To predict what a real request costs, you need one organizing fact: everything in the request counts toward the context window and is billed as input — system prompt, all messages including tool results and images, tool definitions, and, with caching, all three usage categories (input_tokens, cache_read_input_tokens, cache_creation_input_tokens). The output side includes any internal reasoning: thinking tokens are billed as output tokens. The full cost model is just those pieces, each multiplied by its own rate.
The components, priced
Text tokens bill at the model's base rates — Claude Opus 4.8, used throughout this article, is $5 per million input and $25 per million output tokens.
Images bill as input tokens by geometry: ceil(width/28) × ceil(height/28) visual tokens, so a 1000x1000 image is 1,296 tokens. High-resolution models (Fable 5, Opus 4.8/4.7, Sonnet 5) accept up to 2576 pixels on the long edge and up to 4,784 tokens per image. PDFs bill twice per page: extracted text (typically 1,500–3,000 tokens per page) plus each page as an image under the same formula.
Tool schemas and overhead bill as input on every request that carries them. Your own tool definitions cost whatever they tokenize to, and merely enabling tools adds a documented system-prompt overhead — on Opus 4.8, 290 tokens with tool_choice of auto/none, 410 with any/tool. Anthropic-defined tools have published fixed costs too: the bash tool adds 325 input tokens on Opus 4.8, the Claude 4.x text editor tool adds 700. Server tools can carry fees beyond tokens — web search is $10 per 1,000 searches on top of standard token costs.
Cache layers reprice whatever they cover: 1.25x base input to write with the 5-minute TTL (2x for 1-hour), 0.1x to read. On Opus 4.8 that is $6.25/$10 per million to write and $0.50 per million to read.
A worked example
A document-review request on Claude Opus 4.8: a 6,000-token stable prefix (system prompt plus tool definitions) under a 5-minute cache breakpoint; tools enabled with tool_choice: auto (290 overhead tokens); one 1000x1000 image (1,296 tokens); a 500-token user message; an 800-token response. All prices are the listed Opus 4.8 rates.
| Component | Tokens | Rate per MTok | First request | Cached request |
|---|---|---|---|---|
| Prefix — cache write / read | 6,000 | $6.25 / $0.50 | $0.0375 | $0.0030 |
| Tool overhead + image + user text | 290 + 1,296 + 500 | $5.00 | $0.0104 | $0.0104 |
| Output | 800 | $25.00 | $0.0200 | $0.0200 |
| Total | ≈ $0.0679 | ≈ $0.0334 |
Two lessons fall out. First, the cache halves this particular request, but only the stable prefix benefits — the image and user text sit after the breakpoint and always pay full price, which is why image resizing matters even in well-cached systems. Second, 800 output tokens cost more than 2,000 uncached input tokens; output discipline (concise-format instructions, sensible max_tokens, lower effort where appropriate) is frequently the biggest lever in the whole model.
usage object rather than your assumptions. Multipliers stack — batch (50% off) and data-residency (1.1x for inference_geo: "us") apply on top.Platform wrinkles
The token math is identical on all four platforms at matching list prices, with edges worth knowing: on Bedrock and Google Cloud, regional endpoints carry a 10% premium over global ones (Sonnet 4.5 and later models); on Bedrock's Converse API, full visual PDF understanding requires citations enabled, which materially changes PDF token counts; and Claude Platform on AWS and Foundry meter the same dollar amounts through Claude Consumption Units (CCUs) at $0.01 each. Also budget context, not just dollars: cached tokens still occupy the context window, and a request can include up to 600 images or PDF pages on 1M-context models.
Where to go next
Drill into the pieces: image resizing, tool call overhead, cache hit rates, and the anatomy of a token bill.