Conquering the Final Cycle
We did it. The Opus Magnum community has built a solution at minimum cycles to every campaign puzzle. The final cycle, Alcohol Separation 16, fell at the hands of rebix, with significant help from Grimmy, Jammuu, Villfuk, and support from many others on discord. Here’s the gif!
Backstory
Opus Magnum has been out since October 2017. By January 2018 we had pinned down theoretical values for the lowest possible cycle counts, and built a table on reddit. Of the 36 campaign puzzles, we already had 32 of them to the point where we built a solution with the best possible score. 4 of them were “theoretically improvable.”
Those 4 holdouts were Surrender Flare, Alcohol Separation, Litharge Separation, and Universal Solvent.
They were still holding strong 6 years later in 2024, each one cycle away from perfection. Then in September the first fell, with Universal Solvent 38:
Less than a week later, Litharge followed:
As we entered 2025, the list was down to 2. Two puzzles which weren’t perfect just yet. Both needed 16 cycle solutions, and all we had was 17.
But if you are following the blog closely, you know that Surrender Flare just went down. 3 days ago:
In the wake of this, everyone turned to Alcohol Separation once again. As crazybot put it:
In those 3 days, we thought about this puzzle more deeply than ever before. So many of the brilliant players who are active in the server because of the tournament, got nerd sniped. And that turned out to be enough!
What Makes it So Hard
In the lore of the game, Alcohol Separation comes when the protagonists have to recover supplies after a horrible tragedy. Using only Alcohol, they synthesize all of the four base elements: Fire, Air, Earth, and Water. This requires using new mechanics, and so this is how the game introduces the player to the Van Berlo’s Wheel and the glyph of duplication.
The Van Berlo’s Wheel (or as we call it, the Berlo) holds 6 atoms in a ring around a hex arm. These atoms cannot be dropped. You only get one Berlo per solution. The only purpose it serves, is to duplicate elements onto salt.
The glyph of duplication allows a salt to become whatever element is on the other side of the glyph, and both ordinary atoms and Berlo atoms count here.
A basic solution to alcohol separation might look like this:
Each input becomes one of each element, in a clean little loop. The Berlo spins in order to create both earth and air, which happen to lie on opposite sides.
In order for the puzzle to count as complete, there must be 6 atoms dropped into each of the 4 output glyphs. In this solution, the three atoms held by the hex arm are first dropped on cycle 13, and then every 11 cycles afterward as the machine loops. The water is dropped a couple cycles earlier, but this doesn’t improve the cycle count. It comes out to 13 + 11×5 = 68.
Going faster, things get hard.
Full Throughput
You could grab an input every 2 cycles, leading to an average of 2 cycles per output per output glyph. If you do this, it starts to be rather inconvenient to make one earth and one air from each input.
Instead, full throughput solutions tend to make odd inputs into Fire/Water/Earth/Earth, and even inputs into Fire/Water/Air/Air. This still makes all products in equal amount, but it avoids the complication of accessing both sides of the Berlo for each input.
An example of such a solution is here:
Two hex arms grab alternating inputs. The ones that go to the top part of the solution end up being air, and the ones on the bottom end up being fire. However, they don’t all duplicate off of Berlo. Because of the way inputs sit next to each other in the different positions of the hex arm, it becomes possible for air and earth to duplicate onto the salt from the next input in the cycle (called “buddy duping” by the community). This means that the first input around the loop doesn’t get duplicated, which is why there are extra salts sitting off to the side that weren’t able to output.
This rather beautiful loop does successfully reach full throughput, but the cycle count is still 33. The first output of fire isn’t until cycle 23, because of how long it spends on the board.
Lower Cycles
Let’s look at Grimmy’s old cycle record from 2017, which was the best known at the time we came up with the table of theory minima:
Fun fact about this solution by the way. They had built an 18 cycle solution which took an extra cycle on salt, but then had an idea at work.
We can see a few similarities to the other full throughput solution. For one, it has alternating inputs going to the top and to the bottom. But instead of hex arms, the inputs are now grabbed by arms on track loops, allowing for much more specialized motion.
