API rates per million tokens (input, cached input, output) from the Anthropic and OpenAI pricing pages. Defaults copy the average token mix of one Terminal-Bench 4.0 task run by Codex with GPT-6 Astra at max effort: about 4.6M tokens, 94% cached, 1.6% output. Benchmark runs are long and hard, so swap in figures from your own usage dashboard. Ignores cache-write charges, batch discounts and tax. Flat plans cap usage with limits the vendors describe as multipliers, not token counts.