Source: U.S. Census Bureau & HUD, "Monthly New Residential Sales, August 2026" (CB26-155, Sept 24, 2026): https://www.census.gov/construction/nrs/current/index.html

Quoted headline: "Sales of new single-family houses in August 2026 were at a seasonally-adjusted annual rate of 684,000 ... This is 6.4 percent (±19.5 percent)* above the July 2026 rate of 643,000, and is 2.0 percent (±15.7 percent)* below the August 2025 rate of 698,000."

Quoted prices: "The median sales price of new houses sold in August 2026 was $393,700. This is 0.4 percent (±7.4 percent)* above the July 2026 price of $392,200, and is 5.8 percent (±8.2 percent)* below the August 2025 price of $417,900. The average sales price ... was $478,700. This is 9.1 percent (±11.2 percent)* below the July 2026 price of $526,400, and is 8.8 percent (±7.7 percent) below the August 2025 price of $525,100."

The detail that changes the reading: median up, average down, same month, same closings. Opposite signs mean neither number can be read as "prices rose/fell." An average moves on the upper tail while a median barely moves at all, so a -9.1% mean against a +0.4% median points to a change in which homes sold, not in what homes cost. To tell those apart you need Table 2, "New Privately-Owned Houses Sold, by Sales Price," which my text fetch did not render.

Smaller point, same family: every headline change here is smaller than its own printed band except the -8.8% year-over-year average price, and inventory is printed as "virtually unchanged" rather than as a number. Treat each "fell/rose" claim as unresolved until you have checked the parenthetical.

Separate observation from the landing page (not the release): "Due to a processing error, the Year to year percent changes found in Table 1b of the Excel version released at 10:00am ET on 2/20/2026 were incorrect. The corrected tables were posted at 4:35pm ET on 2/20/2026. The pdf version of the tables did not require correction." So the same reference month existed in two versions on the same day — record the download time and file format, not just the month.

Limits: I have not recomputed anything; the asterisked footnote did not render in my fetch, so I cannot state the bands' confidence level; the mix-shift reading is my interpretation, not the Bureau's.