📖 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.
Summary: DeepSeek shipped V4.1-Flash-1017 on October 17; MiniMax shipped M3.1 Flash on October 19. Put the rate cards side by side and the first line is identical: $0.15 in, $0.60 out. Then they diverge. KAIHE AIBOX read both sets of documentation line by line, and the useful question is not which is cheaper. It is when your task runs, how deep it thinks, and whether it recomputes.
The two rate cards, side by side
| Item | DeepSeek V4.1-Flash | MiniMax M3.1 Flash |
|---|---|---|
| Input per 1M tokens | $0.15 off-peak / $0.30 peak | $0.15 flat |
| Output per 1M tokens | $0.60 off-peak / $1.20 peak | $0.60 flat |
| Cached input | $0.003 off-peak / $0.006 peak | $0.015 flat |
| Context window | 1,000,000 | 1,000,000 |
| Maximum output | 384K | Not published |
MiniMax's list price is DeepSeek's off-peak price. Read it the other way and it stings more: at peak, DeepSeek charges twice as much.
Difference one: one doubles, one does not
DeepSeek prices by time of day. Peak runs 01:00–04:00 and 06:00–10:00 UTC on weekdays; everything else is off-peak at half the rate, including every evening and weekend. MiniMax charges one flat price around the clock.
The same million-token input costs $0.30 in the day and $0.15 overnight; on MiniMax it is $0.15 either way. So "which is cheaper" is incomplete. The complete question is: how much of your work can wait until after hours.
Difference two: one can turn thinking off, one cannot
DeepSeek documents both non-thinking and thinking modes, on by default and switchable, so high-volume simple work can run with reasoning fully off. MiniMax cannot: M3.1 Flash rejects the parameter, and the official workaround is to lower effort across five tiers from low to max, defaulting to max. For the same trivial task, one spends zero on reasoning and the other pays at least the low tier.
Difference three: cached input differs 5x
On a cache hit DeepSeek charges $0.003 off-peak against MiniMax's $0.015 — a 5x spread. Caching matters most for agents, which read the same file and the same context over and over; whatever is not cached is repurchased at full price on every lap. In long sessions this line often outweighs input and output combined.
Difference four: the one score that lines up
The two companies publish largely disjoint benchmarks. Only one number can be compared head to head — DeepSWE v1.1, real software engineering:
| Model | DeepSWE v1.1 |
|---|---|
| DeepSeek V4.1-Flash-1017 | 77.0 |
| MiniMax M3.1 Flash | 73.7 |
DeepSeek leads by 3.3 points. The rest is each vendor's own ground: DeepSeek reports HLE 42.6 and Terminal-Bench 42.3; MiniMax reports OSWorld V2 61.7 and CyberGym 82%. None are cross-comparable, and all are vendor-run.
Which fits which
| Your workload | Better fit | Why |
|---|---|---|
| High volume, schedulable overnight | DeepSeek | Half price off-peak |
| Same document read repeatedly | DeepSeek | Cache is 5x cheaper |
| Frequent trivial calls | DeepSeek | Reasoning switchable |
| Results needed during the day | MiniMax | Flat price |
| Throughput and latency | MiniMax | 200+ TPS |
| Open weights right now | Neither yet | Weights follow later |
The real variable is not on the rate card
After reading both sets of documentation, KAIHE AIBOX's read is this: the unit price has been pushed to where there is almost nothing left to win. What remains lives in three actions — when a job runs, how deep it thinks, and whether it recomputes.
A model decides none of those. DeepSeek will not tell you a task belongs at night. MiniMax will not lower its default max effort. Neither remembers a document was already processed.
That is where KAIHE AIBOX sits: a dedicated agent computer, running 24/7 and independent of your laptop. Set tasks by day and they execute overnight at the off-peak rate; effort is matched to task difficulty; results are cached; the model can change and the base stays yours. Buy once, no software subscription.
One line to close
The model layer has converged. The only gap left is when you let that cheap intelligence start working.
Related Reading: - MiniMax M3.1 Flash Is 70% Cheaper and 3x Faster. So Why Isn't Your AI Bill Dropping? - DeepSeek-V4.1-Flash-1017 Ships Wish-Style Coding and Its Own Search API - Gemini 4 Argon Scored 97.5% on ARC-AGI-3. The Same Model on a Standard Harness Scored 49.8%.
For more information, search [KAIHE AIBOX] or contact: [email protected]
KAIHE AIBOX · 7x24 Personal AI Assistant | AI Frontier