The max-temperature charts shrug. Your body doesn't. If you looked up the numbers and came away confused — you were right to be. The thermometer on a weather station pole is measuring one thing. The city around you is a different story.
Every April, the same argument: "It's worse this year." Someone pulls out a chart of peak temperatures. The chart disagrees. Both sides walk away unconvinced.
An average April day in the 1980s had 3½ hours below 22°C — the body's real recovery window. In April 2026 so far, it's under one. The heat-stress band grows by two hours. Not a warmer peak. A longer stay.
Each row is one average April day, stacked by temperature band. Dark blue is the body's recovery window. Deep red is heat stress. Watch which one vanishes and which one extends.
The decade-scale average barely budged. April 2026 is running about 1.5°C above both five-year windows — at every hour of the day. ERA5's 9 km grid averages city and surroundings, so this is a conservative picture of what the thermometer on your balcony is doing.
Source. ECMWF ERA5 reanalysis, ~9 km grid, pulled from Open-Meteo's archive API for the Bangalore coordinate (12.97°N, 77.59°E), Asia/Kolkata timezone. Three windows: April 1985–89, April 2021–25, and April 1–20, 2026.
Why five-year windows. A single April can be a freak — monsoon timing, El Niño, local weather. Averaging five Aprils smooths that out and leaves a climate-scale comparison. Each five-year line is 150 days of hourly data; 2026 so far is 20.
Why April. The pre-monsoon peak. The month Bangaloreans complain about, and the one where urban heat island effects hit hardest — dry air, high solar load, no relief from rain yet.
The metrics. Hours ≥ 26°C uses a common body-stress threshold. Hours ≥ 28°C is the sleep-quality threshold. Degree-hours above 26°C is the sum of (temp − 26) across the 24 hours of a day when that sum is positive — essentially the area under the curve above the comfort line. It's a single number for total daily heat burden.
What this is not. Reanalysis blends station, satellite, and aircraft observations into a gridded model. It's excellent for decade-scale comparison at a location. It is not the thermometer on your balcony — the 9 km grid cell averages concrete and farmland together, so it understates the urban heat island. The real city is hotter than this chart; the direction of change is the point.
A 1°C shift in the mean sounds like nothing. Bodies don't feel a mean — they feel duration. Same 24-hour day; very different number of hours spent above the lines that matter.
How many hours above each threshold · per average April day
How many hours above each sleep threshold · per average April night (22:00–07:00)
Bars to scale: pale orange is 1985–89, deep red is 2026. Same scale across each group. Day thresholds each gained 2–2½ extra hours. The night's cruelest move is at 22°C — the body's real sleep-cool window shrinks from three and a half hours to forty-eight minutes.
Three five-year windows can hide a one-off spike. So here's every April, 1985 to 2026 — forty-two years of hourly data, one dot per year, on the two metrics that matter: hours the night stayed cool enough to sleep in, and hours the day spent above heat-stress.
Each dot is one April. The line is a five-year rolling average — smooths out El Niño / La Niña flukes. 1997 was freakishly cool (look at the blue spike); 1998 came right after and flipped to brutal. Since roughly 2010 the night's recovery window has started drifting down, and the hottest Aprils have clustered: 2016, 2019, 2024, 2026. 2026 is not an outlier. It's the pattern continuing.
Hours are abstract. Clock times aren't. Here's the day — and the night — as an average Bangalorean's body used to experience it, and as it does now. The cool windows have different ends now. One of them has no end at all.
Colored zones: when the average hour sits in each state. Blue = below 22°C (real cool). Red = at or above 32°C (heat stress). Cream between.
The blue band has disappeared. The red band has moved outward at both ends but grown more at the PM end — heat stress now lasts nearly two hours longer into the evening. Clock-times measured by linear-interpolating the hourly mean curve across each crossing.
The city isn't warmer because the planet is warmer. It's also warmer because of what we paved, cut, and filled. Between 1973 and 2023, in Bengaluru:
The cheapest urban-cooling intervention ever measured. A dark concrete roof absorbs sunlight all day and dumps the heat back into your bedroom at night. A white-coated roof throws 70% of it back at the sky. Same building, different day.
Numbers from LBNL / Berkeley Lab cool-roof research and the Ahmedabad Heat Action Plan — which painted 20,000+ slum roofs white between 2017 and 2020 and measured indoor temperatures drop 2–5 °C during heatwaves.
The forest the city lost is gone. You can't replant 88% of canopy this decade. But the roofs — those are still there. Four hundred square kilometres of flat concrete, heating all day and radiating all night. All of it paintable.
Your building's terrace. Your school's. The bus stop's.
The cheapest cooling Bangalore has ever had access to.
The questions that came back most after this went out — on Twitter, in DMs, at dinner. Answered here so the conversation can keep moving.
Both. 2026 is extreme, but not unprecedented — 2024 actually had a hotter average daily max (36.9°C vs 34.6°C) and 8.3 hrs/day above 32°C vs 2026's 6.9. April 2016 had even less sleep-cool than 2026: 0.73 hrs/night vs 2026's 0.85.
What is new is the frequency. Aprils with under 1.5 hours of sleep-cool per night were a one-in-25-year event before 2010 (just 1998). Since 2010 they're one-in-four: 2016, 2017, 2024, 2026. See the year-by-year charts in "Is 2026 just a bad year?" above.
ERA5 is a gridded reanalysis product from ECMWF — it blends observations (including IMD stations, satellites, radiosondes, aircraft reports) with a physical model, producing a consistent hourly record back to 1940. For a 40-year trend analysis it has three advantages over raw station data: it's continuous (no missing days), it's methodologically stable (no station relocations or instrument changes to confound the signal), and it's hourly (station data is usually daily min/max).
