One Call Against One Call
Two cheap-tier models shipped last week with a million tokens of context and multimodal input. The benchmark tables comparing them to frontier models are measuring the wrong thing.
Two cheap-tier models shipped last week with a million tokens of context and multimodal input. The benchmark tables comparing them to frontier models are measuring the wrong thing. What shipped DeepSeek V4-Flash-Vision-Exp — 21 August. An experimental multimodal model taking text and images and returning text. One million tokens of context , 384K maximum output, at $0.14 per million input tokens and $0.28 per million output . GLM-5.3-Flash — 26 August. A 320B-A18B mixture-of-experts model, one million tokens of context , at $0.15 / $0.50 — released under the MIT licence , and revealed as the identity behind the "Ox Alpha" model that had been appearing anonymously on leaderboards. For scale: Grok 4.6, covered elsewhere in this week's scan, is $2 / $6 . That is roughly 14× the input cost and 21× the output cost . "Flash" no longer means text-only, and it no longer means short context. The fact that matters most, and it isn't a benchmark GLM-5.3-Flash is MIT licensed. That is not an API you rent. Those are weights you can self-host, fine-tune, and embed in a shipped product without a licence negotiation. If data residency, latency, or vendor concentration has been blocking something on your roadmap, that block just came off. We argued the other way in August — that most teams should not self-host, because the serving cost and operational burden rarely clear the saving: 2.4 Trillion Parameters, 95 Billion Awake — and Why You Probably Still Shouldn't Self-Host. That reasoning still holds for most teams. What has changed is the size of the model you would be hosting: a 320B-A18B model is a materially smaller operational commitment than a 2.4T one, and it is the first release that genuinely tests the argument rather than confirming it. Almost all coverage led with benchmark scores. The licence is the more consequential fact. Why the benchmark tables are asking the wrong question Every comparison published this month has the same shape. One prompt to model A, one prompt to model B, score the answers, higher number wins. The shape is so familiar it has stopped looking like a choice. It is one — and it encodes an assumption. DeepSeek reports Terminal-Bench 2.1 of 83.9 against Claude Opus 4.8 's 85.0 — its own figure, with no independent verification. At equal prices, 83.9 loses. But these are not equally priced. The honest comparison at a fourteenfold price difference is not one call against one call. It is one call against fourteen calls — and fourteen calls is not merely more of the same thing. It is a different method: - Sample seven times and take the majority. Removes most of the variance that produces wrong answers in the first place. - Generate once, verify twice with fresh context. Catches a class of error that no single pass detects by construction. - Decompose a hard problem into six easy ones , each of which a weaker model handles comfortably. None of these are exotic. They are how anyone actually deploys a cheap model. None of them appear in an