Big Picture 4: Inside the Claude Model

Now the request enters the model. Inside there are four steps: tokenizer, Transformer, thinking and generation.

The boxes in this stage

• Tokenizer: the model does not understand text or images directly, so they are converted into tokens, the model's language.
• Transformer (Claude): the actual model. Using attention, it looks at the entire context at once and understands each token in relation to all the others.
• Haiku, Sonnet, Opus, Fable: the model family. Haiku is fast, Sonnet is balanced, Opus is the default choice and Fable is the max, the most capable. You pick the model to suit the job.
• Adaptive thinking: before answering, Claude can think internally. How much to think is something the model decides.
• Effort: sets the limit on thinking: low, medium, high, xhigh, max. Lower effort means fewer tokens, higher effort means more thinking and more tokens.
• Generation: the model predicts the next token, samples one from that prediction, and repeats until the answer is complete.
• The model runs nothing itself: it only writes tokens. If it needs a tool, it writes a tool_use block and someone outside executes it. This is the key point behind the agent loop in Big Picture 6.
• How the model was built: already done beforehand. 1. Pre-training, 2. Fine-tuning, 3. RLHF, 4. Constitutional AI / RLAIF, 5. Safety evals (red-teaming, system card, ASL). Big Picture 9 explains these.

For depth on tokens, probabilities and sampling, see the Model Internals document later in this Big Picture series.