Frontier AI and Mainstream LLMs

7.1  What “frontier” means

Frontier models are the small set at the capability edge — the handful that define what is currently possible at all. Below them sits a much larger tier of mainstream models: cheaper, faster, smaller, and entirely adequate for most work most of the time.

The tiers matter because they are priced differently by one to two orders of magnitude, and because the capability gap between them is task-dependent rather than uniform. On summarising a document you supplied, a mainstream model is frequently indistinguishable from a frontier one. On a long multi-step task with tools, the gap is large and obvious.

“Frontier” is also a regulatory term of art, used in policy documents to designate the models whose capabilities warrant particular oversight. That is a different use of the same word, and worth not conflating.

7.2  Benchmark saturation, and why it broke the question

A benchmark is useful only while it discriminates. Once most leading models score above 90%, the remaining spread is measuring noise, contamination, and formatting quirks as much as capability.

That is where the field is. MMLU — the broad knowledge-and-reasoning benchmark that dominated model comparisons for years — is effectively uninformative in 2026, with most top models above 90%.1 Attention shifted to harder evaluations: GPQA Diamond for graduate-level reasoning, SWE-bench Verified for real software engineering tasks, Humanity's Last Exam, and arena-style human-preference ratings.2

Those are saturating in turn. As of August 2026 the top GPQA Diamond scores sit in the low nineties and the benchmark is approaching saturation at the very top, though it still separates models in the 60–90% band. SWE-bench Verified is largely saturated, with the top five frontier models between roughly 93% and 97% and the meaningful differentiation moving to the harder SWE-bench Pro.2 On genuinely unsaturated evaluations — ARC-AGI-2, aimed at novel-problem reasoning — scores remain low.

Two structural forces explain the saturation. Contamination: benchmarks published on the open web end up in training data, so a score partly measures exposure rather than capability. Optimisation pressure: when a benchmark becomes the number vendors compete on, effort flows into the benchmark — an instance of Goodhart's law, that a measure which becomes a target ceases to be a good measure. Lesson 6's account of hallucination is the same dynamic seen from a different angle: the scoring rule shapes the system.

FIG-07-01

Saturation, drawn

A small-multiples chart: four benchmarks over time (MMLU, GPQA Diamond, SWE-bench Verified, ARC-AGI-2), each showing top-model score climbing toward a ceiling line at 100%. MMLU flattens against the ceiling early and the plotted models bunch into an indistinguishable cluster; GPQA and SWE-bench are approaching it with the cluster tightening; ARC-AGI-2 sits low with points still well separated. Annotate the bunched regions with no longer discriminates. The visual argument is that a benchmark dies by being answered.

Aspect 16:9 · 3200×1800 px · Brand palette: navy #00040c, silver-blue #a8c0d8, column gold #a87848 · Style: small-multiples line chart, axis labels legible at 90 mm print width

7.3  The 2026 shape of the market

Two facts, both dated, both worth putting on screen with the date visible.

The top is compressed. As of August 2026 the leading models on composite benchmarks sit within roughly half a point of each other, while real-world agentic capability continues to diverge more than the composite scores suggest.3 The practical consequence: a benchmark table cannot tell you what you want to know, because the thing that still differs between these models is not what the benchmark measures.

The open-weight tier has closed much of the gap. Lesson 8 covers this properly, but it belongs in the market picture: the assumption that frontier capability means a closed commercial API no longer holds cleanly.3

7.4  Multi-model routing is the production norm

Organisations have largely stopped standardising on one model. The 2026 production pattern is routing: cheap, fast models for high-volume work; premium models for final output or hard reasoning; task-specific selection in between.14

The reason is economic rather than technical. If a mainstream model handles 80% of your volume at a fraction of the cost with no perceptible quality loss, routing that 80% away from the frontier model is simply free money. The engineering required is a classifier and a fallback, which is much less work than the phrase “multi-model architecture” implies.

So the viewer's takeaway, and the sentence to remember from this lesson:

All three qualifiers are load-bearing. Drop the task and you buy capability you never use. Drop the price and you overspend by an order of magnitude. Drop the quarter and you are running on a decision that expired.

Demonstration protocol

1. Pick three tasks at different difficulties: a formatting or extraction job, a summarisation job, and one genuinely hard multi-step reasoning job.
2. Run all three against three models at three price tiers. Show raw outputs.
3. Show that the cheap model is fine for two of the three. This is the whole demonstration and it lands harder than any explanation.
4. Put the prices on screen alongside the outputs, with the date stamp visible.
5. Show a leaderboard, then show a second leaderboard that disagrees with it. Sixty seconds, and it inoculates the viewer against every ranking they will see for the next year.