This prints, per model family, how many turns you ran in the last seven days, the share of input that came from cache, the list-price split, the median context per turn and the first-turn size. For Haiku it also counts turns over 100K. It only reads your transcripts. It writes nothing and sends nothing.
terminal · python 3python3 - <<'EOF'
import json, glob, os, time, statistics as st
P = {"fable": (8, 10, 16, .20, 40), "opus": (4, 5, 8, .20, 20), "sonnet": (2, 2.5, 4, .10, 10),
"haiku": (.10, .125, .20, .01, .50), "haiku>100k": (.50, .625, 1, .05, 2.50)}
cut, seen, S = time.time() - 7 * 86400, set(), {}
for f in glob.glob(os.path.expanduser("~/.claude/projects/**/*.jsonl"), recursive=True):
if os.path.getmtime(f) < cut: continue
first = True
for line in open(f, errors="ignore"):
try: d = json.loads(line)
except ValueError: continue
m = d.get("message") or {}
if d.get("type") != "assistant" or m.get("id") in seen: continue
fam = next((k for k in ("fable", "opus", "sonnet", "haiku") if k in m.get("model", "")), None)
if not fam: continue
seen.add(m["id"]); u = m.get("usage") or {}; cc = u.get("cache_creation") or {}
w1 = cc.get("ephemeral_1h_input_tokens", 0); w5 = (u.get("cache_creation_input_tokens") or 0) - w1
i, r, o = u.get("input_tokens") or 0, u.get("cache_read_input_tokens") or 0, u.get("output_tokens") or 0
prompt = i + w5 + w1 + r
big = fam == "haiku" and prompt > 100_000; p = P["haiku>100k" if big else fam]
s = S.setdefault(fam, {"turns": 0, "read": 0, "prompt": 0, "big": 0, "first": [], "all": [], "$": [0, 0, 0, 0]})
s["turns"] += 1; s["read"] += r; s["prompt"] += prompt; s["big"] += big
s["all"].append(prompt)
if first: s["first"].append(prompt); first = False
for k, v in enumerate((i * p[0], w5 * p[1] + w1 * p[2], r * p[3], o * p[4])): s["$"][k] += v / 1e6
for fam, s in S.items():
tot = sum(s["$"]) or 1
print(f"{fam:7} {s['turns']:6} turns | cache reads {s['read'] / max(s['prompt'], 1):.0%} of input tokens"
f" | list price ${tot:,.2f}: writes {s['$'][1] / tot:.0%}, reads {s['$'][2] / tot:.0%}, output {s['$'][3] / tot:.0%}"
f" | context per turn median {st.median(s['all']):,.0f}"
f" | first turn median {st.median(s['first']) if s['first'] else 0:,.0f} tokens"
+ (f" | {s['big']} turns over 100K" if fam == "haiku" else ""))
EOF
my outputhaiku 222 turns | cache reads 55% of input tokens | list price $4.18: writes 53%, reads 11%, output 36% | context per turn median 71,300 | first turn median 67,584 tokens | 55 turns over 100K
sonnet 5135 turns | cache reads 98% of input tokens | list price $197.82: writes 42%, reads 43%, output 15% | context per turn median 166,833 | first turn median 77,014 tokens
opus 29513 turns | cache reads 98% of input tokens | list price $2,207.36: writes 39%, reads 47%, output 15% | context per turn median 168,840 | first turn median 82,331 tokens
If your Opus line looks like mine, Opus at medium is probably the right default, as Anthropic says. The cheaper move is a shorter start, not a smaller model.
The next thing that could change this is Haiku's pricing. A Haiku without the 100K line would make the ladder in the guide work for Claude Code helpers too. Until then, the line decides more than the model name does.
