Practical guide. Provider claims are attributed to the sources below. This page does not claim a new independent DARC benchmark.
For: Curious readers and developers making their first model choice
Before you start: No machine-learning background required
⚡ TL;DR — the 30-second version
- Model, context, tools and application policy are different layers.
- Token limits and pricing are provider-specific.
- A tool call requires validation and authorization outside the model.
A useful way to understand an LLM is to separate four things: the trained model, the context passed into a request, the tools the application exposes and the checks performed on the result. Confusing these layers produces common mistakes: assuming the model knows today's news, that a chat product's features exist in its API, or that a long conversation is durable memory.
This guide uses current model documentation to make the boundaries concrete. It does not require knowing how to train a transformer. The goal is to make better product and development decisions before you spend money or give an agent access to real systems.
“A model is one component of a system. Context supplies the evidence; tools supply actions; your application supplies the rules.”
Field guide / 01
01 / Tokens are billing and capacity units
Tokens are pieces of a model's representation, not a universal word count. The same text can tokenize differently across models; code, images and multilingual text make rules of thumb less reliable. Use the selected provider's tokenizer or usage report when a limit or cost matters.
Input and output can have different prices. Reasoning may consume billable output even when only a summary is shown. Cached input is another category; it reuses computation on a prefix, not a previously generated answer.
- → Record input, cached input, cache writes and output separately.
- → Measure the actual language and formats you will process.
- → Reserve output capacity instead of filling every request to its headline limit.
Field guide / 02
02 / Context is a working set, not a permanent brain
A request can include instructions, conversation history, retrieved passages, tool results and files. A large window permits more material, but does not establish that all of it is relevant, correct or equally likely to influence the answer. Too much contradictory context can make the output less useful.
Providers also describe limits differently. GPT-6 documentation quotes total context, while Gemini 3.8 Flash lists separate input and output limits. Store the actual constraint rather than comparing labels alone.
| Layer | Contains | Who owns the decision? |
|---|---|---|
| Model | Learned parameters and capabilities | Provider / model developer |
| Context | Instructions, evidence, recent state | Your application |
| Memory | Persisted facts selected for reuse | Your storage and retrieval policy |
| Tools | Explicit external operations | Your permission system |
Field guide / 03
03 / Tool use is an application contract
A model's tool request is a proposal, not a permission grant. Your application validates arguments, checks the caller's authority and decides whether to execute. The returned tool result becomes additional context for the next turn.
For a calendar assistant, reading availability and sending an invitation should be different operations. A prompt that says 'be careful' cannot replace that separation. Start with read-only tools and require approval before consequential writes.
Field guide / 04
04 / Run a small evidence exercise
Choose a question whose answer you can verify, then ask it with no sources, with one current source and with two conflicting sources. Observe whether the model distinguishes evidence from instructions and acknowledges missing information. Record the exact prompt and model ID.
Your deliverable is a one-page note: what context improved the answer, what made it worse, and which checks still needed a human. This is more transferable than memorizing a particular prompt incantation.
Before you ship
0/4Use this as a review aid. Checks reset when you leave this page.
Questions that come up
Does a bigger context window mean the model remembers me?
No. Durable memory requires storage and a policy for retrieving it. A large window is request capacity, not persistent personal memory.
Can a model know today's news without a search tool?
Do not assume so. Supply current evidence or use a retrieval/search system, and verify the cited sources.
Sources & verification
Last checked 23 Sept 2026. Publication dates and access conditions belong to the linked originals.
- 01GPT-6 Astra API model and pricingOpenAI↗
- 02Gemini 3.8 Flash model specificationsGoogle↗
- 03OpenAI — current prompt-caching behaviorOpenAI↗
- 04MCP — tools, permissions and validationModel Context Protocol↗
