When Intelligence Is Abundant

When intelligence is abundant

Every few weeks a smarter model ships, and the price of last month’s intelligence drops. If that keeps up, raw model capability stops being an advantage, because everyone can rent the same one. The edge moves to what stays scarce. Here’s my view of what that is, and how a team can get ready now.

By Nathan · October 2026 · About 10 minutes

The problem

September 2026 brought GPT-6 Astra, Claude Fable 5.1, Claude Opus 5.5 and Gemini 4 Argon, each claiming the top of some leaderboard. Opus 5.5 arrived cheaper than the model it replaced. Rumors about the next generation are already circulating, though none of it is confirmed.

Most teams respond by chasing each release: rewrite the prompts, rerun the demos, argue about which model is best this month. That’s a treadmill. The more useful question is what a much smarter model would still need from you, because that’s the work that won’t be thrown away next month.

4new frontier models released or announced by three labs in September 2026

−20%price of Claude Opus 5.5 against Opus 5, at $4 / $20 per million tokens instead of $5 / $25

69×cheaper per right answer for a small model with a code tool than the top model asked plainly, in my own test

How I’d know a team is ready

Days to evaluate a new model on your own work, not weeks
Share of AI effort going into tools, tests and data, rather than prompt tuning
Cost per right answer, tracked by task, so each job runs on the cheapest model that does it well
Clear rules for what a model may do alone, written before you need them

Why I wrote it

I ran an experiment to see which prompting habits still help as models get stronger: What Smarter Models Make Obsolete. The prompt tricks faded; tools, verification and tests didn’t. This essay is the thought experiment that goes past what one test can show.

The ideaTreat intelligence as a commodity that gets cheaper every quarter. Invest in the things it can’t supply for itself: knowing exactly what you want, owning the ground truth, giving it access and tools, deciding what it may do, and paying only for as much intelligence as each job needs.

1Where the work moves

Think about how a team spends its AI effort today. Much of it goes into compensating for the model: writing long prompts, catching its mistakes, rephrasing until it behaves. As models improve, that compensation shrinks. What grows is everything around the model.

How a team’s AI effort might shift

My estimates, not measurements. Shares of a team’s AI-related effort; toggle to compare.

Compensating for the modelBuilding what lasts

A smarter model needs fewer instructions. It does not need fewer facts, fewer tools or less accountability.

2What stays scarce

Five things a model can’t generate for itself, however capable it becomes. Each one has a measured hint from the builds on this site.

1 · INTENTKnowing what you wantA brilliant model asked a vague question gives a brilliant answer to the wrong question. The scarce skill is a precise definition of done: who it’s for, what good looks like, what’s off limits.

Hint: in my test, a long expert prompt never beat a plain one. What the model needed was the problem stated clearly, not coaching.

2 · GROUND TRUTHOwning the answers that prove itA graded set of your own real cases is what tells you whether a new model is better for you, not on a leaderboard. Whoever owns the best test set adopts new models fastest.

Hint: calibrating on labeled data was the biggest single gain in Second Opinion.

3 · ACCESSTools, data and permissionIntelligence without access to your systems is advice. The edge is in clean data, well-built tools and the permission to use them: the hands the model works with.

Hint: one code tool took the small model from 13 to 16 of 16, matching the top model.

4 · JUDGMENTDeciding what it may do aloneAs models act on their own, someone must decide where they can, check what they did, and own the result. Accountability doesn’t get cheaper with intelligence.

Hint: The Handoff Line sets that line from measured error rates; The Governor sketches the controls always-on agents will need.

5 · ECONOMICSPaying for the right amountWhen the top model can do everything, the question is what each job is worth. Routing easy work to small models and tools, and saving the frontier for the hard part, compounds across millions of calls.

Hint: maximum thinking matched medium at about 4× the cost per right answer in How Hard Should It Think?

AND · TRUSTShowing your workCustomers and regulators will trust answers they can check. Quotes tied to sources, confidence that means something, and a record of who approved what will matter more as answers get harder to second-guess.

Hint: How Sure Is Sure? found a confidence number is only useful once you know what it actually tracks.

3Prompt craft becomes spec writing

The prompt doesn’t disappear; it changes job. Today’s prompts are full of technique: personas, rules, worked reasoning. Tomorrow’s look more like a good brief to a senior colleague: the goal, the constraints, the materials, the definition of done, and how the work will be checked.

Examples survive, but differently. In my test, two worked examples no longer improved accuracy, yet they cut the top model’s output by 44%, because it copied their format. Examples become a contract for the shape of the answer rather than a lesson in how to think.

Context matters more, not less. A smarter model can use far more of what you give it, so the work shifts from trimming context to curating it: the right documents, current and trustworthy, with the stale ones taken out.

4Token efficiency still matters

Cheaper intelligence doesn’t make waste free. Smarter models tend to think longer when asked hard questions, and agents call models thousands of times a day. A team that routes well spends a fraction of what a team that sends everything to the top model spends, for the same results.

Three habits held up in my tests. Let code do exact work, because a program is short and right where reasoning is long and occasionally wrong. Use the lowest thinking level that holds accuracy, because the top settings mostly bought cost. And measure cost per right answer, not cost per call, because a cheap model that fails costs more than an expensive one that doesn’t.

5How to prepare in the next 90 days

DAYS 1–30

Build the test set

Collect 50–100 real cases from your own work, with known right answers.
Write a short definition of done for each AI use case.
Track cost per right answer for what you run today.

DAYS 31–60

Give models hands

Wrap the exact work, such as calculations, lookups and date math, in tools.
Clean and document the data models will read.
Route each task to the smallest model that passes the test set.

DAYS 61–90

Decide the limits

Write down what models may do alone, what needs approval, and who owns each.
Prune prompts: drop personas and tricks the test set says don’t help.
Practice a model swap: how fast can you evaluate and adopt a new one?

The team that can test a new model on its own work in a day will get every future model’s gains first.

Choices

Bet on the trend, not the releaseModels will keep improving and getting cheaper. Work that assumes otherwise is the work most likely to be thrown away.

Invest around the modelTests, tools, data and rules transfer to every future model. Prompt tuning transfers to none of them.

Keep people accountableMore capable models make the decision about what they may do more important, not less.

Where this can go wrong

This is a forecast. Progress could slow, prices could rise, or new models could bring new quirks that need new techniques. My experiment used exact-answer problems, which favor tools; open-ended work like writing and strategy may keep rewarding careful prompting longer. And “intelligence is abundant” is true only for those with access: the most capable model this month, Gemini 4 Argon, is open only to vetted security teams.