The inputs that go to the top have both salts become air, and the inputs that go to the bottom have both salts become earth. This means the Berlo does the exact same rhythmic rocking back and forth as in the other solution.
Grimmy wrote up quite a bit about this solution in their reddit post here. However, for the purposes of this blog post, the most important discussion points are debonding and latency.
Debonding and Latency
Notice that the atom in the middle, the water, has to break 3 bonds. Whenever a molecule, held by one arm, breaks in two on a debonder, one of those parts is now not held by any arm. To move that part any further, you have to grab it. Grabbing takes a cycle in Opus Magnum. This is one of the main difficulties with any puzzle that involves disassembling a molecule very quickly.
In Grimmy’s solution, these debonds happen on 3 of the first 4 cycles that the input spends on the board. The first time it breaks, the arm keeps holding onto the 3 atom piece, while the 1 atom piece waits a cycle to be grabbed.
After arm 1 pivots, it is only holding the single atom, so now the 2 atom piece waits to be grabbed.
In order to prevent losing any additional cycles from needing another grab later, Grimmy then grabs this 2 atom piece with two different arms at the same time, and they slide in unison. This is the only way to break apart a 2 atom molecule such that both atoms are held immediately after debonding.
We call this a multigrab, and it is very restrictive. Importantly, if two different arms grab the same molecule at the same time, the only allowed moves are translations, so you can only move one tile per cycle.
In Alcohol Separation, the central water atom has to debond 3 times. Less obviously, whichever of the 3 atoms it debonds from last, has to move from the debonder to the output too. Both of these atoms cause a problem when minimizing latency. This solution shows that it all can be done with 4 cycles of latency (debond, debond, multigrab, debond). That leads to 17 cycles. But the theory suggests this puzzle should be doable in 16. So what is the theory thinking?
Theory
A few days before Grimmy made their 17, I had found a solution that made its first complete set of outputs on cycle 6. Gfycat has lost the recording to time, but I have this instead that is equivalent to what I built back then in all the important ways.
Right away, there is a triple multigrab. Arms have to grab the first input by 3 different atoms, preventing it from moving more than one tile at a time. After that, 3 debonder positions are completely forced (up to symmetry), because the input hole blocks all the other options within sliding range.
If you do this, you cannot take an input 2 cycles later. But in theory, this could be your 6th input, while inputs 1-5 could all take a different path which repeats every 2 cycles. I don’t know how you would make them all land on the output glyphs, especially when there only gets to be one Berlo. But the theory doesn’t care.
It was based on this idea, that we put the theory limit at 16. Nobody really seriously considered 16 to be practical, and even Grimmy’s post on reddit had said 17 was probably optimal.
When they built their tierlist of remaining cycles, Grimmy listed Alcohol Separation 16 under “most likely not possible.”
New Faces, New Inspiration
As I also discussed in the Cow-Eating post, that tierlist may have led people to believe that certain cycles were not worth even trying to find. Grimmy and I were two of the most confident players, and if our ideas of puzzles had us thinking they were impossible, others tended to shy away from them.
Things began to change in 2023 with the arrival of newer, even more confident players. Then the record books exploded in 2024. Cycle records began falling. First there were the journal puzzles. Alchemical Slag 35 by spiritual shampoo. Refined Bronze 32 by me. Van Berlo’s Pivots 26 by Grimmy. Preservative Salt 21 by Thought. Visillary Anaesthetic 28 by spiritual shampoo. Sailcloth Thread 25 by spiritual shampoo. Wheel Representation 27 by Thought. General Anaesthetic 38 by Thought. Ablative Crystal 36 by Thought.
In between Visillary Anaesthetic and Sailcloth Thread, the first week of destruction for campaign cycles came through. That was when we saw Thought take down Universal Solvent 38, and Villfuk take down Litharge 23. Now we are at the second such week, and it is the final one due to no other cycles to save. But the point of it all is that more perspectives allowed us to see the puzzles differently. Certainly Alcohol Separation 16 had an aura to it. But there were other options to make a 16, and they were less bleak.
