📖 Glossary
AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.
TL;DR: June 29th, DeepSeek dropped a bomb — V4 official launch this July with peak/off-peak pricing, doubling daytime API costs. For a regular user, that's an extra phone's worth of AI bills per year. Off-peak use seems like the fix, but the math tells a different story. And there's one approach that completely sidesteps the pricing game.
Last week, a friend's group chat blew up with a single screenshot: DeepSeek V4 is launching mid-July with peak-hour pricing. Work hours — 9 AM to noon, 2 PM to 6 PM — your API calls cost double.
Translation: using AI during your actual workday just got twice as expensive.
I grabbed a calculator and ran the numbers on a colleague's API bill.
Breaking Down the Price Hike
DeepSeek V4 Pro's base pricing is ¥3 per million input tokens and ¥6 per million output tokens. During peak hours: ¥6 and ¥12 respectively.
For GPT-5.5 users, the baseline is already ¥11 per million output — peak pushes it to ¥22.
Take a typical developer: 50 AI queries per day, each averaging 2000 input + 800 output tokens, across 22 working days.
Baseline: 50 × 22 × (2000×3 + 800×6) / 1,000,000 ≈ ¥11.9/month. Seems trivial.
But those working hours fall entirely within peak time. Peak rate: 50 × 22 × (2000×6 + 800×12) / 1,000,000 ≈ ¥23.8/month.
Double the price. ¥144 extra per year for light use. Heavy users — 200 calls/day, 5000 tokens each — face a peak premium of ¥1,200 per year. That's a new smartphone.
And it's not just about direct API calls. AI coding assistants like Cursor, Codex, and GitHub Copilot fire off API requests constantly. Every Tab completion is a request. A typical day: hundreds to thousands of calls running silently in the background.
"Just Use AI Off-Peak"
Sounds simple. After 8 PM, or on weekends — prices stay the same as before.
Here's the catch. When do you work?
Option A: Do your day job normally, batch all AI questions for the evening. The cost: code can't be debugged in real-time, afternoon bugs wait until night, and when your colleague says "let's ask AI about this" in a 3 PM meeting, your answer is "I'll get back to you tonight."
Option B: Work nights. Congratulations, you're now time-zone aligned with the West Coast.
The art of peak pricing isn't forbidding savings — it's betting you can't actually save. Your workflow is structurally locked into peak hours. You can pay double, or you can break your rhythm. Those are the options.
Local AI: The Obvious Fix Nobody Mentions
The day after running those numbers, I looked at the Kaihe AIBOX sitting on my desk. It's been there for months, quietly running background tasks on an Ethernet connection. Then it hit me: the models running on it aren't metered.
Put simply: whether you call it 50 times today or 500, the marginal cost is zero. Because you're using your own compute, not rented GPUs.
Let's redo the math.
Run an equivalent of DeepSeek V4 Flash locally on Kaihe AIBOX. Heavy user: 200 calls/day, 5000 tokens each. Cloud peak yearly cost: ≈¥1,200. Local: hardware is a one-time investment. Every subsequent call costs nothing at the margin.
Over three years, the cloud minimum is ¥3,600 — and prices only trend one direction. Local is hardware cost plus zero marginal expense forever.
More importantly: when DeepSeek announces tomorrow's pricing change — up or down — it doesn't touch you. Your AI freedom isn't tied to someone else's pricing announcement.
What's Worth More Than Price
There's something beyond the numbers: creative rhythm.
Cloud AI users develop a subtle habit — mentally calculating token counts before every query. Should I segment this long document analysis? Clear chat history to save context? Will one more question break this month's budget?
This "metering mindset" is the natural enemy of deep work. It's like writing with a ticking clock beside you, a voice constantly urging you to hurry up, save tokens, keep it short.
Local AI removes that ceiling. Read a 200,000-word document into context and discuss it for half an hour — it feels as natural as drinking water. Because it produces no new bill.
Here's a number worth watching: in H1 2026, domestic AI API call volume grew 340% year-over-year. Enterprise AI budgets grew only 40%. This scissors gap means one thing — AI costs will keep rising. Peak/off-peak pricing is the opening move. More dimensions of "granular pricing" are coming.
The real value of local AI is stepping outside this game entirely.
Further Reading
Want to dive deeper into local AI? - Cloud vs Local AI: 4 Real-World Tests - Stop Paying Monthly: Bring LLMs Home with a One-Time Purchase
Want to learn more about Kaihe AIBOX? Contact: [email protected]