Most of the model market is still unbuyable on latency

Only 57 of 396 listed models in this week’s catalogue have a published time-to-first-token measurement, which means 339 listings are unmeasured on latency rather than fast. On the subset that is measured, TTFT runs from 0.56s at the tenth percentile to 36.6s at the ninetieth, with a 1.58s median — a spread wide enough to make “cheap” and “usable” two different questions entirely (OpenRouter for listings and price, Artificial Analysis for the measurements).

That is the most actionable inference finding from the week because it changes how you should shortlist models. If your application has a human waiting on the first token, the market you can safely buy from is not the 396-model catalogue headline; it is the 57-model subset with published TTFT, and even that is not large enough to assume that price, quality and speed line up neatly (OpenRouter for listings and price, Artificial Analysis for the measurements).

The week’s external material pointed in the same direction. AWS said it cut automatic speech recognition serving cost by 75% using NVIDIA MPS with Triton on EC2, a reminder that utilization and concurrency policy can matter more than the model switch people talk about publicly (AWS). Together AI reported more than 2x single-node throughput from its ThunderAgent scheduler by reducing KV-cache thrashing in agentic workloads, again showing that runtime behavior, not just model choice, is where the real performance budget leaks out (Together AI).

But those posts are vendor stories. The more useful result is what the measured catalogue says when you look across providers at once. The cheap end is not uniformly slow, and the expensive end is not uniformly responsive. Ling-3.0-flash sits at about $0.02 per million input tokens and 371 tokens per second, with a published TTFT around 1.64s (OpenRouter for listings and price, Artificial Analysis for the measurements). OpenAI gpt-oss-120b is roughly $0.037 per million with 179 tokens per second and 0.52s TTFT. NVIDIA Nemotron 3.5 Lightning is about $0.08 per million, 283 tokens per second, and 0.54s TTFT. Those are not marginal differences. They are the difference between a product that feels conversational and one that feels queued.

Now look at the other end, and read the variant names carefully. The slowest published first tokens in the catalogue are not slow models — they are extended-reasoning configurations, where the measured TTFT includes the whole thinking phase before a single token is emitted. Claude Sonnet 5 (Adaptive Reasoning, Max Effort) is measured at 121.77s to first token; GPT-5.6 Terra (max) at 119.68s; GPT-5.6 Luna (max) at 69.71s. The same GPT-5.6 Terra at medium effort comes back in 1.59s — seventy-five times faster to start, on the same model (Artificial Analysis). Thirty-three of the 488 measured entries sit above 30s, and effectively all of them are these high-effort variants. So the useful correction is not that some models are unusably slow. It is that reasoning effort is a latency decision as much as a quality one, and a shortlist that ignores which variant it measured will compare a thinking configuration against a default one and call it a model difference.

That distinction matters more now because the catalogue is mature enough that “best value” and “best experience” have started to diverge. Dividing the measured intelligence index by list price — a screen, not a verdict, and only meaningful between models — puts Ling-3.0-flash first at roughly 1,800 index points per dollar, ahead of Upstage Solar Pro 4 near 1,390 and Z.ai GLM 5.3 Flash near 770 (OpenRouter for listings and price, Artificial Analysis for the measurements). That is useful as a first pass. It is not enough to buy on. Solar Pro 4 scores well on quality per dollar, but its published throughput is only 34 tokens per second, far below the faster flash-class options in the same broad price band. If you optimize only for quality per dollar, you can easily buy a model that is economical on paper and expensive in wall-clock.

The practical implication is simple: stop treating missing latency as neutral. This week’s coverage shows that only 38% of listed models are independently quality-scored and only 14% have published TTFT at all (OpenRouter for listings and price, Artificial Analysis for the measurements). In other words, the catalogue is much better at telling you what can be called than at telling you what can be deployed well. That is not a criticism of any one provider. It is a market-structure problem. Measurement is still sparse, and sparse measurement encourages buyers to substitute price for evidence.

That substitution is how teams end up with bad shortlists. A common workflow is: cap the input price, filter for a context window, skim a benchmark aggregate, then test two or three candidates. The flaw is that none of those filters tells you whether a user will stare at a blank cursor. This week’s measured spread — 0.56s to 36.6s from the tenth to ninetieth percentile, with a 1.58s median — says you should invert the order. Start with measured TTFT and throughput, then move to quality, then price, and only then the long tail of secondary features like context and tool syntax (Artificial Analysis).

The week also included a small cluster of price cuts, led by Z.ai GLM 5.1 down 23% from $1.26 to $0.97 per million input tokens, plus DeepSeek V4 Pro 0423 down 17% from $0.55 to $0.46 and Qwen3 Next 80B A3B Instruct down 10% from $0.10 to $0.09 (OpenRouter). Those matter, but less than they look like they do. A price cut without latency evidence narrows your cost estimate and leaves your user experience estimate mostly untouched.

So the answer to the week’s question is blunt: the model market you can actually shop for interactive inference is much smaller than the catalogue suggests, because latency measurement is still scarce and price does not predict responsiveness well enough to stand in for it. Until that changes, the safe procurement rule is not “pick the cheapest model above a quality threshold.” It is “pick from the measured subset first, and treat the rest as unpriced operational risk” (OpenRouter for listings and price, Artificial Analysis for the measurements).

Both projects are on GitHub and PyPI. Install them.