RushSecond
Detached from the broader community, a player known as RushSecond came up with the second ever original 17 cycle solution. They began playing on Christmas 2019 and had the solution below less than a week later:
This solution processes its odd and even inputs using the same glyphs. It also uses a calcification glyph to turn fire into salt! It compensates by turning alternating salt into fire using “buddy duping.”
For debonding latency, there is a handoff for every water. It happens on different debonders for odd and even inputs, but it is always present. This limits it to 17 in the same way as in Grimmy’s, but the design gives more to think about. Namely, the idea that any atom can become any other atom, which could come in handy.
PentaPig adapted this solution to get it down to 59 area:
In terms of ideas, this is the same solution, just tinkered with by an expert. It didn’t get us closer to 16, but it helped break the spell a little bit.
spectria.limina
Back in 2019, the tournament puzzle Creative Accounting became the obsession of one player. Her name is spectria.limina. During the tournament, the best cycle count submitted was 24. The people to find 24 were panic, me, F43nd1r, and spectria. I was pretty sure 23 was possible, but had failed to find it myself. A week later, spectria managed to achieve the 23:
Notice that the input is shaped identically to the Alcohol Separation input. The player must fully debond it to make an output. Some of the techniques carry over. And spectria believed with certainty that 22 cycles was possible. She wasn’t going to do a multigrab though, she was going to steal an extra input at the end.
That extra input would be too late for most of its atoms to be useful. But one quicksilver was useful. It let her ignore iron on the would-be-last input. Because inputs come every 2 cycles, the first debond on the stolen input happens on the same cycle as the third debond on the would-be-last input. So she could keep a hold on this quicksilver, ignore the abandoned iron, and open up the option to do 22 without a multigrab.
And half a year later, she did it.
Come 2024, she was focusing her efforts on Alcohol Separation, having certainty that a 7th input would let players reach 16 without a multigrab. She sent me this spreadsheet of atom timings. It was more promising than the multigrab approach, but she and I worked together and couldn’t figure out an implementation.
W
The Chinese player known as W came up with the third original 17, during the record spree in 2024. With all the space near the input, this got people thinking about 16 again. Spectria considered this new pipeline extremely promising, as you could just slide the 7th input onto the debonder and output its fire immediately. But one would need to send a previous fire to be calcified like in RushSecond’s solution, and have it become something else.
These cross-pipeline moves complicate the whole machine, and turn it into what Grimmy called a “latency carousel”.
It only works to output all atoms by 16, with 6 going to each.
Villfuk
Villfuk made a few attempts, the most promising being this:
It used the 7th fire to get fire out by 16, abandoning the fire from the 6th input. The water already worked, and the air was relatively easy because the final two air came from the 5th input. However, the earth was screwed. It needed two atoms from the 6th input, and was relying on buddy duping so it also needed to output the buddy atom. It was 2 cycles short on earth, despite everything else being a 16.
Villfuk tinkered with this solution to make a 17 in which earth was only one cycle short:
The community tinkered heavily with this layout, believing it was close. I had something based on this that I completely believed could be a 16, but I had broken the air side, and realized too late that my best completion would still require two air outputs both on cycle 16, which can’t happen:
Jammuu
During the 3 day alcohol spree (positive) following Surrender Flare, Jammuu began going live on YouTube whenever she attempted to work on Alcohol Separation 16. Grimmy showed up in many of the streams and the two played a sort of co-op Opus Magnum. Together they found a few more pipelines for the water and fire, each relying on input stealing for the final fire. The biggest problem with Villfuk’s layout is the requirement for Berlo to communicate with the 5th and 6th inputs, despite them being pulled in different directions. A different debonding order would make it possible to use buddy duping on one of the two, allowing for a much simpler completion. This was the goal of these streams:
After success finishing out a promising start on Surrender Flare, reb was interested in any layout promising enough to try to finish. But reb worked mostly independently from the stream, tinkering at home. After a few hours, everyone was feeling the latency carousel. Something felt close until it became impossible to complete due to some contradiction several steps away.
