A post by Calvinfo hit 682 points on Hacker News in late August with a precise economic observation: "With the previous generation of models (Sonnet class), you'd spend ~$1 to get anywhere. With GPT-5.6-luna, the results are pretty decent, and the average cost is ~$0.10. Now we're talking."
He's right. Small models have arrived. And the consequence he didn't fully draw out is that the routing problem — which model for which request — just got an order of magnitude harder.
The economics are straightforward. GPT-5.6 Luna is priced at $0.20/$1.20 per million tokens (input/output). A typical research task — a few thousand input tokens, a thousand output tokens — costs roughly $0.10. The same task on a Sonnet-class model cost ~$1. The quality is "pretty decent" — not frontier, but sufficient for a wide range of tasks that previously required Sonnet-class spending.
But the story isn't "small models are cheap now." That's been true since GPT-4-mini. The story is that small models crossed the quality threshold where they're viable for production workloads. There's a difference between "cheap and bad" and "cheap and good enough." Luna, GLM-5.3-Flash, DeepSeek V4 Flash, and Gemini 3.7 Flash are all in the latter category.
Here's what the model catalog looked like in early 2025:
Five models. Two quality tiers (frontier and everything else). One provider per model. A human could pick the right model for an application and configure it once.
Here's what the catalog looks like in September 2026:
| Tier | Models | Price Range (per MTok) |
|---|---|---|
| Frontier | GPT-5.6 Sol, Claude Opus 5, Gemini 3.7 Pro | $2.50–$5 / $12–$25 |
| Frontier-adjacent | GLM-5.3 (open-weight), Kimi K3, DeepSeek V4 Pro, Qwen 3.8 Max | $0.75–$3 / $3.75–$15 |
| Cost-efficient | GLM-5.3-Flash, DeepSeek V4 Flash, Gemini 3.7 Flash, Claude Sonnet 5 (intro) | $0.20–$2 / $1.20–$10 |
| Ultra-low-cost | GPT-5.6 Luna, Qwen 3.8 27B, Llama 4 variants | $0.04–$0.20 / $0.24–$1.20 |
That's roughly 15 models across 4 tiers, each available from 1-4 providers. The cross-product of models × providers × tasks × quality bars is not 15 routing decisions. It's potentially hundreds of distinct routing paths, each with a different cost-quality tradeoff.
The FinOps Foundation's 2026 data: enterprises with $7M+ annual AI budgets are spending the majority on frontier models. The 4,500× pricing spread between the cheapest production LLMs ($0.04/MTok) and the most expensive frontier reasoning models ($180/MTok) means that organizations defaulting to "the best model for everything" are paying up to 4,500× more than necessary for tasks that cheaper models handle adequately.
The counterargument is always: "But the frontier model does the hard tasks better." That's true. The question is: what fraction of your tasks are hard?
If 70% of your requests are simple, 20% are standard, and 10% are hard, then routing everything to the frontier model costs you 10× more than routing by task complexity. The math is not subtle.
When there were 5 models, routing intelligence was a nice-to-have. When there are 50, routing intelligence is essential. Here's why:
The number of routing decisions scales linearly with model count. Every new model in the catalog creates a new comparison: "is this model better than the current default for this task type, at a lower cost?" With 5 models, that's 5 comparisons. With 50, it's 50. With 50 models × 4 providers × 10 task types, it's 2,000 distinct routing cells — each requiring a quality score and a cost calculation.
The savings per routing decision scales with the pricing spread. The wider the price gap between the cheapest and most expensive viable model, the more each correct routing decision saves. The pricing spread has grown from ~10× in 2024 to 4,500× in 2026. Each routing decision is worth 10-100× more than it was two years ago.
The quality variance between models on the same task scales with model diversity. GLM-5.3 might handle Python refactoring at parity with GPT-5.6 Sol but fall behind on Rust. Luna might handle text summarization at parity with Sonnet 5 but fail on complex code review. The per-task quality matrix is not a single number — it's a matrix of model × task quality scores that changes as models are updated.
Three practical consequences for engineering and finance leaders:
If your routing layer is configured once ("send coding tasks to Claude, text tasks to GPT-4-mini"), it's already stale. New models arrive weekly. Existing models get updated. The quality of a model on your specific task distribution can only be determined by evaluating it on your traffic — not by reading benchmark tables.
The infrastructure that makes this possible: shadow quality scoring on every routed request, coverage data per model per task, and a routing layer that consumes quality signals in real time. This is not a dashboard. It's a feedback loop.
The $0.10 task exists. The $0.01 task exists. The $0.001 task exists. But if your application sends every request to the same model — even a relatively cheap one — you're overpaying on the tasks that could have gone cheaper. The savings from small models are not automatic. They require a routing layer that can distinguish a $0.10 task from a $0.001 task and send each to the appropriate model.
If your 2027 AI budget projection assumes "current spend × growth rate," it's wrong. The cost per task is dropping by 10× per generation. If your routing layer is capturing that decline, your 2027 spend should be flat or declining — not growing. If it's not, you're paying frontier prices for tasks that frontier-adjacent models can handle, and the gap between your bill and your optimal bill is growing every month a new model ships.
"Small models have arrived" is not a victory lap. It's the starting gun for the routing era. When models were expensive and few, the cost of picking the wrong one was bounded. When models are cheap and many, the cost of not routing at all — of sending every request to the same model regardless of complexity — is the most expensive decision you can make.
The teams that build routing intelligence into their stack now will see their cost-per-task curve bend downward as the catalog grows. The teams that don't will see their bills keep climbing, even as the underlying cost of intelligence collapses.
Trimio routes every LLM request to the cheapest model that passes the quality bar — continuously evaluated on your real traffic. As small models arrive, the routing layer captures the savings automatically. See how it works.