The model landscape, minus the marketing

This one is a reading entry, not a building one. I spent the day reading around current AI models – how they actually differ when you use them, and what they cost to run. No code, no repo. The goal was to replace a vague sense of the landscape with something I can make decisions from, because the vague sense was all I had and it was not decision-grade.

Worth saying plainly: the apprenticeship does not touch any of this. The course gives me full-stack development, project process, the academic side of the discipline. Choosing a model or budgeting for one is somebody else’s problem right up until it is mine – and for anything I build that calls an API, it already is mine.

The ladder every vendor sells

Strip the branding and every major lab currently sells the same three-rung ladder: a flagship, a workhorse and a small fast model. Anthropic has Opus, Sonnet and Haiku; Google sells Pro and Flash tiers; OpenAI’s line-up has the same shape. On the current price lists the flagship rung lands around five dollars per million input tokens and twenty-five to thirty per million output, while the small rungs start at or under a dollar in with single-figure output pricing. Inside one vendor’s own ladder that is a spread of roughly four to twenty-five times between top and bottom rung, depending on who you buy from.

Two shapes inside the price card are worth more than any single number. First, output is priced at roughly four to six times input at every tier, so a model that waffles costs more per task than one that answers tersely, even when their headline prices look similar. Second, the reasoning models make that worse on purpose: thinking happens in tokens, those tokens are billed at output rates, and you do not see the thinking – just its share of the bill. One hard question on a reasoning model can cost several times the same question on a plain model, and nothing in the visible answer tells you which one you just paid for.

What the rungs actually buy

My honest takeaway from the practice side of the reading: what separates the rungs day to day is less headline intelligence than reliability. The cheap rung will do the easy version of a task fine and then fail strangely – mangle a tool call, miscount arguments, invent a field name. The expensive rung fails less often and, more usefully, fails legibly, so the fix fits in the next turn. Where a mechanical check exists – does it compile, does the test pass – cheap-model failures are cheap to catch, and that is most of the argument for running cheap models at scale.

Context windows run from a couple of hundred thousand tokens to a million depending on tier and vendor, but size is not free. Input is charged per token on every request, and a fat context resubmitted on every step of a loop stops being a convenience and becomes the design’s main cost.

The economics that change the design

The part I had never thought through is how much usage pattern, rather than model choice, drives the bill. An agent working over a codebase resends most of its context on every step. That is why prompt caching – the provider keeps your stable prefix warm and charges roughly a tenth of the normal input rate to reuse it – is structural rather than a micro-optimisation, and why every vendor sells a batch tier at around half price for work that can wait hours. None of that changes which model is cleverest. All of it changes what an agent-shaped design costs to actually run.

So the shape I ended the day with: a cheap fast model for the bulk steps where a mechanical check catches the failures, the expensive model for planning and for whatever the cheap one flags as hard, caching switched on, and cost reasoned about per task rather than per token. The specific prices here will date this piece inside a year – they were moving while I read. The shapes held across every price list I looked at, and they are what I will actually make decisions with.