From 5ec49de92a058a020a52292711d1fd27f188903d Mon Sep 17 00:00:00 2001
From: xiaoyuyu6420 <93528429+xiaoyuyu6420@users.noreply.github.com>
Date: Mon, 31 Aug 2026 18:27:17 +0800
Subject: [PATCH] fix(provider): count Anthropic cache_creation tokens in input
 usage

Anthropic's input_tokens only covers tokens after the last cache breakpoint;
cache_creation_input_tokens (tokens written into a new cache entry) were
dropped entirely from TokenUsage, so usage.input/total undercounted requests
that create a fresh cache entry. Fold cache_creation_input_tokens into
input_other in both _extract_usage and the streaming _update_usage, matching
the OpenAI provider's accounting where input_other holds all non-cache-read
input (including cache writes).

Official formula: total_input = cache_read + cache_creation + input_tokens.
---
 .../core/provider/sources/anthropic_source.py | 10 ++-
 tests/test_anthropic_source.py                | 77 +++++++++++++++++++
 2 files changed, 85 insertions(+), 2 deletions(-)
 create mode 100644 tests/test_anthropic_source.py

diff --git a/astrbot/core/provider/sources/anthropic_source.py b/astrbot/core/provider/sources/anthropic_source.py
index ee72bc9d44..c861ded6ba 100644
--- a/astrbot/core/provider/sources/anthropic_source.py
+++ b/astrbot/core/provider/sources/anthropic_source.py
@@ -445,15 +445,21 @@ def _extract_usage(self, usage: Usage | None) -> TokenUsage:
         if usage is None:
             return TokenUsage()
         # https://docs.claude.com/en/docs/build-with-claude/prompt-caching#tracking-cache-performance
+        # Anthropic's input_tokens excludes cache served reads AND writes, so
+        # cache_creation_input_tokens must be added back into input_other to
+        # keep total input (and context-occupancy stats) accurate.
         return TokenUsage(
-            input_other=usage.input_tokens or 0,
+            input_other=(usage.input_tokens or 0)
+            + (usage.cache_creation_input_tokens or 0),
             input_cached=usage.cache_read_input_tokens or 0,
             output=usage.output_tokens or 0,
         )
 
     def _update_usage(self, token_usage: TokenUsage, usage: MessageDeltaUsage) -> None:
         if usage.input_tokens is not None:
-            token_usage.input_other = usage.input_tokens
+            token_usage.input_other = usage.input_tokens + (
+                usage.cache_creation_input_tokens or 0
+            )
         if usage.cache_read_input_tokens is not None:
             token_usage.input_cached = usage.cache_read_input_tokens
         if usage.output_tokens is not None:
diff --git a/tests/test_anthropic_source.py b/tests/test_anthropic_source.py
new file mode 100644
index 0000000000..4aa0f2efb3
--- /dev/null
+++ b/tests/test_anthropic_source.py
@@ -0,0 +1,77 @@
+from anthropic.types import MessageDeltaUsage, Usage
+
+from astrbot.core.provider.entities import TokenUsage
+from astrbot.core.provider.sources.anthropic_source import ProviderAnthropic
+
+
+def _provider() -> ProviderAnthropic:
+    return ProviderAnthropic.__new__(ProviderAnthropic)
+
+
+def test_anthropic_extract_usage_counts_cache_creation_input():
+    provider = _provider()
+
+    usage = provider._extract_usage(
+        Usage(
+            input_tokens=10,
+            cache_read_input_tokens=100,
+            cache_creation_input_tokens=50,
+            output_tokens=20,
+        )
+    )
+
+    # Anthropic's input_tokens excludes cache writes, so cache_creation
+    # must be folded into input_other to keep total input accurate.
+    assert usage.input_other == 60
+    assert usage.input_cached == 100
+    assert usage.input == 160
+    assert usage.output == 20
+
+
+def test_anthropic_extract_usage_without_cache_breakpoints():
+    provider = _provider()
+
+    usage = provider._extract_usage(Usage(input_tokens=30, output_tokens=10))
+
+    assert usage.input_other == 30
+    assert usage.input_cached == 0
+    assert usage.input == 30
+    assert usage.output == 10
+
+
+def test_anthropic_extract_usage_none_returns_empty():
+    provider = _provider()
+
+    assert provider._extract_usage(None) == TokenUsage()
+
+
+def test_anthropic_update_usage_counts_cache_creation_input():
+    provider = _provider()
+    token_usage = TokenUsage(input_other=5, input_cached=0, output=0)
+
+    provider._update_usage(
+        token_usage,
+        MessageDeltaUsage(
+            input_tokens=10,
+            cache_read_input_tokens=100,
+            cache_creation_input_tokens=50,
+            output_tokens=20,
+        ),
+    )
+
+    assert token_usage.input_other == 60
+    assert token_usage.input_cached == 100
+    assert token_usage.input == 160
+    assert token_usage.output == 20
+
+
+def test_anthropic_update_usage_omitted_fields_are_preserved():
+    provider = _provider()
+    token_usage = TokenUsage(input_other=5, input_cached=0, output=0)
+
+    # message_delta usage only carries output tokens in practice.
+    provider._update_usage(token_usage, MessageDeltaUsage(output_tokens=7))
+
+    assert token_usage.input_other == 5
+    assert token_usage.input_cached == 0
+    assert token_usage.output == 7