China's AI Model API Race: 165.4B Token Gap, 5200% Surge – Who to Bet On?

Weekly Token Rankings Overview

B.AI (@BAI_AGI) released its Chinese model brand token weekly ranking for the period October 16 00:00 – October 22 00:00 SGT. The ranking is based on token usage volume across Chinese large model brands. Previous weeks typically showed DeepSeek leading with others slowly climbing; this week stands out with the top gap narrowing to a single-digit percentage, mid-tier players fighting closely, and one brand surging over 5200% week-over-week — all while three popular Flash models launch simultaneous discounts.

Top-Tier Battle: 165.4 Billion Token Gap

DeepSeek remains first, but Xiaomi has closed to a visibly striking distance. The official ranking chart gives absolute numbers: DeepSeek-V4.1-Flash accumulated 585.44B tokens (rank 1), while Xiaomi's MiMo-V2.6-Flash accumulated 568.90B tokens (rank 2). The difference is 16.54B tokens (165.4 billion) — in weekly ranking volume terms, this is a one-body-length gap that a few large customers' usage swings could erase.

More telling is the trend. Xiaomi's growth is not a one-week spike; it has posted four consecutive weeks of growth (the chart labels it "4-week increase"), and its platform token usage share rose 8.13 percentage points this week alone. An 8-point share gain in a roughly stable total market means real territory captured from competitors — a genuine "land grab." DeepSeek still holds the top seat, but the "4-week growth +8.13pp share" curve is the first time this year a contender has pressured the leader to this degree.

Mid-Tier Melee: 5200% Surge and 0.18% Gap

Mid-tier competition is even more heated, with two distinct storylines.

First storyline: MiniMax's explosive growth. Week-over-week growth exceeded 5200% , the most eye-catching growth star on this week's list, with its representative model MiniMax-M3.1-Flash labeled "2-week increase." However, a cold-water check on absolute volume: MiniMax accumulated only 32.97B tokens, ranking fifth, less than one-quarter of third-place Zhipu (161.83B). Growing 50x from a tiny base still yields modest absolute increments. This is the basic discipline of interpreting week-over-week data: check growth rate, then base, then continuity. Whether MiniMax's growth-star quality holds depends on whether it can sustain double-digit growth next week; a single-week burst does not count.

Second storyline: Qwen vs. third place in a clinch. Alibaba's Qwen3.8-Flash accumulated 159.05B tokens (rank 4), growing 80.86% week-over-week (also labeled "2-week increase"), narrowing the gap with third-place Zhipu (GLM-5.3-FlashX, 161.83B) to just 0.18% .

0.18% is essentially statistical noise magnitude; next week's ranking could flip at any time. This means the top-three seating order under this statistical methodology has entered a "weekly reshuffle" state — great drama for the ranking, and a concrete signal for developers: the leading pack competes so fiercely that no one can afford to relax.

TOP5 Ranking Table

Official ranking chart reorganized into a table (all absolute numbers from B.AI official chart):

Weekly cumulative tokens: 585.44B / 568.90B / 161.83B / 159.05B / 32.97B. Verification: 585.44 − 568.90 = 16.54B, matching the article's "165.4 billion token gap."

The table reveals a order-of-magnitude chasm between the top two (585B / 569B) and third place onward (162B / 159B) — the top battle is a "duel," while third place downward is a separate war. MiniMax's 32.97B confirms the earlier cold-water point: 5200% growth is fierce, but the pot remains small.

Flash Model Limited-Time Discounts

Beyond the ranking, this week brought a direct wallet impact: three popular Flash models discounted simultaneously.

All three deserve attention:

DeepSeek-V4.1-Flash-1017 : October 17 snapshot version. Distinguish from the earlier V4 Flash (late July) — this is the 1017 snapshot, uses MIT license, and ultra-long context is its flagship feature. A straight 40% price cut is the steepest of the three.

Qwen3.8-Flash : Flash tier of the Qwen3.8 family, positioned for low-cost fast response. Timed with Qwen's 80.86% weekly growth momentum, the discount plus volume push forms a coordinated combo.

GLM-5.3-Flash : Flash tier of the GLM-5.3 family. Zhipu is defending third place with GLM-5.3-FlashX (161.83B weekly), so the discount timing signals clear intent.

"Growth momentum + limited-time discount" appearing together typically indicates vendors using price to offset developers' migration costs — lock your workflow in first, discuss price hikes later. Discounts are time-limited ; those wanting to test should seize the window.

Three Decision Recommendations for Your API Strategy

Rankings are others' battles; the wallet is yours. Three pragmatic suggestions:

1. Watch trends, not single weeks. A 5200% single-week surge may be low-base distortion; 80.86% may be a large customer batch release. The truly reference-worthy trendline is Xiaomi's four consecutive weeks of growth plus sustained share climb . Track 3–4 weekly rankings yourself before making migration decisions; don't let one data point drive you.

2. Discount windows are for "testing," not "betting." Three Flash models discounted together create a low-cost horizontal evaluation window — run your real prompts through all three; that beats any ranking. But don't go all-in on production just because of a discount; prices revert when the promotion ends, while migration costs stay yours.

3. Mind the single-source methodology. This weekly ranking comes solely from B.AI. The official chart notes ranking basis (token usage), scope (Chinese model brands), and period, but platform coverage, deduplication methods, and other details are not fully public. Rankings under different methodologies may differ. Treat it as a "trend thermometer" — appropriate; as a "precise ranking" — handle with care.

Conclusion

This week's ranking — the 165.4B top gap, the mid-tier 5200% and 0.18%, plus three Flash models discounting in sync — composes a complete picture: Chinese models' price war and ranking war have become a head-on war of attrition with no safe zones.

For developers, this is actually the best era — top-tier entanglement means lower prices, denser discounts, faster model iterations. Your job is not to bet on one vendor, but to maintain a multi-vendor architecture and use discount windows to trial all alternatives.

Next week's ranking will reveal: can Xiaomi overturn DeepSeek? Will Qwen or Zhipu take third? Is MiniMax a flash in the pan or a true dark horse?

Note: Ranking data sourced from B.AI (@BAI_AGI) Chinese Model Brand Token Weekly Ranking public post and official chart for Oct 16–22 (data period Oct 16 00:00 – Oct 22 00:00 SGT, ranking basis token usage, scope Chinese large model brands). TOP5 weekly cumulative tokens (585.44B / 568.90B / 161.83B / 159.05B / 32.97B) and growth trend labels extracted from official chart; 16.54B = 165.4B verified by conversion. Interpretation of MiniMax low-base effect is this article's speculation, not official methodology.