The US Harmonized Tariff Schedule is not a yearly file. USITC published the 2026 Basic edition on December 31, 2025, and Revision 23 took effect on November 10, 2026. That is 24 editions in a bit over ten months, one every 13.7 days on average (the closest two were 3 days apart, the longest gap was 42 days).
If you keep a copy of the schedule in a database, the obvious plan is to reload it when a new revision appears and compare the duty rates. I did that for three revisions in a row. Between Revision 21 and Revision 23, not a single code was added or removed, and not a single rate in chapters 1 to 97 changed. Eight rows did change. Each change was a few words of text, and two of them moved the start date of a 100% duty.
Getting the data
USITC has a small REST API. releaseList tells you which revision is current, and exportList gives you the whole schedule as JSON:
import gzip import json import urllib.request BASE = "https://hts.usitc.gov/reststop" def get_json(path): with urllib.request.urlopen(f"{BASE}/{path}", timeout=180) as r: return json.load(r) current = next(r for r in get_json("releaseList") if r.get("status") == "current") rows = get_json("exportList?from=0101&to=9999&format=JSON&styles=false") with gzip.open(f"{current['name']}.json.gz", "wt", encoding="utf-8") as f: json.dump(rows, f) Save every revision you download. The export gives you the current revision only, so if you want to compare against last month you have to have kept last month yourself. Gzipped, one revision is about 640 KB.
Some things to know about the rows before you compare them:
- Revision 23 has 35,805 rows, but only 29,861 have a code in
htsno. The other 5,944 are text rows such asHorses:orCattle:that group the lines under them. Key by code and they disappear, so check them separately. - Every row has a field called
addiitionalDuties(with the typo). It isnullin all 35,805 rows. The real field isadditionalDuties, set on 810 rows. If you pick fields by hand, pick the right one. - The order of rows was identical in all three revisions, but I would not rely on that.
A diff that does not assume which field matters
from collections import Counter def by_code(rows): return {r["htsno"]: r for r in rows if r["htsno"]} def diff(old_rows, new_rows): old, new = by_code(old_rows), by_code(new_rows) changed = [] for code in sorted(old.keys() & new.keys()): fields = [k for k in new[code] if old[code].get(k) != new[code][k]] if fields: changed.append((code, fields)) return { "added": sorted(new.keys() - old.keys()), "removed": sorted(old.keys() - new.keys()), "changed": changed, } def text_rows(rows): return Counter((r["indent"], r["description"]) for r in rows if not r["htsno"]) It compares every field, not only general (the Column 1 General rate), and it reports which fields changed. text_rows turns the rows without a code into a multiset, so an edited or missing heading text shows up even though it has no key.
Results:
| Comparison | Added | Removed | Changed | Text rows |
|---|---|---|---|---|
| Revision 21 → 22 | 0 | 0 | 1 | identical |
| Revision 22 → 23 | 0 | 0 | 7 | identical |
All 8 changes are in chapter 99, and in all 8 the only changed field is description.
What changed: a few words each
A plain string comparison tells you that a 600-character description changed, which does not help much. A word-level diff shows what changed:
import difflib def word_changes(old_text, new_text): a, b = old_text.split(), new_text.split() ops = difflib.SequenceMatcher(a=a, b=b).get_opcodes() return [(" ".join(a[i1:i2]), " ".join(b[j1:j2])) for op, i1, i2, j1, j2 in ops if op != "equal"] Revision 21 → 22 changed one heading, 9903.01.63 (an additional duty on products of China). The text lost one sentence:
("[Compiler's note: provision suspended. See 90 Fed. Reg. 50729.]", "") Its general column read The duty provided in the applicable subheading + 34% before and after. What that means legally is a question for the Federal Register notices, not for this post. But for a program the result is clear: the only sign of the change was a bracketed note that disappeared from a description.
Revision 22 → 23 changed seven headings, and the word diff is the same for every one of them:
[("November", "January"), ("2026,", "2027,")] -
9903.88.69and9903.88.70cover products of China with an exclusion granted by the U.S. Trade Representative. Their end date moved from November 9, 2026 to January 9, 2027. -
9903.91.12to9903.91.16cover intermodal chassis and their parts, and ship-to-shore gantry cranes. Two of them carry+ 100%. The text used to say "on or after November 10, 2026" and now says "on or after January 10, 2027".
November 10, 2026 is also the date Revision 23 took effect. So a 100% duty that was due to start that day was moved two months later in the same revision, and in the data all of that is two words in a description field.
Pull the dates out
Once you know the dates are in the text, you can extract them. None of the 26,246 coded rows in chapters 1 to 97 has a date in its description. 168 rows do, and 166 of them are in chapter 99, which has 3,112 coded rows in total.
import re from datetime import date, datetime MONTHS = "January|February|March|April|May|June|July|August|September|October|November|December" DATE = re.compile(r"\b(" + MONTHS + r") (\d\d?), (\d{4})") def dates_after(rows, today): for r in rows: for m in DATE.finditer(r["description"]): when = datetime.strptime(" ".join(m.groups()), "%B %d %Y").date() if when > today: yield r["htsno"], when, r["general"] Run on Revision 23 with today = date(2026, 11, 14), it finds this:
| Date | Heading | What the text says |
|---|---|---|
| 2026-11-30 | 9903.54.02 | end of a window for 300,000 t of extra lean beef trimmings |
| 2026-12-31 | 9903.54.01 | end of a window for 80,000 t of extra lean beef trimmings from Argentina |
| 2026-12-31 | 9819.11.12 | apparel from certain lesser developed countries, fabric of any origin, "through December 31, 2026, inclusive" |
| 2027-01-09 | 9903.88.69, 9903.88.70 | USTR exclusions for products of China end |
| 2027-01-10 | 9903.91.12 to 9903.91.16 | chassis and crane provisions start (100% on two of them) |
| 2027-04-18 | 9903.91.16 | entry deadline for cranes bought under contracts dated before April 17, 2025 |
That is a small list, short enough to put in a calendar. It is also a list that the rate column never shows.
The regex only catches dates written as "Month D, YYYY". 9819.11.12 starts "on the date announced in a Federal Register notice", and no pattern finds that. Use the list as a reminder to read the text, not as the legal answer.
What I would do differently from the start
- Diff every field, then filter. I started by comparing
generalonly. It reported "no changes" for both revisions, which was technically true and not useful. - Keep chapter 99 in scope. It is easy to drop as "not real products", but it is where the time-limited duties live. A
Freerate in chapter 85 does not tell you what chapter 99 adds for a given country of origin. - Archive everything. Two revisions are enough to compare, but only if you kept the older one.
- Expect quiet revisions. Revisions 22 and 23 changed 1 and 7 rows out of 35,805. Most of the work is noticing that something changed at all.
I run this diff for every new revision and publish the result at hs-code.net/changes, next to the classification guides. If you build your own, start with the word diff. It took the 7 changed descriptions in Revision 23 from 7 long paragraphs down to one line each.
