On October 23, 2026, the China Academy of Information and Communications Technology released its "Large Model Token Service Performance Monitoring Report (2026)", covering January through August. The headline numbers pointed in the same direction: the normalized average price of domestic token services rose from 6.8 yuan in the second half of 2025 to 13.46 yuan in 2026, an increase of 97.9 percent. Average daily token consumption reached 500 trillion in June 2026, nearly four times the December 2025 level.
Multiply the two and the total spend on token services rose roughly eightfold in six months.
The arithmetic is straightforward; the explanation less so. A closer reading of the same report shows that inference unit prices have continued to fall quarter over quarter even as the normalized average price doubled. The two statements are not contradictory. They describe a market where the composition of demand changed faster than the price of any individual unit.
The price war is over. What replaced it is a different kind of market.
What the Normalized Price Measures
The 6.8 yuan and 13.46 yuan figures are not list prices. They are normalized benchmarks — the result of converting different vendors' pricing schemes, model tiers, and billing units onto a single capability-adjusted baseline so that prices can be compared across providers.
That methodology matters for interpretation. A normalized price can rise for reasons that have nothing to do with a vendor raising rates. If the mix of models being used shifts toward higher-capability, higher-priced tiers, the normalized average rises even if every individual model's list price stays flat or falls. If context lengths expand and more tokens are consumed per task at the same per-token rate, the cost per task rises even if the cost per token does not.
The report's own framing supports this reading. It describes 2026 as a year of "price recovery" following two years of price war, during which some leading vendors' quotes fell below 0.01 yuan per thousand tokens in March 2026 — a decline of more than 98 percent from normal levels. The recovery is measured against that trough, not against a stable baseline.
The more informative finding appears in CAICT's companion "2026 AI Industry Ten Keywords" publication: inference unit prices continued to decline every quarter, while total user spending kept rising. The normalized average went up because the demand mix shifted toward tasks that consume more tokens, not because the underlying price of a token went up.
What Changed in the Demand Mix
Three structural shifts explain why consumption grew fourfold while unit prices fell.
Context length. Among monitored models, 47.8 percent now support a context window of 1 million tokens, and 95.7 percent support 200,000 tokens or more. That is a more than 76 percentage-point increase from the end of 2025. Longer context windows are a capability decision that changes the token math of every task. A document analysis task that once consumed 8,000 tokens of context now consumes 80,000. The per-token price can fall and the per-task cost can still rise.
Agentic workloads. A Goldman Sachs analysis cited in industry coverage put the difference in concrete terms: a standard programming question consumes far fewer tokens than the same task executed through an automated agent workflow, where planning, tool calls, observation, and iteration across a multi-step loop multiply consumption by orders of magnitude. Goldman Sachs Research separately projected that global token consumption will grow 24-fold by 2030, reaching roughly 120 quadrillion tokens per month. The volume is not waste. It is the cost of planning, tool use, and iteration across a multi-step loop.
The inference-training split. Token consumption on the inference side now exceeds training by a ratio of more than 6:4, according to industry estimates. Inference runs continuously. Training is episodic. The shift toward inference-heavy workloads means the token bill is no longer a one-time capital expense tied to model development. It is an operating expense tied to every user interaction and every agent task.
The report's latency data corroborates the demand-side story. Average time to first token rose from 0.78 seconds in 2025 to 1.26 seconds in August 2026. CAICT attributed the increase to larger parameter counts, longer contexts, and higher call volumes, not to a degradation in serving efficiency. Throughput actually doubled: average output speed rose from 38.25 tokens per second to 78.23. The system is serving more tokens, faster, to more callers, with longer inputs. The latency increase is the signature of demand outpacing capacity, not of technology getting worse.
The Price War's Aftermath
The two-year price war that preceded this recovery was a customer acquisition strategy, not a sustainable business model.
From 2024 through 2025, major Chinese platforms — Baidu, Alibaba, ByteDance — opened their model APIs at or near zero cost. The logic was ecosystem lock-in: get developers to build on your model, and the switching cost of changing an API endpoint would eventually produce revenue. That logic did not hold. Switching a model endpoint requires changing a few lines of code. When every major vendor priced near zero, there was no lock-in to speak of, only a shared subsidy.
The subsidy had a real cost. Every API call consumes GPU time, power, bandwidth, and depreciation. As call volumes grew exponentially under free pricing, those costs began to erode cash flow. Smaller model startups, without a cloud business to absorb the losses, were the first to exit. Larger platforms began to withdraw free tiers and introduce tiered billing.
The recovery in normalized prices reflects that exit. The vendors that remained stopped competing on the lowest possible list price because the lowest possible list price had already been tested and found unsustainable. What they compete on now is the price of a capable token, sold to customers who need it for production workloads.
The demand-side data explains why they could. A customer running an agent workflow that consumes millions of tokens per task is not choosing a model on the basis of a 20 percent difference in per-token price. The customer is choosing on the basis of whether the task completes, whether the output is reliable, and whether the total cost per completed task is acceptable. The price sensitivity of a chat user and the price sensitivity of an agent operator are different. The market shifted toward the second.
The Margin Problem
The recovery in normalized prices has not solved the industry's profitability problem.
