This page exists because the rest of the guide is written to age slowly and this material does not age at all, it simply expires. It is revised most often and it is the page to be most skeptical of. Categories are stable. Names are not.
Read this for the shape of the market rather than for recommendations. The useful skill is recognizing which category a new product is in, because that tells you what question to ask about it.
The categories that persist
- Frontier models. A small number of providers, sold by the token, competing on reasoning and on enterprise terms. Ask about retention, training, and subprocessors before capability.
- Open weight models. Downloadable and self-hostable, trailing the frontier by a variable margin that is narrow on classification and extraction and wider on open-ended reasoning. The relevant question is chapter nine's arithmetic, not the leaderboard.
- Harnesses. The software wrapped around a model that gives it tools, files, and a loop. This category moves fastest and is the least worth committing to.
- Tool protocols. Standards for how a model connects to systems. Worth caring about because they reduce lock-in, and because a connector is a permission decision wearing a convenience costume.
- Retrieval and memory. Everything in chapter eight, sold as a product. The differentiator that matters is whether source permissions survive indexing. Most demos do not answer this.
- Orchestration. Running multi-step work reliably: queues, retries, state, human approvals. Frequently a job for infrastructure you already own.
- Evaluation. Tooling for chapter ten. Undersold, underbought, and the clearest signal of a serious team.
- Governance tooling. Registers, logging, policy enforcement, spend controls. Newest category, thinnest offerings, and mostly still assembled by hand.
What to ignore
Benchmark movements of a few percent. Anything described as a platform without naming which of the categories above it occupies. Products whose demo depends on a task nobody in your company performs. Agent frameworks that require you to model your business in their abstractions before you have automated a single task.
What is worth your attention instead
Changes to vendor terms on retention and training. Anything that makes permissions survive retrieval. Anything that makes evaluation cheaper. Published incident writeups, which teach more in twenty minutes than any vendor material. And the small number of independent people who publish their own failures, because the successes are all marketing and the failures are the curriculum.
Revision trail