Wrong Numbers — defect notes https://wrong-numbers.dev/ LLM cost accounting and eval correctness. Each note is a defect class found and fixed in production libraries. en Why your Claude bill is higher than your dashboard https://wrong-numbers.dev/articles/claude-bill-higher-than-dashboard/ https://wrong-numbers.dev/articles/claude-bill-higher-than-dashboard/ Tue, 01 Sep 2026 12:00:00 GMT Four causes of under-reported LLM spend, ranked by frequency, and the ten-minute reconciliation that identifies which one is yours. Anthropic cache tokens are not in total_tokens https://wrong-numbers.dev/articles/anthropic-cache-tokens-not-counted/ https://wrong-numbers.dev/articles/anthropic-cache-tokens-not-counted/ Tue, 01 Sep 2026 11:00:00 GMT cache_read_input_tokens sits outside input_tokens; libraries that sum input and output under-report every cached request. OpenAI prompt_tokens vs Anthropic input_tokens https://wrong-numbers.dev/articles/openai-prompt-tokens-vs-anthropic-input-tokens/ https://wrong-numbers.dev/articles/openai-prompt-tokens-vs-anthropic-input-tokens/ Tue, 01 Sep 2026 10:00:00 GMT One is inclusive of cached tokens, the other exclusive. Adapters that treat them as interchangeable are wrong in one direction or the other. Streaming message_delta erases accumulated token usage https://wrong-numbers.dev/articles/streaming-message-delta-erases-token-usage/ https://wrong-numbers.dev/articles/streaming-message-delta-erases-token-usage/ Tue, 01 Sep 2026 09:00:00 GMT Assigning the final delta instead of accumulating wipes the input side, including cache reads. Streamed calls under-report; non-streamed are fine.