Downside: it represents a ~30km grid cell, so it smooths out the very heat-island effect we're writing about. The trends here are therefore conservative — the lived experience in built-up Bengaluru is almost certainly hotter than ERA5 shows.
22°C is a widely cited threshold for comfortable thermoregulation during sleep — below it, the body can shed metabolic heat passively without sweating or a fan. It's also where most sleep-science papers place the lower end of the "thermoneutral zone" for adults under a light sheet.
"Sleep-cool hours" here means the count of hours between 22:00 and 07:00 local time when temperature is below 22°C. It's not a verdict on whether you slept — it's a measure of whether the outside air was doing any of the work for you.
In the April 2026 data through the 20th: the average hourly night temperature never crosses below 22°C — the minimum of the 24-hour mean curve is 22.65°C. Individual nights do still go below 22°C at specific hours, but they're rare. In the 1985-89 composite, the average night curve spent roughly 3½ hours below 22°C; in 2026 it spends about 48 minutes.
Partially, yes — outdoor AC units dump heat into the street, which is a measurable contributor in dense neighbourhoods. But the bigger drivers in Bengaluru over this period are the ones in the "What changed on the ground" section: built-up area up 1,055%, vegetation down 88%, water bodies down 79% (1973-2023, per the ICCEAS 2025 paper). Concrete holds heat; trees and lakes used to release it. Losing the second is a much larger effect than gaining the first.
All reproducible. The fetch scripts hit the
Open-Meteo archive API
for ECMWF ERA5 hourly data at 12.97°N, 77.59°E;
compute_comfort.py turns the hourly series into the
comfort bands and threshold crossings used in the charts. Python
stdlib only — no external dependencies.
Scripts:
fetch_bangalore_april.py — pulls March & April hourly data for the three composite windows (1985-89, 2021-25, 2026).fetch_yearly_april.py — pulls April every year 1985-2026 for the trend charts.compute_comfort.py — computes comfort bands, exceedance curves, and threshold-crossing times.Data (source of truth for every number on this page):
bangalore_raw.json — raw ERA5 hourly series.bangalore_data.json — 24-hour mean curves per era.bangalore_comfort.json — comfort bands, exceedance, crossings.bangalore_yearly.json — per-year metrics, 1985-2026.Claude (Anthropic's LLM) was a genuine collaborator on this piece — not a spellchecker and not the only author. In practice: Prasanna set the question and pushed back on the framing; Claude wrote most of the Python, pulled the data, iterated on chart designs, drafted and rewrote prose, and cross-checked the numbers. Prasanna made the calls on what the story was and whether it was honest.
Every number on this page has been verified against the source JSON. Every claim has been edited by a human who reads ERA5 output for fun.
Yes and yes. The scripts are city-agnostic — change
LAT and LON and everything works.
Easiest path: open this shared Claude conversation. Fork it, tell Claude your city, latitude, longitude, and timezone, and it will pull the ERA5 data and generate the scripts and the HTML for you. No copy-paste needed.
Prefer the raw prompt? It's here: REMIX_PROMPT.md. Paste it into a fresh Claude conversation if you'd rather start clean.
Either way: runs entirely locally. No API keys, no hosting keys, no npm. Open-Meteo's ERA5 archive is free and unauthenticated.
If you'd rather I did it for Chennai, Hyderabad, Delhi, or anywhere else, ping me — an afternoon of work.
The thermometer at the airport reads 37°C today, same as it did a decade ago. The chart looks boring. But the experience of living in Bangalore in April has changed — and not by a rounding error.
Nights don't cool off. Summer starts in March. The heat reaches you at 8:30 am. The trees that used to be overhead aren't. The lake that used to breathe onto your neighbourhood is a drainage canal. The wall facing your window has been absorbing sun since 11am and will radiate it into your bedroom at 1am.
What to watch instead of the max: the minimum overnight temperature, the number of consecutive days above 35°C, and the percentage of the day your tile floor is warmer than your body. Those are the numbers that match what you're feeling.
The city has quietly become a different climate. The charts just haven't caught up.
The error
An earlier version of this piece — and the carousel that
accompanied it — cited
"27.1°C · Bangalore's record-high minimum, set 9 June 2025"
as "the number you feel" in the discrepancy section. That figure
was unsupported by primary data and has been removed.
How it happened The number was introduced by Claude (Anthropic), our co-author, during the initial drafting of the article. It was not cross-checked against primary observations before publication — a failure of the very "don't invent numbers" rule the piece itself is supposed to embody. In short: it was hallucinated, landed in an attributed citation box, and went out the door. Responsibility for the oversight is ours jointly. Claude wrote the number; a human editor failed to verify it.
How it was caught West Coast Weatherman (@RainTracker) pointed to the Weatherspark record for 9 June 2025, which showed a daily minimum of ~22°C — nowhere near 27°C. We then re-pulled the ERA5 hourly series for Bangalore for every March–June from 2015 to 2025 to find the actual warmest night in the record. Thanks to him for the catch.
The correction The discrepancy card now reads "25.0°C · The warmest night in Bangalore's ERA5 record — 14 April 2016". Four of the five warmest overnight lows in the last decade fell in a single week of April 2016; another cluster appears in April-May 2024. Those are the same years the year-by-year chart flags as brutal, so the substitution makes the underlying story land harder, not softer. The 39.2°C / April 25, 2016 all-time-max claim was separately verified against the IMD Bengaluru Meteorological Centre's April extremes PDF and stands unchanged.
What we changed upstream The remix prompt used by others building "Is [their city] hotter?" pages has been updated with an explicit warning not to quote aggregator "record-high minimum" figures without verification against IMD station data or a direct ERA5 pull. We'd rather no one inherits this mistake.