MODELS — which one, and who's telling the truth
What this section is
The Foundation layer explains how a model gets made — pretraining, alignment, quantization, the whole manufacturing floor. This section answers the other question, the one you actually have on a Tuesday afternoon: which one do I use, and how much of what I'm being told about it is true?
Those are different problems. Knowing how a transformer works tells you almost nothing about whether this quarter's flagship is worth switching to. That decision runs on pricing, context limits, licensing, latency, refusal behavior, deprecation risk, and benchmark claims of wildly varying honesty. It is a purchasing decision dressed up as a technical one.
So: four dossiers, one per family. Each is a buying-and-using guide — what the family is good at, where it isn't, what it costs, and when to reach for it. This page tells you which dossier to open.
The four families
Claude (Anthropic). The Opus / Sonnet / Haiku tiering — flagship, workhorse, fast-and-cheap. Reputation built on long-context work, code, and agentic reliability. Read the dossier if your work is code, long documents, or agents that have to stay coherent across many steps.
ChatGPT / OpenAI. The incumbent by mindshare, and the broadest ecosystem — the default enterprise checkbox. General chat, vision, audio, plus a reasoning tier that spends inference-time compute thinking and prices accordingly. Read the dossier if you want maximum tooling and integrations, or you're evaluating the reasoning tier's cost profile before committing.
Gemini (Google). The vertical-integration play: Google's own silicon, its own cloud, distribution through Search, Android and Workspace. The technical signature is context length and native multimodality, with aggressive price-performance in the lighter tier. Read the dossier for very long documents and video, or if you're already committed to Google's cloud.
The open-weight ecosystem. Not a vendor — a category. Weights you download and run: on your own hardware, on rented GPUs, or through a host that charges by the token. You trade some frontier capability for control, privacy, and a cost structure that doesn't move when someone else's pricing page does. Read the dossier if data can't leave your building, if you're at volume where per-token pricing stops making sense, or if you want a floor under your costs that no vendor can raise.
How to actually choose
Start from the job, not the leaderboard. The leaderboard is a marketing surface; your job is not.
Name the task first. Coding, deep reasoning, long-context retrieval, cheap bulk classification, and vision are five different procurement problems with five different winners. A model that tops a general intelligence index can lose badly on your specific workload, and frequently does.
Then set the constraint. Latency budget, privacy requirement, monthly ceiling, whether an outage takes your product down with it. Constraints eliminate more candidates than capability does.
Then price the whole thing, not the token. Output tokens usually cost several times input tokens. Reasoning tiers bill for the thinking you never see. Long-context surcharges can apply to every token in a request, not just the ones past the threshold. The sticker price and the invoice are different documents.
Then hold your choice loosely. The capability lead changes hands release by release, and models get deprecated on the vendor's schedule, not yours. Build model-agnostic. Assume the thing you picked will be retired, and that the migration will land at an inconvenient moment.
For the task-by-task version of this — top pick, runner-up, and budget pick per job — see Models by Task.
Where it breaks
This is the layer with the widest gap between what's claimed and what's true, so it earns its own warning.
Benchmarks measure benchmarks. A score tells you how a model performed on a fixed set of problems under conditions the publisher chose. It does not tell you how it performs on your work. Treat every number as indicative, never as gospel — and prefer third-party evaluations to vendor announcements, because the vendor picked which numbers to publish.
Contamination is routine and largely unmeasurable. When the test set is on the public internet and the model trained on the public internet, a high score may mean the model has seen the answers. Nobody can fully rule this out from the outside, and the labs' own contamination checks are self-graded homework.
Cherry-picking is the house style. Launch posts compare against a competitor's older model, or a carefully chosen subset, or on a metric where the result flatters. Read what was not compared.
The demo-to-production gap is real. Impressive launch demos are curated. Sustained performance across thousands of messy real requests — including the ones with bad formatting, hostile input, and ambiguity — is a different measurement, and almost nobody publishes it.
And the naming is deliberately slippery. Version numbers, tiers, "fast" variants and reasoning modes are product decisions, not technical ones. Two things with nearly the same name can differ substantially in cost and capability.
·
What we will and won't tell you
Every pricing table and specification in the four dossiers carries a verification tag until an editor has confirmed it against the vendor's own live page. Figures that haven't cleared that check are marked, visibly, on the page. We would rather show you an unverified figure flagged as unverified than launder it into a fact.
Where a number is contestable we publish the conservative end and show the range. The market moves monthly; these tables are re-verified on a standing cadence, and every one is date-stamped. If a figure looks stale, it probably is — check the vendor page and tell us.
The dossiers
- Claude (Anthropic) — the coding and agent workhorse
- ChatGPT / OpenAI — the incumbent and its reasoning tier
- Gemini (Google) — context length and vertical integration
- The open-weight ecosystem — control, privacy, and a cost floor
Where this connects
Foundation — how these artifacts get built in the first place: pretraining, alignment, quantization, open versus closed weights. Read it if the why behind the differences matters to you.
Playbook — how to actually use what you pick: prompting, best practices, integrations, IDEs.
Resources — the full tool and model directory, broader than these four families.