Research#Price comparison#Model inference
CAICT releases token-service performance monitoring report: China consumed ~500T tokens a day in June, normalized prices nearly doubled since H2 2025
CAICT's token-service report (Oct 23): China burned ~500T tokens a day in June; normalized prices nearly doubled to 13.46 yuan, overseas models ~117.8 yuan.
On October 23, the AI institute of China’s state-backed Academy of Information and Communications Technology (CAICT), together with a Ministry of Industry and Information Technology evaluation key lab and the China Artificial Intelligence Industry Alliance (AIIA), released the Token Service Performance Monitoring Report 2026. It distills six dimensions of measurement — call success rate, throughput, first-token latency, stability, price and context window — from monitoring run between January and August 2026. Its most quotable number: in June 2026, China consumed roughly 500 trillion tokens a day, nearly four times the December 2025 level. For teams buying model capacity through an API, it is a rare like-for-like comparison in a market that otherwise runs on vendor benchmarks.
Key points
- Per the report, domestic average throughput reached 78.23 TPS in August, double the 2025 mean of 38.25 — still below the 112.82 TPS that typical overseas models averaged in the same month.
- Normalized average prices nearly doubled from 6.8 yuan in H2 2025 to 13.46 yuan in 2026. Overseas models launched since 2026 average about 117.8 yuan on the same basis — roughly 8.7 times the domestic level, by Global Times’ conversion.
- The same open-weight model can vary by close to 2x in TPS depending on which platform serves it; DeepSeek-V4-Pro’s August TPS was double that of DeepSeek-V3.2.
Background
Chinese LLM APIs spent two years in a price war, and this report documents the turn: prices are climbing again while usage keeps growing. CAICT’s token-service monitoring started in 2025; the 2026 edition is the first to line up throughput, latency, stability and price in one comparable frame — effectively a ruler for the question every buyer asks and no marketing page answers honestly.
The 500 trillion figure is not itself news. TechNode ran “China’s daily AI token usage tops 500 trillion” back on August 28, and Chinese outlets picked up the scale at the time. What the report adds is the comparative layer: performance and price measured the same way across providers, not another aggregate.
Inference performance has also become a selling point in its own right. We covered Inferact running Kimi K3 on 16 TPU v7s, 57% faster than a same-sized GB200 setup — serving vendors compete on speed, and this report hands them an official scoreboard. The open-weight camp keeps building out its own infrastructure too: DeepSeek has open-sourced its Ascend kernel components, and how well a platform optimizes on top of that stack now decides how fast the same weights actually run.
The facts
Per the report’s official summary and reposts by The Paper (via Science Net) and IT Home, the headline numbers look like this. We could not item-check the full PDF, so treat each figure as “per the report”:
| Metric | China | Typical overseas models |
|---|---|---|
| Avg. throughput (Aug) | 78.23 TPS (2025 mean: 38.25) | 112.82 TPS (Dec 2025: 51.35) |
| Avg. first-token latency (Aug) | 1.26 s (2025 mean: 0.78 s) | 2.08 s (April: 1.04 s) |
| Normalized avg. price (2026) | 13.46 yuan (H2 2025: 6.8 yuan) | ~117.8 yuan |
| Context ≥200K share | 95.7% (up 76+ points vs end-2025) | — |
A few readings worth separating:
- Throughput: domestic doubled (38.25 to 78.23), but overseas slightly more (51.35 to 112.82), so the domestic average still trails by roughly a third. Every repost we checked gives the figures without a causal explanation.
- First-token latency is getting worse on both sides, which the report attributes to bigger models, longer contexts and heavier traffic rather than pure regression. Domestic still leads at 1.26 seconds vs 2.08; the overseas figure doubled in just four months, from 1.04 in April.
- Context length: 95.7% of monitored models now take 200K+ tokens, 47.8% take 1M. Long context has become table stakes — Kimi K3.1 Pro, launched the same week, ships 1M as a standard capability.
- Price: 6.8 to 13.46 yuan is close to a doubling; the report’s own phrasing is that prices are “rebounding noticeably, evolving toward capability-driven pricing.” Global Times supplies the comparison: ~117.8 yuan for typical overseas models launched this year.
What others say
Wang Peng, a researcher at the Beijing Academy of Social Sciences, read the report for Global Times (Chi Jingyi, October 24). Buyers facing multiple providers “have no easy way to compare which service is faster, more reliable or better value, so choosing one is difficult,” he said. The industry is moving “from ‘burning cash to win customers’ to ‘earning through service.’” And on the double rise: “customers are not paying more for more expensive tokens, but for models that solve more problems.”
The report’s cross-platform finding is its sharpest angle. DeepSeek, Qwen, GLM and Kimi — four open-weight families — are served on many third-party platforms. After the May wave of new versions, TPS rose and latency compressed across the board; by August, DeepSeek-V4-Pro ran at twice DeepSeek-V3.2’s throughput. Yet the same model can differ by nearly 2x in TPS across platforms, and some third-party platforms beat the model makers’ own first-party endpoints. The report attributes the spread to each platform’s inference acceleration, optimization strategy and cost controls. IT Home, The Paper and Guandian all report the same figures with no contradictions between them.
Our take
For teams doing MaaS selection, three things change. First, the domestic price advantage now has an official number behind it: 13.46 yuan vs 117.8 yuan normalized is an 8x-plus gap, not marketing copy. Second, the throughput gap is real — for latency-insensitive batch workloads, overseas models still belong on the shortlist. Third, the near-2x spread across platforms serving identical weights means “which platform” and “which model” are two separate decisions, and switching the former may buy you more than switching the latter.
The honest catch: the monitoring covers the platforms and models the report chose, not the whole market. Our figures come from the official summary cross-checked against four outlets’ reposts; the full PDF was not item-checked, so verify against the original before a procurement decision. Whether this becomes a recurring index is also open — if CAICT keeps it monthly, it has a shot at becoming the de facto benchmark for Chinese inference serving.