he-yufeng · octo-patch · Jul 13, 2026
diff --git a/README.md b/README.md
@@ -69,8 +69,14 @@ Give it a model and a key and it goes. It speaks the OpenAI-compatible API by de
 |---|---|
 | OpenAI (default `gpt-5.5`) | `OPENAI_API_KEY=sk-...` |
 | DeepSeek | `OPENAI_API_KEY=sk-... OPENAI_BASE_URL=https://api.deepseek.com CORECODER_MODEL=deepseek-chat` |
+| MiniMax (global, OpenAI) | `OPENAI_API_KEY=sk-... OPENAI_BASE_URL=https://api.minimax.io/v1 CORECODER_MODEL=MiniMax-M3 CORECODER_MAX_CONTEXT=1000000` |
+| MiniMax (China, OpenAI) | `OPENAI_API_KEY=sk-... OPENAI_BASE_URL=https://api.minimaxi.com/v1 CORECODER_MODEL=MiniMax-M3 CORECODER_MAX_CONTEXT=1000000` |
+| MiniMax (global, Anthropic via LiteLLM) | `CORECODER_API_KEY=sk-... CORECODER_BASE_URL=https://api.minimax.io/anthropic CORECODER_MODEL=anthropic/MiniMax-M3 CORECODER_MAX_CONTEXT=1000000 CORECODER_PROVIDER=litellm` |
+| MiniMax (China, Anthropic via LiteLLM) | `CORECODER_API_KEY=sk-... CORECODER_BASE_URL=https://api.minimaxi.com/anthropic CORECODER_MODEL=anthropic/MiniMax-M3 CORECODER_MAX_CONTEXT=1000000 CORECODER_PROVIDER=litellm` |
 | Local Ollama | `OPENAI_API_KEY=ollama OPENAI_BASE_URL=http://localhost:11434/v1 CORECODER_MODEL=qwen2.5-coder` |
 
+MiniMax-M2.7 is available through the same four endpoints. Replace `MiniMax-M3` with `MiniMax-M2.7` and set `CORECODER_MAX_CONTEXT=204800`; keep the `anthropic/` prefix when using LiteLLM. The Anthropic-compatible examples require `pip install "corecoder[litellm]"`.
+
 Kimi, Qwen and the like are the same two variables; for providers that don't even offer an OpenAI-compatible endpoint, the optional LiteLLM backend (`pip install "corecoder[litellm]"`) routes to a hundred-plus of them. The third essay goes into this in detail. The key can be `export`ed directly or dropped into a `.env` at the project root, which is loaded on startup. Then:
 
 ```bash

diff --git a/README_CN.md b/README_CN.md
@@ -69,8 +69,14 @@ pip install -e .
 |---|---|
 | OpenAI（默认 `gpt-5.5`） | `OPENAI_API_KEY=sk-...` |
 | DeepSeek | `OPENAI_API_KEY=sk-... OPENAI_BASE_URL=https://api.deepseek.com CORECODER_MODEL=deepseek-chat` |
+| MiniMax (global, OpenAI) | `OPENAI_API_KEY=sk-... OPENAI_BASE_URL=https://api.minimax.io/v1 CORECODER_MODEL=MiniMax-M3 CORECODER_MAX_CONTEXT=1000000` |
+| MiniMax (China, OpenAI) | `OPENAI_API_KEY=sk-... OPENAI_BASE_URL=https://api.minimaxi.com/v1 CORECODER_MODEL=MiniMax-M3 CORECODER_MAX_CONTEXT=1000000` |
+| MiniMax (global, Anthropic via LiteLLM) | `CORECODER_API_KEY=sk-... CORECODER_BASE_URL=https://api.minimax.io/anthropic CORECODER_MODEL=anthropic/MiniMax-M3 CORECODER_MAX_CONTEXT=1000000 CORECODER_PROVIDER=litellm` |
+| MiniMax (China, Anthropic via LiteLLM) | `CORECODER_API_KEY=sk-... CORECODER_BASE_URL=https://api.minimaxi.com/anthropic CORECODER_MODEL=anthropic/MiniMax-M3 CORECODER_MAX_CONTEXT=1000000 CORECODER_PROVIDER=litellm` |
 | 本地 Ollama | `OPENAI_API_KEY=ollama OPENAI_BASE_URL=http://localhost:11434/v1 CORECODER_MODEL=qwen2.5-coder` |
 
+MiniMax-M2.7 is available through the same four endpoints. Replace `MiniMax-M3` with `MiniMax-M2.7` and set `CORECODER_MAX_CONTEXT=204800`; keep the `anthropic/` prefix when using LiteLLM. The Anthropic-compatible examples require `pip install "corecoder[litellm]"`.
+
 Kimi、Qwen 这些同样是改这两个变量；连 OpenAI 兼容接口都不给的 provider，装上可选的 LiteLLM 后端（`pip install "corecoder[litellm]"`）能路由一百多家。第三篇文章把这块讲得更细。key 可以直接 `export`，也可以在项目根目录扔个 `.env`，启动时自动加载。然后：
 
