Every LLM pricing page looks similar, yet the same model can differ by more than 10x across vendors and plans. Let me run the numbers on a real Agent coding bill to show where the gap comes from.

The usage: what this bill looks like

This distribution is typical: 99.3% cache hits. In Agent coding, system prompts and project files are re-read every turn, but with identical prefixes so they hit cache. Cache-hit pricing is 50x cheaper than cache-miss, and it alone decides the bill.

First stop: DeepSeek official pay-as-you-go (baseline)

All figures are ¥ per million tokens. Applying this bill (off-peak):

空闲：55,084,928×0.02/1e6 + 212,119×1/1e6 + 177,492×4/1e6 = ¥2.02
高峰：55,084,928×0.04/1e6 + 212,119×2/1e6 + 177,492×8/1e6 = ¥4.05

官方价区间 = ¥2.02 ~ ¥4.05

Second: Volcengine Agent Plan (the plan costs more)

Volcengine bills in AFP: AFP = (input×0.5 + output×0.5) / 10000. The catch: AFP makes no distinction between cache hits and misses — everything counts at the same coefficient.

AFP = 55,474,539 × 0.5 / 10000 = 2,773.73 AFP

Comparing the tiers

What if you use the full quota? 20,000 AFP = 400M tokens, which at this bill's 99.3% hit rate converts to official pay-as-you-go:

The conclusion is counter-intuitive: at full quota the plan costs 1.37–2.74x official pay-as-you-go. Official cache-hit pricing is dirt cheap at ¥0.02/M, while AFP ignores caching entirely. A plan's value isn't savings — it's model switching plus a fixed budget.

Third: Alibaba Bailian (noticeably more expensive)

¥ per million tokens. Bailian's current version is 2.9–5.7x more expensive than DeepSeek official — they didn't follow the Sept 10 price cut, and their cache price (0.2) is 10x the official 0.02.

When vendors hide their coefficients

Bailian's Token Plan uses Credits, and the docs say: "Credits consumed depend on model type, token usage, thinking mode and tool calls" — no coefficient table provided. I tried back-solving one, realized it was speculation, and deleted it.

Fourth: Zhipu GLM Coding Plan (no DeepSeek support)

Zhipu's plan doesn't support DeepSeek models, so I can only compare with GLM-5.3-Flash. Its points do distinguish cache: points = (miss×2.3 + hit×0.56 + output×8) / 10000

积分 = 212,119×2.3 + 55,084,928×0.56 + 177,492×8
     = 487,874 + 30,847,560 + 1,419,936
     = 32,755,370 → /10000 = 3,275.54 积分

Zhipu's cache coefficient is 0.56 while Volcengine charges a flat 0.5, so the same 55.5M tokens yields 3,275 points vs 2,773 AFP. But these aren't directly comparable — points and AFP have different exchange rates, so convert to money first.

All four routes side by side

Four practical takeaways

1. Look at the cache-hit price first. It can differ 50x (0.02 vs 1+), which matters more than input and output prices combined.
2. Plans aren't always cheaper. Volcengine at full quota is 1.37–2.74x official because it ignores caching. Run your own hit rate through the numbers before buying.
3. Watch price-cut cadence. Bailian missed DeepSeek's Sept 10 cut and ended up 5.7x dearer for the same model. If you're model-locked, going direct to the vendor is safer.
4. Exploit caching deliberately in Agent workloads. Put stable system prompts and project files first and never interleave changing content — once the prefix changes, everything after it misses. This alone can cut 99% off the bill.