DeepSeek Flash input: $0.15/1M tokens on a cache miss, $0.003 on a cache hit. 50x. Same agent loop, very different bill depending on whether your prompt prefix stays stable. Do you track cache-hit rate on your agent runs?

Built a guide on why Claude Code burns so many tokens: system prompt, MCP tool schemas, cache writes vs reads, subagents. A 30-turn session priced out per model, with the assumptions labeled. tokentab-ai.vercel.app/guides/why-i... #buildinpublic #aiagents

Hi Bluesky! TokenTab is a free calculator for what AI agents really cost: 28 models, prompt caching, retries, and now subscription vs API break-even for Claude and Codex. tokentab-ai.vercel.app/tools/subscription-vs-api