The Maturity Curve: How AI Models Stopped Competing on Intelligence

Somewhere in the software that publishes this site, a small piece of code runs every morning whose entire job is to answer one question: which model, today? It does not hardcode an answer. It calls Anthropic’s API, asks which model currently answers to the name it wants, and falls back to a default if nothing better replies. I did not write it that way out of elegance. I wrote it that way because pinning a specific model name felt, even a year ago, like a bet I didn’t want to keep making by hand. I think that small piece of infrastructure is a more honest description of where AI models are in 2026 than most of the benchmark charts published about them, including some I’ve written myself. This piece is an attempt to explain why, using real numbers rather than the vibe.

I should say upfront what I’m not claiming. I’m not claiming the models are the same, or that capability no longer matters, or that the frontier has stopped moving — it hasn’t, and anyone who has watched a coding agent go from “occasionally useful” to “occasionally needs supervision” in eighteen months knows better. What I am claiming is narrower and, I think, more useful: that the question a builder actually has to answer has quietly changed. It used to be “which model is smartest.” It is increasingly “which model is priced, hosted, and contracted in a way that fits what I’m building” — and that a market that asks the second question instead of the first is a market that has matured, in the specific, technical sense economists use the word. This report is about the evidence for that claim, some of which surprised me even though I look at model pricing tables most weeks.

Part I · What Maturity Looks Like

The Convergence

Every technology category that lasts goes through the same phase change, and it rarely gets remarked on while it’s happening because it doesn’t look like progress — it looks like boredom. Database engines went through it: for a decade, “which database” was a religious argument with real performance stakes, and then at some point Postgres, MySQL, and half a dozen managed variants converged on “good enough for almost everything,” and the interesting questions became operational — cost, managed-service quality, ecosystem, who’s on call at 3 a.m. Cloud compute went through it. Web browsers went through it twice. The pattern is always the same: a period of genuine, benchmarkable capability differences, followed by a period where the leaders cluster so tightly that the capability gap stops being the thing anyone actually decides on.

I think large language models crossed that line sometime in the past eighteen months, and I think most of the public conversation about them hasn’t caught up, because the conversation is still organized around leaderboards that reward exactly the kind of differences that stopped mattering. A model that’s three Elo points ahead of another is not a better product decision. It is noise dressed as a ranking.

What’s unusual about this particular convergence — and this is the part I don’t think has an earlier analogue — is that it happened simultaneously at the closed frontier and in the open-weight long tail, for different reasons. At the frontier, GPT-5.x, Claude’s Opus line, Gemini 3.x, and Grok cluster within a few points of each other on the hardest reasoning benchmarks, the pattern I wrote about in more detail in The Age of AI. In the open-weight tail, something more interesting happened: a market big enough that convergence shows up not as three or four labs agreeing, but as hundreds of models agreeing, priced by hundreds of independent hosts, in a live, continuously-updated market. That market is the more honest place to look for evidence, because nobody is managing the narrative. The prices are just prices.

Part II · The Market, By the Numbers

What the Board Actually Shows

AgenticFlow runs an index of open-weight models — 545 of them at last count — pulled and re-ranked every six hours, priced per host¹. Only a subset carry a real LMArena Elo score (88 of the 545, by the index’s own count) — the rest are ranked by proxy signals like how many providers bother to host them. That caveat is worth sitting with for a second, because on any leaderboard the illusion of precision is easy to produce and mildly annoying to audit, and I’d rather flag the limits of the data than borrow false confidence from it.

Here is what the top of that board looked like when I pulled it:

Ten models. A top-to-bottom Elo range of exactly two hundred points — tight enough that, given how noisy crowd-preference Elo is at the margins, I’d hesitate to say the top model is meaningfully smarter than the tenth. And a price range, on output tokens, of $0.80 to $11.00 — nearly fourteen times. Widen the lens to the full price table and the ratio between the cheapest and most expensive model in this cluster is closer to eighty-eight times¹. Same weight class. Fourteen-to-eighty-eight-times difference in what you pay to use it.

The single most legible data point on that board, to me, is the Kimi K3 pair. Same model family, same Elo score — 1674, identical to three digits — hosted in two regions, priced differently: $2.00 in / $8.00 out globally, $2.25 in / $11.00 out in the EU hosting tier¹. Nothing about the intelligence changed between those two rows. What changed was infrastructure, region, and whatever margin the host builds in. That is not a capability story. That is a commodity story, and commodities are what markets look like after they mature.

Widen the aperture past this cluster and the pattern holds in the other direction too: a 31-billion-parameter open model you can run on a single workstation clears a respectable Elo score for a few cents per million tokens, in the same index¹. It will not out-argue the frontier on a genuinely hard reasoning problem. For the large majority of what a publishing pipeline, a customer-support bot, or an internal tool actually asks a model to do — summarize, classify, extract, draft, restructure — it is close enough that the gap is not worth what the frontier costs. AgenticFlow’s own framing for its index is an 80/20 split: routine work on a free local model, hard work escalated to a metered frontier one¹. I think that split is roughly right, and I think most teams still aren’t organized around it, for reasons that are procurement and habit, not technology.

"A model that's three Elo points ahead of another is not a better product decision. It is noise dressed as a ranking."

Part III · What Maturity Demands

Picking a Model Stopped Being the Interesting Question

Back to that small piece of code. The reason it doesn’t hardcode a model name isn’t caution about the future — it’s an admission about the present. If the identity of the specific model were the load-bearing decision, hardcoding it would be correct: you’d want to know exactly, and change it deliberately, on your terms, after your own evaluation. Instead the code asks “what currently answers to this family name” and moves on, because for the workload it’s running — a daily research-and-write pass over a fixed set of sources — the difference between this month’s version and last month’s version is smaller than the difference between having the pipeline run reliably and not. That’s not a statement about Anthropic’s models specifically. I’d write the same code against any lab’s API today. A year ago I don’t think I would have.

What this means in practice, for anyone building rather than benchmarking: the durable decisions have moved up a layer. They’re not “which model” anymore; they’re “which relationship” — which provider’s rate limits and uptime you can live with, whose pricing model matches your usage shape, whose data-handling terms your legal team will actually sign, whose ecosystem (evals, guardrails, observability, agent frameworks) you’d rather build inside. Those are boring decisions compared to “which model is smartest,” and boring decisions are exactly what a mature market hands you, because it has already done the hard work of making the exciting decision not matter as much.

I don’t think this makes the underlying research boring, and I want to be careful not to imply that. Someone is still pushing the actual frontier — the labs racing on the hardest reasoning and agentic benchmarks are doing real, difficult work, and the compute and energy constraints I wrote about in The Age of AI haven’t gone anywhere. What’s changed is who that frontier work is for. Fewer and fewer of the decisions an ordinary team makes in a given month depend on it. The frontier keeps moving. Most of the market it used to matter for has already left the room and gone shopping for infrastructure instead.

That’s the honest version of “AI model maturity,” and it’s less dramatic than the version that gets posted daily — another leaderboard screenshot, another “model X beats model Y” post, indistinguishable from the one before it. If there’s a thesis buried in this piece, it’s that the interesting story left the leaderboard a while ago and nobody updated the headline.