
+
+Direct Flash advertises a 1,000,000-token context and 524,288-token maximum
+output. Flash costs $0.15 input / $0.015 cached input / $0.60 output per million
+tokens. M3 standard bands are $0.30/$0.06/$1.20 and $0.60/$0.12/$2.40; priority
+bands are $0.45/$0.09/$1.80 and $0.90/$0.18/$3.60 (input/cache/output).
+Until the M3 input-band boundary is verified, local estimates conservatively
+use the upper band. Flash has no priority tier or Coding Plan promise.
+
+Gateway catalogs remain authoritative for IDs, rates, context, and modalities.
+For example, OpenRouter currently publishes `minimax/minimax-m3.1-flash:free`,
+with zero catalog pricing and a 1,000,000-token context; requests use its
+`reasoning.effort` field and preserve `reasoning_details`. Commonstack models
+are taken from its live `/models` discovery; no unverified Flash slug is added.
+These gateway values are never replaced with direct MiniMax prices.
+
+Sources: [MiniMax OpenAI API](https://platform.minimax.io/docs/api-reference/text-openai-api.md),
+[MiniMax pricing](https://platform.minimax.io/docs/guides/pricing-paygo.md),
+[OpenRouter model catalog](https://openrouter.ai/api/v1/models).
+
 > UncommonRoute doesn't automatically read `OPENAI_API_KEY` or `ANTHROPIC_API_KEY`. Use `init`, a saved connection, or one of the manual setup paths above.
 
 ### Routing Modes

tests/test_anthropic_compat.py

@@ -1181,10 +1181,7 @@ def handler(request: httpx.Request) -> httpx.Response:
             )
 
             assert resp.status_code == 200
-            assert set(routed["available_models"]) == {
-                "deepseek/deepseek-v3.2",
-                "minimax/minimax-m2.1",
-            }
+            assert set(routed["available_models"]) == {"deepseek/deepseek-v3.2"}
             assert routed["capability_lane"] == CapabilityLane.GENERAL
             assert resp.headers["x-uncommon-route-requested-transport"] == "anthropic-messages"
             assert resp.headers["x-uncommon-route-transport"] == "openai-chat"
@@ -1675,10 +1672,7 @@ def handler(request: httpx.Request) -> httpx.Response:
             )
 
             assert resp.status_code == 200
-            assert set(routed["available_models"]) == {
-                "minimax/minimax-m2.7",
-                "deepseek/deepseek-v3.2",
-            }
+            assert set(routed["available_models"]) == {"deepseek/deepseek-v3.2"}
             assert captured["url"] == "https://api.commonstack.ai/v1/chat/completions"
             body = captured["body"]
             assert isinstance(body, dict)
@@ -1779,8 +1773,6 @@ def handler(request: httpx.Request) -> httpx.Response:
                 "deepseek/deepseek-v3.2",
                 "anthropic/claude-haiku-4-5",
                 "anthropic/claude-sonnet-4-6",
-                "minimax/minimax-m2.1",
-                "minimax/minimax-m2.7",
             }
             assert captured["url"] == "https://api.commonstack.ai/v1/chat/completions"
             body = captured["body"]
@@ -1931,9 +1923,18 @@ def handler(request: httpx.Request) -> httpx.Response:
         finally:
             asyncio.run(async_client.aclose())
 
+    @pytest.mark.parametrize(
+        ("base_url", "expected_url"),
+        [
+            ("https://api.minimax.io/v1", "https://api.minimax.io/anthropic/v1/messages"),
+            ("https://api.minimaxi.com/v1", "https://api.minimaxi.com/anthropic/v1/messages"),
+        ],
+    )
     def test_messages_use_minimax_anthropic_endpoint_for_direct_provider(
         self,
         monkeypatch: pytest.MonkeyPatch,
+        base_url: str,
+        expected_url: str,
     ) -> None:
         captured: dict[str, object] = {}
 