 ```bash

diff --git a/corecoder/llm.py b/corecoder/llm.py
@@ -49,9 +49,9 @@ def message(self) -> dict:
         return msg
 
 
-# pricing per million tokens: (input, output)
+# pricing per million tokens: (input, output, optional cache read, optional cache write)
 # sources: openai.com/api/pricing, api-docs.deepseek.com, platform.claude.com,
-#          platform.moonshot.ai, alibabacloud.com/help/en/model-studio
+#          platform.moonshot.ai, alibabacloud.com/help/en/model-studio, platform.minimax.io
 _PRICING = {
     # OpenAI - current flagships
     "gpt-5.5": (5, 30),
@@ -78,9 +78,55 @@ def message(self) -> dict:
     "qwen-max": (0.78, 3.9),
     # Moonshot Kimi
     "kimi-k2.5": (0.6, 3),
+    # MiniMax
+    "MiniMax-M2.7": (0.3, 1.2, 0.06, 0.375),
 }
 
 
+# tier entries: (maximum input tokens, input, output, cache read, cache write)
+_TIERED_PRICING = {
+    "MiniMax-M3": {
+        "standard": (
+            (512_000, 0.3, 1.2, 0.06, None),
+            (None, 0.6, 2.4, 0.12, None),
+        ),
+        "priority": (
+            (512_000, 0.45, 1.8, 0.09, None),
+            (None, 0.9, 3.6, 0.18, None),
+        ),
+    },
+}
+
+
+def _pricing_key(model: str) -> str:
+    """Strip a LiteLLM provider prefix before looking up model pricing."""
+    return model.rsplit("/", 1)[-1]
+
+
+def _cost_for_usage(
+    model: str,
+    prompt_tokens: int,
+    completion_tokens: int,
+    service_tier: str = "standard",
+) -> float | None:
+    key = _pricing_key(model)
+    pricing = _PRICING.get(key)
+    if pricing:
+        input_rate, output_rate = pricing[:2]
+    else:
+        model_tiers = _TIERED_PRICING.get(key)
+        if not model_tiers:
+            return None
+        tiers = model_tiers.get(service_tier, model_tiers["standard"])
+        _, input_rate, output_rate, _, _ = next(
+            tier for tier in tiers if tier[0] is None or prompt_tokens <= tier[0]
+        )
+    return (
+        prompt_tokens * input_rate / 1_000_000
+        + completion_tokens * output_rate / 1_000_000
+    )
+
+
 class LLM:
     def __init__(
         self,
@@ -94,19 +140,34 @@ def __init__(
         self.extra = kwargs  # temperature, max_tokens, etc.
         self.total_prompt_tokens = 0
         self.total_completion_tokens = 0
+        self._estimated_cost = _cost_for_usage(
+            model, 0, 0, kwargs.get("service_tier", "standard")
+        )
 
     @property
     def estimated_cost(self) -> float | None:
         """Rough cost estimate in USD. Returns None if model not in pricing table."""
-        pricing = _PRICING.get(self.model)
-        if not pricing:
-            return None
-        input_rate, output_rate = pricing
-        return (
-            self.total_prompt_tokens * input_rate / 1_000_000
-            + self.total_completion_tokens * output_rate / 1_000_000
+        if "_estimated_cost" in self.__dict__:
+            return self._estimated_cost
+        return _cost_for_usage(
+            self.model,
+            self.total_prompt_tokens,
+            self.total_completion_tokens,
+            getattr(self, "extra", {}).get("service_tier", "standard"),
         )
 
