System prompts and developer instructions enter first. The counter is already moving.
The Conversational Loop
Every turn rebuilds the packet. Rules, tools, files, history, reasoning, and the visible answer all stack before the next send starts the loop again.
Skills, schemas, and connector descriptions join the prefix as reusable behavior.
Files, images, transcripts, and retrieved snippets become tokenized context.
Previous messages return so the model can remember what happened last turn.
Hidden planning and verification adds output tokens before an answer appears.
The visible response lands on top of the stack, but the loop is not finished.
The answer becomes context. On the next send, it is loaded back into the packet.
Cost stacks as the loop repeats
rules + tools + files + history + reasoning + answer, then history again
Run the Cost Loop
The visible message is only one layer. System prompts, skills, files, multimodal input, cache, and reasoning are rebuilt into the next paid turn.
Context: 0
Cost: €0.0000
System prompt
0 prompt tokens
Public prompt example
Model selection loads a provider prompt cartridge.
The system prompt is hidden setup. It tells the model how to behave before the user message is read, and those tokens sit inside the input context.
Cost impact
Loaded every turn
Loading prompt preview...
Skills files are reusable instructions and tool playbooks. Loading more of them gives the assistant abilities, but each file adds prompt tokens.
Cost impact
0 tokens
Multimodal input means screenshots, photos, or visual references. The model reads them as visual tokens before it can answer.
Cost impact
0 tokens
Uploaded files are parsed into pages, chunks, or retrieved snippets. Useful parts still need to fit into the prompt packet.
Cost impact
0 tokens
Previous turns loaded: 0
Model work area
Reasoning and cache controls
Reasoning is hidden model work: planning, checking, and comparing candidate paths before the visible answer is written.
Selected hidden work
0 tokens
Reasoning controller
1.3x
Higher effort spends more hidden output tokens to explore paths, test assumptions, and reduce shallow answers.
Cache state
Prefix pricing
Cache means a repeated prompt prefix can be reused at a lower input rate when the next loop sends similar instructions and context.
Prompt, messages, files, and image tokens.
Hidden tokens before the visible answer.
The response tokens on screen.
Selected cache discount from the repeated prefix.
Calculated from used tokens.
Anthropic
Claude Sonnet 5 rates
Rates are per 1M tokens and converted to euros. Reasoning is counted as output.
- Input
- €0.00
- Cached
- €0.00
- Output
- €0.00
- Breakdown
- €0.0000
Pricing snapshot checked July 8, 2026: OpenAI, Anthropic, Gemini. USD prices converted with ECB rate: €1 = $1.1404.
Rules for reading token cost
Use these mental shortcuts when a chat, agent, or multimodal workflow suddenly feels expensive.
System prompts and skill files are useful, but every loaded instruction needs room.
Long conversations often re-send enough prior context to keep the next answer coherent.
Uploads must be parsed, chunked, retrieved, or summarized before they can help.
Images and other modalities become model-readable units, not just tiny thumbnails.
More effort can improve answers, but it adds hidden generated tokens.
Caching can lower the bill for repeated prefixes. It does not erase the context.