# Token cost control > VI: https://5ac.vn/wiki/ai-token-cost/ · EN: https://5ac.vn/en/wiki/ai-token-cost/ · lang: en · spec_as_of 2026-10-19 (spec no formal spec) · cluster C8 What a model charges per pass: the three token price lines, the audit hole in pay-per-token billing, and the controls that turn an AI bill into a forecast. --- ## Definition Token cost is what a provider charges for one pass of a model: the text you send in and the text you get back are cut into tokens (units smaller than a word) and multiplied by a rate card. Token cost control means turning that number from a surprise into a forecast — knowing what each workflow will cost, getting an alert before the budget breaks, and being able to explain to the owner why this month cost more than the last. 5ac's pieces add a line that pricing pages tend to skim: the cache hit, the machine re-reading exactly what it already read. ## Mechanism Real rate cards carry three lines, all of them published openly. Qwen3.8-Flash on Aliyun lists 0.8 RMB input, 2.7 RMB output and 0.1 RMB cache hit, per million tokens — the cache line costs one eighth of standard input, and it is the line a business can actually move. | Price line | Where it shows up in the bill | What to measure | | --- | --- | --- | | Input | prompts and documents fed in | context length per call | | Output | answers and generated text | retry count, answer length | | Cache hit | re-reading what was already read | hit rate held above 60–70% | Academic work flags a second hole in pay-per-token: it gives providers a financial incentive to misreport how many tokens a model used, while users cannot prove they were overcharged. The martingale-based auditing framework in arXiv:2510.05181 catches the misreporting after fewer than roughly 70 reported outputs, with a false-flag rate below 0.05. On the price curve, a 2020–2026 study documents roughly a 600-fold decline in token prices, economy-tier prices halving every 1.10 years (arXiv:2603.28576) — today's rate card rarely outlives the contract that quotes it. ## Application context 5ac's pricing piece settles on four controls: a monthly ceiling per workflow, telemetry by workflow so you can see which agent, task or team is spending, approval gates before sensitive actions run, and model-agnostic routing that lets you swap models without breaking the pricing promise. The market context is the Copilot billing backlash, which surfaced an x57 multiplier spread and 171 credits for a single complex prompt. The owner's fear is not a weak model; it is a bill nobody can explain. Up the stack, Alibaba CEO Wu Yongming's line at Apsara 2026 — tokens are the electricity of the AI age — explains why rates fall so fast: Ulanqab, which burned 68 TWh on its machine halls in 2023, holds PUE under 1.2 and grew compute power 89.5% in the first half of the year. The practical move this week is not to renegotiate the rate card; it is to measure your own pipeline's cache-hit rate first. ## Sources fetched - [What Alibaba's Machine Intelligence Pitch Means for AI Token Cost](/en/blog/token-cheap-as-electricity-apsara-2026/) — Qwen3.8-Flash rate card, 0.1 RMB cache hit, the 60–70% threshold — accessed 2026-10-19 - [Why predictable AI agent pricing is a competitive advantage for Vietnamese SMBs](/en/blog/predictable-ai-agent-pricing-vietnam-smb/) — cost ceiling, workflow telemetry, the x57 multiplier and 171 credits — accessed 2026-10-19 - [Ulanqab: The AI Megacity Rising From the Steppe](/en/blog/ulanqab-ai-data-center-inner-mongolia/) — 68 TWh, PUE under 1.2, compute power up 89.5% — accessed 2026-10-19 - — Auditing Pay-Per-Token in Large Language Models (v2) — verified via arXiv API 2026-10-19 - — Tiered Super-Moore's Law: price evolution in LLM inference services (v1) — verified via arXiv API 2026-10-19 ## Related - [AI Agent](/en/wiki/ai-agent/) - [Agent orchestration](/en/wiki/agent-orchestration/) - [One-Person Company](/en/wiki/one-person-company/) - [Local LLM Deployment](/en/wiki/local-llm-deployment/) - [Business automation](/en/wiki/business-automation/)