Where the records come from
Every grade is worked out from ANAC's public records of each scheduled flight in Brazil (Voo Regular Ativo, VRA): when it was due, when it actually left and arrived, and whether it was cancelled. The grades you see use records from September 2025 – August 2026.
What a grade means
A grade sums up how a flight or connection has held up on comparable days in the past. For a connection, that means how often the first flight landed too late to make the second, and how long you'd wait for the next way on. For a single flight, it means how often it arrived an hour or more late or was cancelled. A grade is an estimate from history, never a promise about your day.
On time
A flight counts as on time when it arrived within 15 minutes of its scheduled time. On-time shares are out of the flights that operated; cancelled flights are shown separately.
“1 day in 10”
“1 day in 10 it was more than 52 minutes late” means that on the worst tenth of the days it flew, it arrived more than 52 minutes after schedule.
The time from landing to the domestic gate (international arrivals)
We estimate it by adding up each step:
- leaving the aircraft
- passport control
- collecting bags
- customs
- checking bags in again
- the walk to the domestic gate
Each step's time comes from the dated sources listed below, and the card shows our typical total. These are estimates. We don't measure passengers' actual times.
The time you need at a domestic connection adds the time to leave the aircraft, the walk between gates (from OpenStreetMap distances) and boarding closing 15 minutes before departure (modelled).
How we tested the grades
Before looking at any result, we wrote down how the grades would be judged, then tested them once on records the model had never seen: September 2025 to May 2026. The model was built only on January 2024 to August 2025. The grades on GateCast are built the same way from the newer records above; this test did not check them. The full test plan, with every change and its date, is available on request (contact us).
Four kinds of trip were tested separately:
- connections within Brazil
- international flights to and from Brazil
- domestic flights within Brazil
- international arrivals connecting to a domestic flight at GRU and VCP
For each, the test asked:
- Do worse grades come with more missed connections (or more late and cancelled flights), grade by grade?
- Do the model's own missed-connection estimates match what happened, within 3 points on average? (Connections only.)
- Do the grades put connections in the right order, with the ranking clearly better than chance?
- Is there enough history: at least 2,000 tested connection-days (or flight-days) per airport?
- For connections within Brazil: do Brazil's own figures, measured on alternate weeks, hold within 5 points on the other weeks?
What the test found (25 October 2026)
As tested: 2,025,340 connection-days within Brazil, 144,375 international-arrival connection-days, 386,917 domestic flight-days and 50,020 international flight-days, September 2025 to May 2026. Full report: available on request (contact us).
| Result | On the card | |
|---|---|---|
| Connections within Brazil | Passed every check | Grade, with Brazil's figures: A 5%, B 8%, C 9%, D 9%, F 17% missed connections |
| Domestic flights | Passed every check, but the grades separate only weakly: neighbouring grades B and C, and C and D, did not separate, and at F the model expected 19.0% against 10.1% that happened | Grade |
| International flights | Did not pass: B-graded flights ran late or were cancelled more often than C-graded ones | Grade marked Provisional |
| International arrival to domestic | Did not pass: the model expected far more missed connections than happened. Two of its inputs also lacked a source dated before the test, so it wasn't a clean test | Grade marked Provisional |
Grades marked Provisional are shown because our owner decided Brazil should have grades at launch. We quote no percentage for them.
Connections within Brazil, missed-connection rate by grade (whole test period, 2,025,340 connection-days): A 5.1%, B 7.8%, C 8.8%, D 9.0%, F 16.8%. The model's own estimates were within 1.6 points of what happened, on average; grade by grade, it expected 3.0% at A against 5.1% that happened, and 20.7% at F against 16.8%. Each grade was clearly worse than the one above it, except C and D, which had about the same chance of a missed connection.
Connections of 40 to 50 minutes: of the ones we grade, we estimate a missed connection on 9.3% of connection-days at Brasília, 11.5% at Congonhas, 12.9% at Guarulhos, 13.1% at Campinas and 13.5% at Confins (whole test period). Between a third and over half of the connection-days in this range, depending on the airport, have no grade and aren't counted. A missed connection is estimated from the flights' real times and our modelled transfer time; there are no passenger records. Source: ANAC, VRA.
About a third to over half of the connections within Brazil in the test had no grade (shown “not graded yet”), probably because at least one of their flights had too little matching history in the records we built the model from (our analysis; not yet verified).
International arrival to domestic, what this test can and can't show: the test checks how often the first flight landed late or the second was cancelled, against our estimated time from landing to the domestic gate. It can't check that estimate itself, because no public record says how long passengers actually took. Since the test, we've changed those estimates. Passport control and customs now use the Ministry of Ports and Airports' measured queue times, and the wait at the bag belt is worked out from its measured bag times (2016–2021). At Guarulhos, the move between terminals starts from an airline's published time (about 40 minutes); for the terminal reached by shuttle the move is our estimate (about 45 minutes), which is our judgement. Re-checking bags and the move at Campinas are our judgement too. Getting off the aircraft is our assumption (about 10 minutes); walks use OpenStreetMap distances. The grades you see use these newer figures, which this test did not check.
Things we disclose:
- Brazil's own figures were chosen as the yardstick after a practice month (August 2025) fell short of the US figures.
- One row of the test records was seen early by our data team, who don't build the model; nothing from it was used.
- The program that produced the test predictions was changed twice during the run, to save memory. Both times we proved it gave byte-identical results first.
- Boarding closing time and some steps of the international arrival are modelled, with no dated source.
Sources
- Flight records: ANAC, VRA, September 2025 – August 2026, downloaded 21–23 October 2026 (the full ANAC line is in the page foot).
- Walking distances: OpenStreetMap contributors (ODbL).
- Guarulhos terminal move: GOL's connections page (voegol.com.br/conexoes-em-voos), as published 6 August 2025 (Internet Archive capture 20250806003924).
- Passport control, customs and bag times: Ministry of Ports and Airports, quarterly airport performance reports, 2016–2019 and Q4 2021.
- Boarding closing 15 minutes before departure: modelled, with no dated Brazilian source.