Definition
Model spec is a model spec is a public document in which an AI lab describes how its models are supposed to behave, including their goals, limits, and how they weigh conflicting instructions.
A model spec is a written statement from an AI lab about how its models should behave. It covers what the model should try to do, what it should refuse, how it should handle conflicting instructions, and how it should act when no rule fits.
Most behavior in a chat model comes from training, which is hard for outsiders to inspect. A model spec puts the intended behavior in plain language. That gives users, developers and researchers something to point to when a model does something unexpected, and it gives the lab a document to argue with, revise and be held to.
OpenAI published a document it calls the Model Spec in 2024 and has revised it since. Anthropic publishes a similar document, which it calls a constitution, that describes the values and priorities it trains its Claude models toward. The names differ, but the purpose is the same.
A spec generally includes:
Think of it as an employee handbook for a very fast new hire. It does not list every situation, but it explains the priorities well enough that the hire can make a sensible call.
A published spec states intent. It does not prove the model follows it. Labs check this with evals, red teaming and published system cards, and gaps between the written behavior and the actual behavior are common. When you evaluate a model for a real product, test the behavior you depend on instead of trusting the document.
If you build on a model, the spec tells you what the model is designed to do with your instructions, where its limits are and what you can customize. If you only use AI, it explains why a model pushes back in some places and goes along in others.
A model spec is part of AI alignment in practice: it is the lab's written answer to the question "aligned with what?" It shapes post-training, where models are tuned toward the described behavior, and it informs the AI guardrails builders add on top.
It is a document where an AI lab explains in plain language how its models are meant to behave, including goals, limits and how to resolve conflicting instructions.
Not always. A spec states the intended behavior, and labs use testing to measure how closely the real model matches it.
Mistral opened a public preview of Mistral Large 4 on Tuesday, a mixture-of-experts model with about 1 trillion parameters and 49 billion active, and promised weights by the end of October. The API is a preview. The weights are the launch.
Vals AI built 50 web apps with GPT 6 Sol and Claude Opus 5.5 and found agent teams cost 1.8 to 5.1 times as much as one agent, with one significant gain in four setups. Try higher effort on a single agent before you pay for a team.
Anthropic released Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 output for prompts up to 100,000 tokens, a tenth of Haiku 4.5's price. Above that line the deal changes, so check your prompt lengths before you celebrate.
Figma released Specs in open beta for paid plans on Oct. 15, a doc on the canvas that its own agent and outside coding agents can read through MCP. Agents fail on missing context, and Figma wants to be where that context lives.
Apple's new MacBook Pro with M5 Pro or M5 Max adds a touch OLED screen and up to 128GB of memory, from $2,899, shipping Nov. 6. Apple says M5 Max runs models with hundreds of billions of parameters locally but gives no speeds, so wait for real numbers.