A startup founder in Chengdu checks her monthly model bill and realizes it barely moved after she switched models. The reason is not the model price — it is the plan she is on. China's cloud providers are quietly rewriting how model access is packaged, and the newest move comes from ByteDance's Volcano Engine (火山引擎).
The four tiers
Volcano Engine's Agent Plan, its personal subscription package for large-model services, measures usage in a single unit called the Agent Fuel Point (AFP). Instead of billing each call by the token, it sells capacity in four fixed tiers: Small, Medium, Large and Max.
The entry tiers carry the sharpest discount. In a promotion running through November 8, the Small tier lists at ¥9.9 for the first two months against a ¥40 (≈US$5.6 / ≈HK$44) list price, and Medium at ¥49.9 against ¥200 (≈US$28 / ≈HK$220) — both about a 75% cut (a 2.5-fold markdown, or 2.5折) for the first two months, reverting to full price in the third. In rough conversion the promo prices are ≈ US$1.4 (≈HK$11) for Small and ≈ US$7.0 (≈HK$55) for Medium.
Time-of-day and per-model discounts
On top of the subscription, Volcano applies "deduction coefficients" per model, and it discounts those coefficients in rotating promotions. During October 24 to November 9, the newly launched Kimi K3.1 Pro carries a 50% (5折) coefficient discount inside Agent Plan, with usable quota doubled to 200%. DeepSeek-V4.1-Flash gets the same 50% cut from September 15 to December 30, and its economy variant from October 20. Kimi-K2.8-Preview runs a 40% (6折) cut through mid-December.
There is also a standing "Auto" mode, active through November 8, that routes requests to Kimi K3 at a 0.5 coefficient and pushes even more Kimi K3 traffic into the 00:00–08:00 window. The message is clear: the cheapest tokens are the ones you burn while most of China is asleep.
Why ByteDance is doing this
Subscription tiers trade flexibility for predictability. A developer on a flat per-token API pays for exactly what she uses; a subscriber pays a fixed fee and hopes volume justifies it. By discounting the entry tiers so aggressively, Volcano lowers the barrier for individuals and small teams to commit to its model roster and its agent tooling — OpenClaw, Hermes Agent, Claude Code, OpenCode are all listed as compatible.
The per-model coefficient discounts do double duty: they push adoption of the models Volcano most wants to promote (a freshly launched Kimi flagship, a new DeepSeek flash tier) and they soak up spare inference capacity in off-peak hours. For a cloud provider, filling the night shift is pure margin. The deeper prize is lock-in: once a team's agents, prompts and tool integrations live inside Volcano's plan, the cost of leaving is no longer the model price but the workflow around it.
What it means for buyers
The headline "discount" is not a flat API price cut. It is a discount on a coefficient inside a subscription, and the effective per-token cost still depends on each model's coefficient and your actual usage mix. A team that mostly runs flagship models may see little benefit from the Small tier's headline number. The promotion windows are also time-limited, with end dates scattered across early November and December. The procurement lesson is to read the coefficient table, not the banner: the same model can look cheap or expensive depending on which plan you hold and which hour you query it.
How the packaging reshapes the buyer's math
The shift from per-token billing to subscription tiers is less a discount than a change in what a buyer is actually purchasing. A flat per-token API sells a commodity, tokens at a known price, while a subscription sells commitment: a fixed monthly seat in a model roster, a set of agent tools, and a reason to consolidate work in one place. For a founder weighing the Chengdu bill, the real comparison is not the headline promo number against a rival's rate card, but the total of what she would have spent on tokens versus what she now spends on a plan plus the tokens the plan does not cover. The promotion is the hook; the retained workflow is the product.
The longer game is that packaging trains buyers to think in plans rather than prices. Once a team's prompts, agents and integrations are bound to one provider's tiers and tooling, the monthly fee stops feeling like a cost and starts feeling like rent on infrastructure the team already depends on, which is exactly the position a cloud vendor wants to occupy. Reading the coefficient table instead of the banner is the only reliable defense.
Honest limitations
- The 2.5折 figure applies only to the first two months of Small and Medium; the third month reverts to the full ¥40 / ¥200 list. It is an acquisition discount, not a permanent price.
- Per-model coefficient discounts are temporary. The Kimi K3.1 Pro 50% cut ends November 9; others lapse in December. Budget around the promo, not past it.
- Subscription tiers tie you to Volcano's model catalog and tooling. Switching later means re-integrating, not just changing a base URL.
- "Cheaper" is contingent on actually hitting the usage that the tier assumes; light users can pay more per token than on straight API pricing.
- All prices and dates above are drawn from Volcano Engine's own documentation; we did not independently audit the coefficient math.
What readers can do now
- Compare on your real volume. If you run steady agent workloads, stack Volcano's Small (¥9.9) and Medium (¥49.9) first-two-month promo against plain per-token API for your actual call mix before committing.
- Shift batch jobs to the night window. Schedule heavy tasks into the 00:00–08:00 Auto-mode window to ride the 0.5 Kimi K3 coefficient and the off-peak routing.
- Track promo end dates. The Kimi K3.1 Pro 50% coefficient cut ends November 9 — note it before you build a workflow that assumes it.