Study hall
Glossary / FAQs
Plain-English words and questions for people who do not already live in AI. Same wording as the papers. This page grows when a new paper adds a word or a question.
From: The job picks the model.
* This is from Bench's own experience. It is not what any of the frontier models say about themselves. It will change as new models roll out, or as the ones already here get better.
Glossary
- Orchestra
- A few models taking turns, each doing the part they're best at. One writes the plan. One builds. One checks. Not three chat windows with no jobs.
- Frontier model
- One of the current top models from the companies that make them: Anthropic, OpenAI, or xAI.
- Opus
- The Anthropic model for a normal Tuesday. Nearly as good as Fable, costs less, handles most of the work.
- Fable
- Anthropic's top model. Expensive. Save it for building a pipeline or a big multi-day project. Don't use it to knock out one small job.
- Pipeline
- A system that keeps doing the work. Building a video machine for every job is a pipeline. Asking for one video right now is not.
- One-shot
- One quick job. "Check this email." "Make me one video." Then you're done. Not a project that has to keep running.
- Sol
- OpenAI / Codex's top model. Named for the sun. Not an agent's soul — that's the rules the agent lives by.
- Tokens / budget
- The monthly thinking allowance. Like minutes on a phone plan. Bigger jobs use it up faster.
- Adversarial
- A second model looks for mistakes before you trust the work. Don't call it done because the first one said so.
FAQs
Which model should I start with?
When Bench builds, Grok first. When Bench starts a new client on their Mac, one chat. Opus is the model for a normal Tuesday.
What is an orchestra?
A sequence of models, each with one job. Not three chat windows. At Bench, Grok starts, Claude plans, Codex builds, and a second model checks.
When should I use Claude, Codex, and Grok together?
When the job needs a photo, a video, and a page. You do not need three companies to write an email.
What is the difference between Claude and Codex?
Claude explores and designs. Codex builds, checks, and talks to customers. Do not put Claude in front of a customer. Do not ask Codex to figure out what to build.
When do I use Fable vs Opus?
Opus is the model for a normal Tuesday. Fable is for building a pipeline — a system that keeps doing the work, like a video machine for every job. Don't use Fable to knock out one quick job. That's a one-shot. Use Opus for that.
What is a one-shot?
One quick job, then you're done. Check an email. Make one video. Not a system that has to keep running.
What is a pipeline?
A system that keeps doing the work. Not "make me one video right now." A machine that can make the next ones too.
What is Sol?
Sol is OpenAI and Codex's top model — the sun. It is not an agent's soul. Soul is the rules the agent lives by.