@@ -1947,7 +1948,7 @@ def handler(request: httpx.Request) -> httpx.Response:
                     "id": "msg_minimax_byok",
                     "type": "message",
                     "role": "assistant",
-                    "model": "minimax/minimax-m2.1",
+                    "model": "MiniMax-M3",
                     "content": [{"type": "text", "text": "ok"}],
                     "stop_reason": "end_turn",
                     "stop_sequence": None,
@@ -1964,32 +1965,44 @@ def handler(request: httpx.Request) -> httpx.Response:
                 "minimax": ProviderEntry(
                     name="minimax",
                     api_key="mm-key-123",
-                    base_url="https://api.minimax.io/v1",
-                    models=["minimax/minimax-m2.1"],
+                    base_url=base_url,
+                    models=["minimax/minimax-m3"],
                 ),
             })
             app = create_app(
-                upstream="https://api.commonstack.ai/v1",
+                upstream="https://primary.example/v1",
                 providers_config=providers,
             )
             client = TestClient(app, raise_server_exceptions=False)
             resp = client.post("/v1/messages", json={
-                "model": "minimax/minimax-m2.1",
+                "model": "minimax/minimax-m3",
                 "max_tokens": 32,
                 "messages": [{"role": "user", "content": "hello"}],
             })
 
             assert resp.status_code == 200
-            assert captured["url"] == "https://api.minimax.io/anthropic/v1/messages"
+            assert captured["url"] == expected_url
             headers = captured["headers"]
             assert isinstance(headers, dict)
             assert headers["x-api-key"] == "mm-key-123"
+            body = captured["body"]
+            assert isinstance(body, dict)
+            assert body["model"] == "MiniMax-M3"
         finally:
             asyncio.run(async_client.aclose())
 
-    def test_chat_completions_keep_openai_transport_for_minimax_models(
+    @pytest.mark.parametrize(
+        ("base_url", "expected_url"),
+        [
+            ("https://api.minimax.io/v1", "https://api.minimax.io/v1/chat/completions"),
+            ("https://api.minimaxi.com/v1", "https://api.minimaxi.com/v1/chat/completions"),
+        ],
+    )
+    def test_chat_completions_use_minimax_openai_endpoint_for_direct_provider(
         self,
         monkeypatch: pytest.MonkeyPatch,
+        base_url: str,
+        expected_url: str,
     ) -> None:
         captured: dict[str, object] = {}
 
@@ -2003,7 +2016,7 @@ def handler(request: httpx.Request) -> httpx.Response:
                     "id": "chatcmpl_minimax",
                     "object": "chat.completion",
                     "created": 1,
-                    "model": "minimax/minimax-m2.1",
+                    "model": "MiniMax-M3",
                     "choices": [{
                         "index": 0,
                         "message": {"role": "assistant", "content": "done"},
@@ -2016,13 +2029,23 @@ def handler(request: httpx.Request) -> httpx.Response:
 
         async_client = httpx.AsyncClient(transport=httpx.MockTransport(handler))
         monkeypatch.setattr("uncommon_route.proxy._get_client", lambda: async_client)
-        monkeypatch.setenv("UNCOMMON_ROUTE_API_KEY", "env-key-123")
 
