The arithmetic on the gap is the finding. At 95.5% of 299,668 August calls, about 13,485 requests were not successes [5] [17]. At 98.9%, about 3,296 were not [6] [18]. The difference is roughly 10,190 calls, near three quarters of the whole failure bucket, and those were customer prompts the upstream model refused on policy grounds rather than anything the gateway did wrong [19].
For that split to be a measurement instead of an opinion, the forwarding layer has to distinguish an upstream policy refusal from an upstream 5xx and from its own timeout. The post attributes the difference to refusals but does not describe how they are identified [24]. If refusals arrive as generic HTTP errors from some providers and as structured refusal payloads from others, 98.9% is a judgement call with a decimal point on it.
The July-to-August bridge has the same shape of problem. Volume grew 2.6x, from 116k to 300k calls, while the reported success rate went from 93.9% to 98.9% [7]. But 98.9% is the moderation-excluded number and 95.5% is the raw one, and the post does not say which basis the July figure uses [21]. If July is raw and August is adjusted, part of that improvement is a change in definition. That matters, because "volume nearly tripled and stability went up" is the load-bearing claim of the whole systems section.
The retention number does something a counter cannot. The author wired a very cheap upstream for a popular video model, and several users tried it once and never came back, with no errors and every request returning 200 [10]. Running the same prompt through the official channel and comparing output side by side showed a watered-down clip, most likely a lower resolution upscaled or a cheap tier passed off as the premium one [11]. A 200 came back regardless of whether the pixels behind it were right. The rule that came out of it is official model channels only, never a reseller priced far below market [12], and the instrument that caught the substitution was retention, not the error rate [13].
The 70.8% figure needs care before anyone borrows it. It is a share of all revenue ever collected, not of the $8,880 month, and the cumulative total is not published, so you cannot back out dollars per repeat customer [8] [22]. What survives is a count of people who chose to pay a second time, which no dashboard configuration can flatter.
The build decision follows from the same place. The product is one API key in front of 182 models behind an OpenAI-compatible interface [9], and off-the-shelf open-source gateways were rejected because price and stability both live in the forwarding logic between request and upstream, which is exactly the layer such a gateway occupies [14]. A bug in a dependency means filing an issue and waiting; a bug in your own code gets fixed that afternoon [15]. That trade only pays if you can genuinely ship that afternoon, and here every mechanism was retrofitted after a specific production failure rather than designed up front [16], with an AI coding agent writing it under a product manager's direction [1].
Whether 98.9% transfers to your gateway depends on whether you can prove which bucket a given call landed in.