Goldman Sachs Research estimated that China's large model vendors will generate approximately $13 billion in annual recurring revenue in 2026, against training and inference costs of $11 billion, according to the firm's published analysis. API gross margins sit at 20 to 30 percent. Earnings before interest and tax remain negative. Goldman's model does not project a profitability inflection until 2030, when it estimates EBIT margins of 18 percent on $125 billion in revenue.
That forecast describes a business that is growing quickly and losing money at scale. The revenue trajectory — from roughly 35 billion yuan in 2026 to 879 billion yuan in 2030, with daily token consumption rising from 350 trillion to 4,600 trillion — assumes that agent and coding workloads continue to expand and that Chinese models retain their cost advantage. Goldman estimated that Chinese models' combination of sparse mixture-of-experts activation and cache-hit pricing — DeepSeek's cache hits bill at 2 percent of list price — compresses the daily fuel cost of a top American model from $52 to $104 down to about $13, a 4-to-8x gap that widens into a several-hundred-fold difference in billed cost at scale.
The margin problem and the price recovery are not in tension. A vendor can raise its normalized average price and still lose money if its costs rise faster. The price war suppressed revenue. The recovery restores some pricing power. Neither changes the underlying cost structure of serving agentic workloads at 500 trillion tokens per day.
What the Report Does Not Cover
The CAICT figures describe a specific slice of the market, and the boundaries matter.
The report measures token services supplied through public cloud APIs. It does not cover private deployments, enterprise self-built inference clusters, or the actual settlement prices after discounts, prepayments, and committed-spend agreements. The 6.8 yuan and 13.46 yuan figures are comparison benchmarks, not invoices.
That limitation cuts in both directions. It means the headline price increase may overstate what a large enterprise actually pays, because enterprises negotiate volume discounts that are not reflected in the normalized benchmark. It also means the consumption figures may understate total token usage, because a growing share of inference runs on private infrastructure that the report does not monitor.
The report's most useful contribution is not the absolute number. It is the direction of travel across three variables measured on the same basis: price up, volume up faster, unit inference cost down. That combination describes a market where demand has outgrown the pricing strategy that was designed to stimulate it.
What the Recovery Means
The end of the price war does not signal that the market is contracting. It signals that the market has changed shape.
The customers driving volume in 2026 are not consumers trying a chatbot. They are agents running multi-step tasks, coding assistants iterating against test suites, and applications that stuff a million tokens of context into every request. Those customers are less price-sensitive per token and more sensitive to whether the task completes. The vendors that can serve them reliably can charge more per normalized unit than the vendors that were giving tokens away to chat users.
The eightfold increase in total token spend is the clearest measure of that shift. Prices doubled, usage quadrupled, and the bill grew eightfold. A market where that combination holds is a market where demand is being driven by capability, and price is following capability rather than leading it.
The price war is over. The bill is larger than ever. Those two facts describe the same market.
Sources: China Academy of Information and Communications Technology, Large Model Token Service Performance Monitoring Report (2026), released October 23, 2026, via Securities Times (October 24, 2026), DTInsight (October 23, 2026), and Phoenix Tech (October 23, 2026); CAICT "2026 AI Industry Ten Keywords" via Tencent News (October 25, 2026); Business Times China "AI large model price war ends" (October 23, 2026); CSDN "Average price doubles, usage quadruples" (October 23, 2026); Goldman Sachs Research via Yahoo Finance Hong Kong (August 3, 2026); Economic Weekly on China Telecom Research Institute token consumption forecast (September 14, 2026).
Disclaimer
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations
This analysis is based on the CAICT monitoring report as reported by state and trade media; the full report was not reviewed directly.
Normalized average prices are comparison benchmarks, not invoice prices, and cover January through August 2026 only.
Goldman Sachs projections for 2030 are forecasts, not observed data.
Daily token consumption figures aggregate different providers and workloads and are not audited.
Reporting standards differ across jurisdictions; figures may not be directly comparable.
Market size and contract value figures come from third-party research and are estimates.
This sector is changing quickly; claims can become outdated soon.
Sources
- China Academy of Information and Communications Technology, Large Model Token Service Performance Monitoring Report (2026), released October 23, 2026, via Securities Times (October 24, 2026), DTInsight (October 23, 2026), and Phoenix Tech (October 23, 2026)
- CAICT "2026 AI Industry Ten Keywords" via Tencent News (October 25, 2026)
- Business Times China "AI large model price war ends" (October 23, 2026)
- CSDN "Average price doubles, usage quadruples" (October 23, 2026)
- Goldman Sachs Research via Yahoo Finance Hong Kong (August 3, 2026)
- Economic Weekly on China Telecom Research Institute token consumption forecast (September 14, 2026).
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations: This analysis is based on the CAICT monitoring report as reported by state and trade media; the full report was not reviewed directly.; Normalized average prices are comparison benchmarks, not invoice prices, and cover January through August 2026 only.; Goldman Sachs projections for 2030 are forecasts, not observed data.; Daily token consumption figures aggregate different providers and workloads and are not audited.; Reporting standards differ across jurisdictions; figures may not be directly comparable.; Market size and contract value figures come from third-party research and are estimates.; This sector is changing quickly; claims can become outdated soon.