Which Claude model did Claude Code actually serve you?

You pinned a model, but you suspect a different (cheaper) tier is quietly being served. Stop guessing from the UI — your own session logs record, per turn, the exact model ID that was served. Two copy-paste commands turn that into a timestamped timeline and a per-model tally. Everything runs locally against files already on your disk.

Where the ground truth lives

Claude Code writes one JSON object per line into a session log:

~/.claude/projects/<project>/<session-id>.jsonl

For every assistant turn, the line has type":"assistant", and inside it message.model is the model ID that was actually served for that turn. The same line carries a top-level timestamp (UTC, ISO 8601). So message.model — not the UI, not the status line — is the canonical record of what you got.

1 · Per-turn timeline of the served model (with timestamps)

Across every session, list each assistant turn's served model next to its timestamp, sorted in time order. Paste as-is:

for f in ~/.claude/projects/*/*.jsonl; do
  python3 - "$f" <<'PY'
import sys, json
for line in open(sys.argv[1]):
    try: d = json.loads(line)
    except: continue
    if d.get("type") == "assistant":
        m = d.get("message", {}).get("model")
        if m and m != "<synthetic>":
            print(d.get("timestamp",""), m)
PY
done | sort

Each line is one served turn: 2026-06-07T01:23:45.678Z  claude-opus-4-8. If a stretch of turns shows a different (lower) tier than the one you pinned, that's your evidence — with the exact UTC time you can correlate against your activity.

2 · Per-model turn tally (the strong number for a report)

Collapse all of that into a count of turns served by each model — the single figure that makes a report concrete:

cat ~/.claude/projects/*/*.jsonl | python3 -c '
import sys, json, collections
c = collections.Counter()
for line in sys.stdin:
    try: d = json.loads(line)
    except: continue
    if d.get("type")=="assistant":
        m=d.get("message",{}).get("model")
        if m and m!="<synthetic>": c[m]+=1
for m,n in c.most_common(): print(f"{n:7d}  {m}")
'

Output is one row per served model, most-served first — for example a column of turn counts beside claude-opus-4-8, claude-sonnet-4-6, and so on.

3 · Recent-window breakdown (catch a stale "favorite model")

An all-time tally can be the opposite of what you're getting right now — which is also why a stats "favorite model" that ignores recency misleads (cf. anthropics/claude-code #65899). On one real machine the two windows fully inverted:

So tally only the recent window. Change DAYS to taste:

python3 - <<'PY'
import json, glob, os, collections, datetime
DAYS = 7
cut = (datetime.date.today() - datetime.timedelta(days=DAYS)).isoformat()
c = collections.Counter()
for f in glob.glob(os.path.expanduser("~/.claude/projects/*/*.jsonl")):
    for line in open(f):
        try: d = json.loads(line)
        except: continue
        if d.get("type") == "assistant":
            m = d.get("message", {}).get("model"); t = d.get("timestamp", "")
            if m and m != "<synthetic>" and t[:10] >= cut:
                c[m] += 1
tot = sum(c.values()) or 1
for m, n in c.most_common():
    print(f"{n*100//tot:3d}%  {n:6d}  {m}")
PY

Each row is the recent-window share per model — e.g. 73%  24773  claude-opus-4-8 — so a drifted "favorite" stands out at a glance.

Reading the model IDs

The IDs are explicit, so you can see when a turn crossed into a different tier. Real current examples:

What's actually worth reporting

Not every difference is equal. Sort what you find by how much it matters:

• Opus-to-Opus (e.g. 4-7 ↔ 4-8) is a weak signal. Both are top tier; a swap between them rarely changes output quality enough to matter, and can happen for ordinary reasons.
• A silent drop to a non-Opus tier (e.g. claude-sonnet-4-6) while you pinned Opus is the sharp evidence. A cross-tier downgrade is what produces a felt quality difference, so that's the case worth raising.

And turn it into something reproducible rather than an anecdote:

• Include the timestamped table from command 1 for the window in question — exact UTC times, not "yesterday afternoon."
• If you have a Feedback ID for the affected turn, attach it. A timestamped served-model table plus a Feedback ID is a report a maintainer can actually act on.

This page describes a local verification method against your own session logs; it is not account or legal advice and makes no claim about Anthropic's routing intent. For model and Usage Policy questions, confirm against Anthropic's own documentation and support. Issues referenced for context: anthropics/claude-code #65889 (silent routing under a pinned model), #65899 (stats "favorite model" ignores recency).