         try:
-            app = create_app(upstream="https://api.commonstack.ai/v1")
+            providers = ProvidersConfig(providers={
+                "minimax": ProviderEntry(
+                    name="minimax",
+                    api_key="mm-key-123",
+                    base_url=base_url,
+                    models=["minimax/minimax-m3"],
+                ),
+            })
+            app = create_app(
+                upstream="https://primary.example/v1",
+                providers_config=providers,
+            )
             client = TestClient(app, raise_server_exceptions=False)
             resp = client.post("/v1/chat/completions", json={
-                "model": "minimax/minimax-m2.1",
+                "model": "minimax/minimax-m3",
                 "max_tokens": 32,
                 "messages": [{"role": "user", "content": "hello"}],
             })
@@ -2031,10 +2054,13 @@ def handler(request: httpx.Request) -> httpx.Response:
             assert resp.headers["x-uncommon-route-requested-transport"] == "openai-chat"
             assert resp.headers["x-uncommon-route-transport"] == "openai-chat"
             assert resp.headers["x-uncommon-route-transport-source"] == "ingress-policy"
-            assert captured["url"] == "https://api.commonstack.ai/v1/chat/completions"
+            assert captured["url"] == expected_url
+            headers = captured["headers"]
+            assert isinstance(headers, dict)
+            assert headers["authorization"] == "Bearer mm-key-123"
             body = captured["body"]
             assert isinstance(body, dict)
-            assert body["model"] == "minimax/minimax-m2.1"
+            assert body["model"] == "MiniMax-M3"
         finally:
             asyncio.run(async_client.aclose())
 

tests/test_cache_support.py

@@ -2,8 +2,15 @@
 
 import json
 
-from uncommon_route.cache_support import apply_anthropic_cache_breakpoints, parse_stream_usage_metrics
-from uncommon_route.router.types import ModelPricing
+import pytest
+
+from uncommon_route.cache_support import (
+    apply_anthropic_cache_breakpoints,
+    estimate_usage_cost,
+    parse_stream_usage_metrics,
+    parse_usage_metrics,
+)
+from uncommon_route.router.types import ModelPricing, ModelPricingTier
 
 
 def test_anthropic_cache_breakpoints_do_not_upgrade_after_existing_5m() -> None:
@@ -88,3 +95,74 @@ def test_stream_usage_falls_back_to_reasoning_content_tokens_when_usage_missing(
     assert usage.output_tokens > 0
     assert usage.total_tokens == usage.output_tokens
     assert usage.actual_cost is not None
+
+
+def test_usage_cost_applies_long_context_pricing_tier() -> None:
+    pricing = ModelPricing(
+        0.30,
+        1.20,
+        pricing_tiers=(
+            ModelPricingTier("standard", 0.30, 1.20, input_tokens_lte=512_000),
+            ModelPricingTier("standard", 0.60, 2.40, input_tokens_gt=512_000),
+        ),
+    )
+
+    cost = estimate_usage_cost(
+        input_tokens_uncached=600_000,
+        output_tokens=100_000,
+        cache_read_input_tokens=0,
+        cache_write_input_tokens=0,
+        pricing=pricing,
+    )
+
+    assert cost == 0.60
+
+
+def test_usage_cost_applies_priority_pricing_tier() -> None:
+    pricing = ModelPricing(
+        0.30,
+        1.20,
+        pricing_tiers=(
+            ModelPricingTier("standard", 0.60, 2.40, input_tokens_gt=512_000),
+            ModelPricingTier("priority", 0.90, 3.60, input_tokens_gt=512_000),
+        ),
+    )
+
+    cost = estimate_usage_cost(
+        input_tokens_uncached=600_000,
+        output_tokens=100_000,
+        cache_read_input_tokens=0,
+        cache_write_input_tokens=0,
+        pricing=pricing,
+        service_tier="priority",
+    )
+
+    assert cost == pytest.approx(0.90)
+
+
+def test_usage_metrics_prefers_response_service_tier() -> None:
+    pricing = ModelPricing(
+        0.30,
+        1.20,
+        pricing_tiers=(
+            ModelPricingTier("standard", 0.60, 2.40, input_tokens_gt=512_000),
+            ModelPricingTier("priority", 0.90, 3.60, input_tokens_gt=512_000),
+        ),
+    )
+    content = json.dumps({
+        "service_tier": "priority",
+        "usage": {
+            "prompt_tokens": 600_000,
+            "completion_tokens": 100_000,
+            "total_tokens": 700_000,
+        },
+    }).encode()
+
+    usage = parse_usage_metrics(
+        content,
+        "test/model",
+        {"test/model": pricing},
+    )
+
+    assert usage is not None
+    assert usage.actual_cost == pytest.approx(0.90)