readme.md

右上のアイコン->設定->データコントロール->エクスポートする

conversations.jsonを保存する

↑のスクリプトを実行する

chat-gpt-conversation-if-using-api.py

#!/usr/bin/env python
"""
conversations.json → 日次 API コスト試算
  * モデル別 token
  * Web Search クエリをカウントして課金
"""
import json, sys, pathlib, re, collections, zoneinfo
from datetime import datetime
import tiktoken, pandas as pd

# ── 単価 ───────────────────────────────────────────────────────────
PRICE = {
    "gpt-4o":      dict(inp=5.00,  out=20.00),
    "gpt-4o-mini": dict(inp=0.60,  out=2.40),
    "o3":          dict(inp=10.00, out=40.00),
    "o4-mini":     dict(inp=1.10,  out=4.40),
}
SEARCH_RATE = {
    "gpt-4o":      0.030,      # 1 query ≈ $0.03
    "o3":          0.030,
    "gpt-4o-mini": 0.0275,
    "o4-mini":     0.0275,
}
# ──────────────────────────────────────────────────────────────────

JST = zoneinfo.ZoneInfo("Asia/Tokyo")
enc = tiktoken.get_encoding("cl100k_base")

def daykey(ts):               # epoch → "YYYY-MM-DD"
    return datetime.fromtimestamp(ts, JST).strftime("%Y-%m-%d")

def get_model_slug(msg):
    md = (msg or {}).get("metadata", {})
    return md.get("model_slug") or md.get("default_model_slug")

# --- 検索クエリ検出 ------------------------------------------------
def count_search_queries(raw: str) -> int:
    if '"search_query"' not in raw:
        return 0
    return raw.count('"q"')            # "q": の個数を単純に数える

def find_parent_model(node_id, mapping):
    cur = node_id
    while cur:
        par = mapping.get(cur, {}).get("parent")
        if not par:
            break
        pmsg = mapping.get(par, {}).get("message") or {}
        if pmsg.get("author", {}).get("role") == "assistant":
            slug = get_model_slug(pmsg)
            if slug:
                return slug
        cur = par
    return "gpt-4o"
# ------------------------------------------------------------------


data = json.loads(pathlib.Path("conversations.json").read_text())

tok_tot  = collections.defaultdict(lambda: collections.defaultdict(lambda: [0, 0]))
web_cost = collections.defaultdict(float)     # day → $

for conv in data:
    mapping = conv.get("mapping", {})
    for nid, node in mapping.items():
        msg  = node.get("message") or {}
        role = msg.get("author", {}).get("role")
        if role not in ("user", "assistant"):
            continue

        ts = msg.get("create_time") or conv.get("create_time")
        if ts is None:
            continue
        dkey = daykey(ts)

        # 本文テキスト
        cont = msg.get("content") or {}
        raw  = "\n".join(cont.get("parts", [])) if cont.get("content_type") == "text" else cont.get("text", "")

        # token
        tokens = len(enc.encode(raw))

        # モデル
        slug = get_model_slug(msg)
        if role == "user" and not slug:
            slug = find_parent_model(nid, mapping)
        slug = slug if slug in PRICE else "gpt-4o"

        io = 0 if role == "user" else 1
        tok_tot[dkey][slug][io] += tokens

        # assistant メッセージに search_query があれば課金
        if role == "assistant":
            q = count_search_queries(raw)
            if q:
                web_cost[dkey] += q * SEARCH_RATE.get(slug, 0.03)

# DataFrame 化
rows = []
for d, models in tok_tot.items():
    for slug, (in_tok, out_tok) in models.items():
        p = PRICE[slug]
        cost = in_tok/1e6*p["inp"] + out_tok/1e6*p["out"]
        rows.append(dict(date=d, model=slug,
                         input_tok=in_tok, output_tok=out_tok, cost=cost))

for d, c in web_cost.items():
    rows.append(dict(date=d, model="web_search",
                     input_tok=0, output_tok=0, cost=c))

df = pd.DataFrame(rows).sort_values(["date", "model"])

# 日合計
day_sum = df.groupby("date")["cost"].sum().reset_index()

print("\n### 日次コスト (USD)")
print(day_sum.to_string(index=False, formatters={"cost": "${:,.4f}".format}))

print("\n### 詳細（モデル別）")
print(df.to_string(index=False, formatters={
        "input_tok": "{:,}".format,
        "output_tok": "{:,}".format,
        "cost": "${:,.4f}".format}))