Every seat has a name and a model.

Where each seat shows up

The mode decides who sits.

The models above do not all run on every question. The mode you choose decides which seats are convened.

Solo

Chair answers directly.

Review

Generator drafts and revises; Chair critiques and judges.

Debate

Generator argues the Affirmative, Sceptic the Negative; Chair decides.

Moot

Generator proposes, Sceptic challenges, Specialist checks; Chair rules.

The specifications

What each model brings.

Context is how much of your question, documents and the panel’s own working a model can hold at once. Latency is how long it waits before speaking, throughput how fast it speaks once it starts.

Model	Category	Context	Latency	Throughput	Capabilities
GPT-5.6 TerraOpenAI	plus / generator	1.05M	1.8s	147tps	Extended reasoning (to max) Reads images Reads PDFs and documents Calls tools Searches the web
GLM 5.2Z.ai	light / sceptic	1M	2.6s	139tps	Extended reasoning (to xhigh) Calls tools
Claude Sonnet 5Anthropic	plus, pro / sceptic, chair	1M	3.4s	112tps	Extended reasoning (to xhigh) Reads images Reads PDFs and documents Calls tools Searches the web
Claude Opus 4.8Anthropic	pro / chair	1M	3.5s	92tps	Extended reasoning (to xhigh) Reads images Reads PDFs and documents Calls tools Searches the web
GPT-5.6 SolOpenAI	pro / generator	1.05M	4.2s	70tps	Extended reasoning (to max) Reads images Reads PDFs and documents Calls tools Searches the web
DeepSeek V4 ProDeepSeek	light / generator	1.05M	2.5s	69tps	Extended reasoning (to xhigh) Calls tools
Grok 4.5xAI	plus, pro / specialist	500K	1.6s	59tps	Extended reasoning (to high) Reads images Reads PDFs and documents Calls tools Searches the web
Claude Sonnet 4.6Anthropic	light / chair	1M	1.5s	56tps	Extended reasoning (to high) Reads images Reads PDFs and documents Calls tools Searches the web
Mistral Large 3Mistral	light / specialist	256K	0.6s	56tps	Reads images Calls tools

Context and capabilities come from the Vercel AI Gateway model catalogue. Latency is p50 time to first token and throughput is p50 output tokens per second, both as published by Vercel from live traffic on the same gateway every Dr Moot run goes through. Both sources were read on 12 August 2026; they are measurements of a live system and they move. A dash means nothing is published, which we prefer to an estimate. Speed is not quality: the slowest seat on this page is often the one that catches the error.

Straight from the makers

Every lab says how its model should be driven.

Each maker publishes its own guidance on how to ask - what to spell out, what to leave alone, where it wants structure instead of prose. We apply those conventions strictly, so a seat holds its persona whichever model is sitting in it: one job per seat, carried in that seat’s own system prompt and restated every time it speaks; ballots and critiques returned as structure rather than prose about structure; and anything a model cannot accept stripped before the call instead of left for it to trip over.

OpenAI

A stated instruction hierarchy, explicit reasoning effort, and separate guidance for the GPT-5.6 line on agentic work.

Generator on Plus and Pro

Prompt engineering(OpenAI documentation, opens in a new tab)GPT-5.6 prompting(OpenAI documentation, opens in a new tab)

Anthropic

Be clear and direct, give context rather than cleverness, use worked examples - with its own prompting notes per model.

Chair on every plan, Sceptic on Plus and Pro

Prompting best practices(Anthropic documentation, opens in a new tab)

xAI

Grok 4.5's API notes: what the reasoning effort settings do, and the details that matter when it is asked to check someone else's work.

Specialist on Plus and Pro

Grok 4.5(xAI documentation, opens in a new tab)

DeepSeek

Strict JSON output, documented as its own mode - the difference between a seat returning a ballot and a seat returning prose about a ballot.

Generator on Light

JSON output(DeepSeek documentation, opens in a new tab)

Z.ai

GLM-5's own model card: what it is positioned for, which modalities it takes, and where its context and output limits actually sit.

Sceptic on Light

GLM-5(Z.ai documentation, opens in a new tab)

Mistral

A system prompt with a stated purpose, few-shot examples, structured output - and a long list of what not to do, including subjective scales and asking a model to count.

Specialist on Light

Prompt engineering(Mistral documentation, opens in a new tab)