From cc37cdd54bc402329348eafbc05b93d796b63afe Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 22 May 2026 17:41:50 +0000
Subject: [PATCH] Studio: read OpenAI cached tokens from prompt_tokens_details
 too

Codex flagged that the chat-style OpenAI envelope Studio re-emits
via _build_usage_chunk surfaces cached prompt tokens under
prompt_tokens_details.cached_tokens, not input_tokens_details. The
OpenAI branch only checked input_tokens_details, so a cache-heavy
chat-style turn billed every cached token at the full input rate
instead of the 0.1x cache_read discount.

Walk both keys when discovering the cached count. New regression
test pins that the two envelopes price identically for a turn with
80k of 100k tokens cached.
---
 studio/backend/core/inference/pricing.py | 16 ++++++++--
 studio/backend/tests/test_pricing.py     | 37 ++++++++++++++++++++++++
 2 files changed, 50 insertions(+), 3 deletions(-)

diff --git a/studio/backend/core/inference/pricing.py b/studio/backend/core/inference/pricing.py
index d9033e73ea..9f344059fd 100644
--- a/studio/backend/core/inference/pricing.py
+++ b/studio/backend/core/inference/pricing.py
@@ -230,9 +230,19 @@ def calculate_cost(
     else:
         output_tokens = int(usage.get("completion_tokens") or 0)
     if provider == "openai":
-        details = usage.get("input_tokens_details") or {}
-        if isinstance(details, dict):
-            cache_read = max(cache_read, int(details.get("cached_tokens") or 0))
+        # Cached prompt tokens live on different sub-objects depending on
+        # which envelope landed here. Raw OpenAI Responses usage uses
+        # ``input_tokens_details.cached_tokens``; the OpenAI Chat
+        # Completions envelope Studio re-emits via ``_build_usage_chunk``
+        # uses ``prompt_tokens_details.cached_tokens``. Check both so
+        # cache-heavy chat-style turns get the 0.1x cache_read_mult
+        # discount instead of full input pricing.
+        for key in ("input_tokens_details", "prompt_tokens_details"):
+            details = usage.get(key) or {}
+            if isinstance(details, dict):
+                cache_read = max(
+                    cache_read, int(details.get("cached_tokens") or 0)
+                )
         # OpenAI: cache_read already counted inside input_tokens.
         out["billable_input_tokens"] = input_tokens + cache_creation
     else:
diff --git a/studio/backend/tests/test_pricing.py b/studio/backend/tests/test_pricing.py
index 662866187f..f5f25c1aa3 100644
--- a/studio/backend/tests/test_pricing.py
+++ b/studio/backend/tests/test_pricing.py
@@ -554,6 +554,43 @@ def test_openai_chat_style_prompt_tokens_keeps_cache_read_semantics():
     assert _isclose(chat["total_usd"], raw["total_usd"]), (chat, raw)
 
 
+def test_openai_chat_style_envelope_reads_cache_from_prompt_tokens_details():
+    """``_build_usage_chunk`` emits cached tokens under
+    ``prompt_tokens_details.cached_tokens`` (chat-style), not
+    ``input_tokens_details``. The cost calculator must honour
+    both so cache-heavy chat-style turns get the discounted rate."""
+    base = OPENAI_PRICING["gpt-5.5"]["input_per_mtok"]
+    raw = calculate_cost(
+        "openai",
+        "gpt-5.5",
+        {
+            "input_tokens": 100_000,
+            "input_tokens_details": {"cached_tokens": 80_000},
+            "output_tokens": 0,
+        },
+    )
+    chat_style = calculate_cost(
+        "openai",
+        "gpt-5.5",
+        {
+            "prompt_tokens": 100_000,
+            "prompt_tokens_details": {"cached_tokens": 80_000},
+            "completion_tokens": 0,
+        },
+    )
+    # Both envelopes must price identically.
+    assert _isclose(chat_style["input_usd"], raw["input_usd"]), (chat_style, raw)
+    assert _isclose(chat_style["cache_read_usd"], raw["cache_read_usd"]), (
+        chat_style,
+        raw,
+    )
+    # And the discount is real: 80k charged at 0.1x base, 20k at full.
+    assert _isclose(
+        chat_style["cache_read_usd"],
+        80_000 / 1_000_000.0 * base * OPENAI_CACHE_READ_MULT,
+    )
+
+
 def test_explicit_zero_output_tokens_wins_over_stale_completion_tokens():
     """``output_tokens: 0`` must beat a stale ``completion_tokens: 50``.