Data-grounded analysis of AI coding plans, model economics, prompt caching, reasoning budgets and the DeepSeek peak-hour schedule. Every figure is either pulled from the verified dataset or linked to a first-party source.
The Intelligence Index leaderboard, median output speed, first-answer-token latency, published token prices and cost per task for the current frontier — measured rather than described.
Published cache-write and cache-read pricing across Anthropic, DeepSeek, Z.ai and Moonshot, plus a worked 20-turn agent session showing why stable prefixes matter more than sticker prices.
Reasoning effort is now a price dial. Compare measured intelligence, cost per task and latency across frontier, mid-tier and open models, with a three-tier routing policy for real repositories.
Model-level medians and per-provider endpoint benchmarks: DeepSeek V4.1 Flash spans 67–1,504 t/s for identical weights, plus the routing pattern that keeps interactive agents fast without dropping final-patch quality.
Current seat plans from Copilot, Cursor, Devin, Claude, Command Code, Cline and Zed against bring-your-own-key workflows, with the token arithmetic for a 5M-input / 500K-output month.
OpenCode connects 75+ providers for free; Command Code sells hosted open-model credits from $1/month with project-level taste learning. A current tier comparison plus the DeepSeek peak-hour impact.
The full current ladder: ChatGPT Free / Go / Plus / Pro 100 / Pro 200 / Pro 500 / Business against Claude Free / Pro / Max 5x / Max 20x / Team Standard / Team Premium, including Codex and Claude Code access.
GLM Coding Plans from $18/month, Alibaba's Qwen Coding Plan Pro at $50/month, DeepSeek's weekday peak schedule and current Kimi K3 API pricing — with the arithmetic for off-peak batch runs.