The person who has to act on this owns a routing table: the config that decides which requests go to the expensive endpoint and which get answered locally. For that person the interesting figure is not the midpoint. It is what sits at each end of the line the midpoint divides.
Start with the money, because it is the only unambiguous part. One-twentieth of the cost is a 95 percent saving [12], which means the budget that buys one call to the full GLM-5.2 buys twenty bridged ones [14]. The parameter gap being spanned is roughly 188 to 1 [13]. Those are good numbers on their own terms.
The quality claim is softer than it looks. Halfway is a position on a line, and Wired's account does not say which benchmark or task drew the line [15]. Between a 753-billion-parameter model and a 4-billion one, the distance could be twenty points on something you care about or it could be two. Halfway across a narrow gap is close to free money. Halfway across a wide one may still land below the level you are allowed to ship.
Karl Tuyls, formerly of Google DeepMind and familiar with the work, describes the result as approaching large-model quality without the large model handling the entire loop, with a smaller model running alongside [8]. That is worth reading literally. The frontier model is still in the loop, it is just doing less of it. The pitch is models talking without words, and in practice that means shedding work off the expensive model onto a cheap one without paying for a text round trip in between [2].
There is also a boundary condition sitting inside the mechanism. The bridge operates on values found in the weights [1], the demonstration used two Chinese open-weight models [3], and Wired frames the payoff as making open-weight models more competitive against the closed offerings of labs like Anthropic and OpenAI [11]. Read together, that points away from the endpoint most teams are actually billed for. Most teams hold an API key, not the weights the bridge needs.
The evidence is thinner than the enthusiasm around it. The ARC-AGI 3 result arrives with no details, because Mostik wants to win the contest [5]. Vladimir Arustamian of Lovable, who knows the team, says they have something running that he would have guessed was years away [9], while Mostik's own chief scientist, the Fields medalist Stanislav Smirnov, says there is still no appropriate mathematical language for finding common ground between two models [10]. None of the people quoted here have reproduced the numbers themselves.
The forcing function is a two-by-two you can build from your own logs. One axis is the width of the gap between your small model and your frontier model on your evaluation set. The other is whether your acceptance threshold falls above or below the midpoint of that gap. Wide gap with a threshold under the midpoint is where a bridge pays, and it pays well. Narrow gap means you were overpaying for the frontier model already and should simply drop down. A threshold above the midpoint means the discount buys you answers you will reject, however wide the gap. A gap you have never measured means you are pricing a midpoint between two numbers you do not have.
So run both endpoints against your own set, then price the middle. Sasha Malysheva's bet is that capability will come from combining models rather than scaling a monolithic one [7], and if she is right this becomes ordinary plumbing within a year. The cost of going early is that you inherit a quality level defined by somebody else's benchmark.