An hour later, Jammuu found another layout and pulled everyone into the “we’re so back”. Poor Grimmy was up to 6am by this point but admitted “i was planning to sleep but now i obviously can’t”.
But at the same time, reb had found something of his own:
Final Stretch
This particular gif is using “scaffolding mod” to add atoms to the starting state, providing air and earth catalysts and initial outputs. The machine was only programmed to use 5 inputs, with the assumption that the previous 2 were going to be routed to the places scaffolding was needed. That could be easy, it could be impossible, it’s Alcohol Separation and by this point nobody knew the difference between the two. What we knew was that reb had solved all but the first two inputs in optimal cycles, and that was a lot of progress.
Grimmy tinkered with this solution to replace most of the catalysts, but it made one extra air and one too few earth. One more crossover and it would work… but it being 7am, Grimmy went to bed. And reb solved it 20 minutes later.
The first input has to do a monster swing that drops off a fire to be calcified and reused later as earth. It also exchanges water for air in-place by buddy duping within the molecule. The arm that moves that input lives on a track all by itself, doing odd jobs. Because earth is now ahead, the 6th input ignores its would-be-earth, allowing it time to output its fire. The 7th input is used for only fire. And it all makes a 16.
It does just the right amount of shenanigans to reach the goal.
Post-Solve Clarity
As soon as we see where all the atoms go, skilled optimizers are always quick to find ways to simplify the handling. This is over twice as expensive and twice as big as the respective cheapest and smallest 17c. Maybe we could find better. Reb did the initial wave of optimizations, bringing the cost down to 945:
A big group came together to try to find the lowest area version though. Featuring contributions from gooey, bred, prayformerci, Psi looks away, Grimmy, and of course reb, the area fell down below 100. Here’s a much more compact implementation of the same design:
Just as PentaPig improved RushSecond’s solution, it has all of the same assignments from atoms to destinations, but it manages to put them there in a very different way.
The world of secondary optimization is a different world, and many more people are able to see things in it. Getting to the point where 16 exists at all is much more difficult. The specific assignments that make it work still feel like magic.
It’s especially funny compared to Surrender Flare 16, where we knew exactly what to do but were bashing our heads against geometry. Here we were churning on a much earlier conceptual phase. There are so many possible ways to put atoms in these holes, but finding one without conflict took serious collaboration, as well as a determined faith that we would get there eventually.
And as a result, the entire campaign is now solved in minimum cycles. This could be framed as “the quest for perfection,” straight out of a Summoning Salt video. Except instead of watching real time video gaming, we are watching machines that people spent hours building, spinning and whirring into place across 2-3 seconds. Such is Opus Magnum.
What is Next?
It felt surreal to update the table of minimum cycles, and write that the first puzzle one encounters in the game with any possible cycles to save is post-game content. Specifically Refined Bronze.
I can speak from experience that the cycle records for that level are kind of insane already. But hey, theory doesn’t stop you from solving it faster. As I write, Jammuu is taking a crack at Hyper-Volatile Gas, a journal puzzle with one cycle theoretically savable. Another option is Synthesis from Alcohol, which once again uses the input from Alcohol Separation. It may have as many as 2 cycles to save, but right now it would be a triumph to even save 1.
Some folks are looking at saving area now. Curious Lipstick is not provably at its minimum area..!
There’s also still a puzzle left in the 2025 tournament, and it’s quite a challenge! Building a computer with an arbitrary encoding from binary to base 6. Haxton has been a good sport, watching his final puzzle drowned out in community effort on cycles that have been there for years and years.
Or we could wait for the DLC. I know that there are going to be metallurgy puzzles with absolutely unreasonable cycle theory minima, and it may never be the case that they are brought to their limits. But I can’t be certain. The DLC doesn’t have a release date, but if you want to play it in its unofficial version, come to discord and check out the modding framework!