Round stats

GET /v1/{game}/matches/{id}/rounds returns two rows per round — one per team. Exactly one has won: true.

Join to a map with (match_id, map_number); rounds are numbered from 1 within a map and overtime keeps counting up, so round 25+ on a standard map is overtime rather than a restart.

economy_level is a signed tier, not a sentinel

Read it alongside equipment_value, enemy_equipment_value and money_spent — the raw dollar figures, with the opponent's value denormalised onto the same row so you can judge the buy matchup without a join.

A zero money_spent is usually a save, not a gap

About a quarter of rounds have money_spent: 0, and in most of them that is the truth: a team that saves buys nothing and carries the previous round's kit, so a zero spend beside a five-figure equipment_value is a save. Roughly 208,000 such rounds follow a round the team lost, which is exactly the shape of a save.

The field is genuinely absent on about 0.7% of matches, where every round of the match reads money_spent: 0. utility_value is absent the same way on about 2% of matches. Check the match, not the round: a scattered zero is a save, a match-wide zero is missing data.

Sides

team_side is lowercase ct or t. Sides swap at the half, so a team's side changes partway through a map — never assume a team's side from the map alone.

pistol_round flags round 1 and the second-half opener.

Counters

All integers, never null:

Two further fields, end_reason and round_duration_ms, describe the round rather than the team and can be null — see below.

How the round ended, and how long it took

Two fields describe the round itself rather than either team, so both rows of a round carry the same value: end_reason and round_duration_ms.

end_reason is an open vocabulary, not a fixed set. These are the values in the data today:

round_duration_ms is wall-clock milliseconds between round boundaries, which is not the same as the round's live length: it includes the freeze time before the round and whatever else happened in that window — a timeout, a technical pause, or a scheduled break. The median round is 112 s. About 13% exceed 155,000 ms, which is longer than a CS2 round can physically run (a 1:55 round plus a 0:40 bomb timer), and 1% run past five minutes.

Those long rounds are not scattered at random: the round after a break carries the break. Round 13 — the half-time swap — is over 155 s on 19.1% of maps against a 13% average, and round 25, the first overtime round, on 19.3%; rounds 12 and 14, either side of half-time, are the calmest on the map at 6.2% and 5.8%. The five-minute tail is headed by round 1 and round 13 (map start and half-time) and is ordinary rather than exceptional — 756 of the 2,152 matches with duration data carry at least one, usually exactly one. So a long round is far more often a break than a slow round. Read the field as a pace signal — comparing pistol rounds to gun rounds, or one team's rounds to another's, ideally excluding rounds 1, 13 and 25 — never as an exact duration.

Both fields are null on rounds recorded before they were captured. They are populated for every match ingested from 2026-09-19 and were backfilled across roughly the three months before that, so filter on end_reason != null rather than assuming a cutoff date.

rounds = get(f"/v1/cs2/matches/{match_id}/rounds")["data"]

# One row per round: the two team rows agree on both fields.
timeline = {
    (r["map_number"], r["round_number"]): (r["end_reason"], r["round_duration_ms"])
    for r in rounds
}

Rebuilding a scoreline

rounds = get(f"/v1/cs2/matches/{match_id}/rounds")["data"]

from collections import Counter
score = Counter(r["team_id"] for r in rounds if r["won"] and r["map_number"] == 1)

That should reconcile with map 1's score_a/score_b from map results. If it doesn't, you're probably summing across maps.