Work out where the 0.0026 in the mid tier's retry line comes from. It is 156.00 divided by 60,000 conversations, the tier's own blended cost per call [2]. Retries are therefore billed at the average, which is a fair simplification and also why the retry term vanishes: 6,000 second attempts come to 15.60, against 7,200.00 of escalated human time [10].
Then compare the two deltas. Going from mid to frontier adds 609.00 of token spend and removes 4,800.00 of escalation cost, so the human term moves about 7.9 times more money than the rate card does [3].
That ratio is why the escalation rate is the only line worth arguing about, and you can solve for the point where it stops mattering. Hold the frontier tier at 1 percent and its monthly total at 3,165.00 [11]. The mid tier's non-escalation costs are 156.00 plus 15.60 [10]. That leaves 2,993.40 of escalation budget, and at 4.00 per escalation across 60,000 conversations the mid tier breaks even at 1.25 percent [4]. Its winning window is a quarter of a percentage point wide [5]. Measuring an escalation rate to that resolution means labelling real transcripts, which is more work than the four multiplications and is the work that actually picks the tier.
The table is illustrative and says so [3]. For its ranking to carry into your system, four things have to hold: an input to output ratio near 4.3 to 1 [6], a two point escalation gap between the tiers [10], six-minute human handling, and a 40.00 loaded support hour [9]. The ratio is doing more than it looks. The frontier input rate is 6.25x the mid tier while its output rate is only 3.75x [8], and the blended multiple that falls out, 765.00 over 156.00, is 4.9x [7], sitting near the input multiple because this workload reads about four times as much as it writes [6]. Push the mix toward long generations and the blended multiple slides toward 3.75x [11], and the mid tier keeps more of the token advantage it started with.
The input count is where estimates rot. A 40 token user message with a 900 token system prompt and three retrieved documents in front of it is billed in full on every call [14]. Pull the transcript out of your logs and count everything you send.
The article declines to print a single real rate, on the grounds that published prices move and a number baked into an article goes stale silently [12]. That is the right call for a method piece, and it means the inversion in the table is a property of the placeholders rather than of anyone's current price list [5]. The author also refuses the easy summary that the frontier tier always wins, and names the escalation rate as the deciding variable instead [13]. A cent per conversation has never blocked a launch [7]; it is review meetings, not cents, that decide what gets contested.
Four multiplications produce a number a product manager can act on. Getting the decision itself right still takes a day of transcript labelling.
