New Super Mario Bros. record breaks speedrunning’s “four-minute mile”

You’ll likely never see a human beat Super Mario Bros. faster than this.

Earlier this week, speedrunner Niftski became the first player to ever beat Super Mario Bros. in under four minutes and 55 seconds (4:54.948, to be precise). That might not sound too impressive on the surface; it’s only about a quarter-second under the world record set by Miniland just two months ago, after all, and less than a second under the first sub-4:56 time (4:55.913) set by Kosmic a year ago.

But once you understand everything that needed to come together to break SMB‘s 4:55 barrier, the feat becomes something akin to speedrunning’s version of the four-minute mile. Niftski’s performance is within spitting distance of the machine-generated perfection of tool-assisted speedruns, which use emulator-recorded frame-perfect inputs to push a game to its limits. Niftski’s performance approaches the theoretical limits of what a human can achieve in this seminal game.

What goes into a feat like this? Join me for a quick primer.

Tricks of the trade

In the early 2000s, many Super Mario Bros. speedrunners struggled to beat the game any faster than in 5:08. That’s primarily because they were sticking to Twin Galaxies’ rules, which still prohibit the use of unintended glitches to complete the game faster.

In the 2010s, though, the speedrunning community by and large started ignoring those rules and exploiting a number of useful but hard-to-pull-off tricks to fool the game and save time. Discussing these glitches in full detail would take an entire separate article (this video and this tool-assisted speedrun resources page provide some good breakdowns), but here’s a quick summary of the most important glitches exploited in SMB speedruns these days:

  • Wall clip: Moving Mario with sub-pixel precision allows a player to get Mario’s foot partially (and briefly) stuck in a wall. Jumping repeatedly from that position will get Mario fully stuck, at which point he can clip through and run to the other side. This is useful in saving time getting to the World 1-2 warp zone and setting up a “wrong warp” in World 4-2.
  • Flagpole glitch: Possibly the most important glitch in Super Mario Bros. speedrunning. By performing a version of a wall clip inside the block at the bottom of the flagpole at the end of most levels, the game skips the “lowering the flagpole” animation, saving significant time. A separate version of this glitch used in World 8-3 bounces Mario off a Bullet Bill and also skips the “walking to the castle” animation.
  • Wrong warp: Super Mario Bros. can only load one value at a time in the memory slot for “where does a pipe/vine go?” By carefully scrolling Mario’s position on the screen or backtracking just as the next pipe value is loaded, you can fool the game into loading the wrong value and warping Mario to an unintended location. This is useful for skipping the vine-climbing animation in World 4-2 and skipping a large section of World 8-4.
  • Backward acceleration: From a standing start, Mario gets to full running speed faster if you jump backward. Pulling this off requires a number of frame-perfect button presses in a row but is important for shaving a few frames off in special cases.
  • Wall jump: By hitting the exact right pixel between two blocks in a wall and performing a frame-perfect jump, Mario can jump off the wall and save significant time in one part of World 8-4.

The other important thing to remember in Super Mario Bros. speedrunning is the so-called “frame rule.” This vagary of Super Mario Bros. programming adds a delay to the “blackout” at the end of a level so it lines up with a 21-frame timer. Think of it like a bus that leaves for the next level every 21 frames (0.35 seconds); even if you get to the stop at the end of the level a little earlier, you may end up waiting for the same bus and not saving any time.

This means that saving just a few frames over the best possible performance in a level is often fruitless and that many records come as 0.35 second improvements over the previous best. The exception to this is World 8-4, where time stops as soon as you hit the final axe, and frame rules don’t apply.

Approaching perfection

Back in 2018, YouTuber Bismuth put out an excellent video looking at the speedruns that used these glitches to break Super Mario Bros.‘ 4:56 barrier. Bismuth’s video does a good job of laying out just how difficult it is to perform all these glitches perfectly in a single run, even with practice. One-time record-holder somewes at one point performed over 6,000 streamed attempts on the game and only achieved sub-4:56 pace up to World 8-4 on two of them (both of which failed at the wall jump later in that level).

Back in 2018, many thought this sub-4:56 performance was approaching Super Mario Bros. perfection.

Bismuth also laid out the differences between these human runs and the pre-recorded tool-assisted speedrun (TAS). That TAS, which has remained unimproved since 2011, sets a theoretical limit of 4:54.03 to finish Super Mario Bros. if every input is entered with mechanical perfection. That suggests we’ll never see a Super Mario Bros. performance under 4:54 and that Niftski’s sub-4:55 was the last “full second” barrier to be broken in the game.

But the Super Mario Bros. TAS exploits some acceleration tricks that require holding both left and right on the d-pad at the same time, which isn’t possible on a standard NES controller. Without those tricks (which aren’t allowed under human speedrun rules), the theoretical best is 4:54.282.

By eliminating other TAS tricks that are probably impossible for a human to perform in real time, Bismuth calculated that the best time a human could ever likely achieve on Super Mario Bros. was 4:54.798 (unless other major glitches are found). Niftski’s time earlier this week is just nine frames off from that theoretical human limit.

This handy illustration from YouTuber Bismuth (circa 2018) shows just how little practical difference there is in these completion times.
Enlarge / This handy illustration from YouTuber Bismuth (circa 2018) shows just how little practical difference there is in these completion times.

Approaching that human limit required a few refinements to the speedruns of years past. The most significant came in World 8-1, where Niftski performs a fast acceleration at the beginning of the level combined with a slightly faster, more difficult version of the sub-pixel alignment needs for the flagpole glitch (which requires letting go of right for precisely two frames in the middle of the level). Combined, these moves save just enough frames to reach past the next frame rule barrier, knocking 0.35 seconds off a more reliable method used by many speedrunners.

That, plus some technical jumping improvements in World 8-2, as well as a fast acceleration when exiting a pipe in World 8-4, saved Niftski just enough time to reduce a sub-4:56 performance to a sub-4:55 performance. And don’t forget, those come on top of the “standard” flagpole glitches in Worlds 1-1, 4-1, 8-1, and 8-3; wall clips in World 1-2 and 4-2; and crucial wall jump and wrong warp glitches in World 8-4.

Pulling any of those off individually is difficult. Pulling them all off in a single run, while also playing the rest of the game at full speed and without any errors, is practically miraculous.

Niftski’s full run is worth watching in its entirety, if for no other reason than to hear his exclamations of “come on dude, stay calm… OK, don’t throw up” and “this isn’t real—it doesn’t feel real.” But Niftski isn’t resting on his achievement; his list of 2021 goals for Super Mario Bros. include setting new world records for Minus World and blindfolded runs of the game, among other feats. It just goes to show, when it comes to speedrunning, there are always new frontiers to be crossed.