+    def _record_usage(self, prompt_tokens: int, completion_tokens: int):
+        self.total_prompt_tokens += prompt_tokens
+        self.total_completion_tokens += completion_tokens
+        request_cost = _cost_for_usage(
+            self.model,
+            prompt_tokens,
+            completion_tokens,
+            self.extra.get("service_tier", "standard"),
+        )
+        if request_cost is not None:
+            self._estimated_cost = (self._estimated_cost or 0.0) + request_cost
+
     def chat(
         self,
         messages: list[dict],
@@ -180,8 +241,7 @@ def chat(
                 args = {}
             parsed.append(ToolCall(id=raw["id"], name=raw["name"], arguments=args))
 
-        self.total_prompt_tokens += prompt_tok
-        self.total_completion_tokens += completion_tok
+        self._record_usage(prompt_tok, completion_tok)
 
         return LLMResponse(
             content="".join(content_parts),
@@ -236,6 +296,9 @@ def __init__(
         self.extra = kwargs
         self.total_prompt_tokens = 0
         self.total_completion_tokens = 0
+        self._estimated_cost = _cost_for_usage(
+            model, 0, 0, kwargs.get("service_tier", "standard")
+        )
 
     def chat(
         self,
@@ -300,8 +363,7 @@ def chat(
                 args = {}
             parsed.append(ToolCall(id=raw["id"], name=raw["name"], arguments=args))
 
-        self.total_prompt_tokens += prompt_tok
-        self.total_completion_tokens += completion_tok
+        self._record_usage(prompt_tok, completion_tok)
 
         return LLMResponse(
             content="".join(content_parts),

diff --git a/tests/test_core.py b/tests/test_core.py
@@ -1,5 +1,7 @@
 """Tests for core modules: config, context, session, imports."""
 
+import pytest
+
 from corecoder import Agent, LLM, Config, ALL_TOOLS, __version__
 from corecoder import session as session_module
 from corecoder.context import ContextManager, estimate_tokens
@@ -150,6 +152,61 @@ def test_cost_estimation_unknown_model():
     assert llm.estimated_cost is None
 
 
+def test_minimax_pricing_registry_preserves_official_tiers():
+    from corecoder.llm import _PRICING, _TIERED_PRICING
+
+    assert _PRICING["MiniMax-M2.7"] == (0.3, 1.2, 0.06, 0.375)
+    assert _TIERED_PRICING["MiniMax-M3"] == {
+        "standard": (
+            (512_000, 0.3, 1.2, 0.06, None),
+            (None, 0.6, 2.4, 0.12, None),
+        ),
+        "priority": (
+            (512_000, 0.45, 1.8, 0.09, None),
+            (None, 0.9, 3.6, 0.18, None),
+        ),
+    }
+
+
+@pytest.mark.parametrize(
+    ("model", "prompt_tokens", "service_tier", "expected"),
+    [
+        ("MiniMax-M3", 500_000, "standard", 0.27),
+        ("MiniMax-M3", 600_000, "standard", 0.60),
+        ("MiniMax-M3", 500_000, "priority", 0.405),
+        ("anthropic/MiniMax-M3", 600_000, "priority", 0.90),
+    ],
+)
+def test_minimax_m3_cost_estimation_uses_request_tier(
+    model, prompt_tokens, service_tier, expected
+):
+    from corecoder.llm import LLM
+
+    llm = LLM.__new__(LLM)
+    llm.model = model
+    llm.total_prompt_tokens = prompt_tokens
+    llm.total_completion_tokens = 100_000
+    llm.extra = {"service_tier": service_tier}
+    assert llm.estimated_cost == pytest.approx(expected)
+
+
+def test_minimax_m3_cost_tracks_each_request_separately():
+    from corecoder.llm import LLM
+
+    llm = LLM.__new__(LLM)
+    llm.model = "MiniMax-M3"
+    llm.extra = {}
+    llm.total_prompt_tokens = 0
+    llm.total_completion_tokens = 0
+    llm._estimated_cost = 0.0
+
+    llm._record_usage(400_000, 100_000)
+    llm._record_usage(400_000, 100_000)
+
+    assert llm.total_prompt_tokens == 800_000
+    assert llm.estimated_cost == pytest.approx(0.48)
+
+
 # --- Changed files tracking ---
 
 def test_edit_tracks_changed_files(tmp_path):