LLM API Price Weekly: China-US Model Price Gap at 5.2x, Over 150 AI Funding Deals

AI infrastructure data platform TANDIU has released its first weekly data note, covering the week of October 13–19. The value of this note lies not in the news itself, but in how it demonstrates a data publishing approach "built for AI"—every data point has a JSON endpoint, a source link, a verification date, and can be accessed directly through an MCP server that requires no API key.

For developers and analysts who need to do model selection, cost accounting, or funding tracking, this level of structure is worth paying attention to.

LLM API Prices: A Quiet Week

No vendor adjusted official list prices this week. TANDIU's token market page tracks 69 models from vendors in the US, China, and other regions. The only source of price movement was exchange rates—RMB-denominated Chinese models show small changes in their USD figures due to USD/CNY fluctuations (6.6992 on October 15, 6.6898 on October 19).

China-US Model Price Gap Around 5.2x

Calculated by blended average price per million tokens: US models average ¥56.48, Chinese models ¥10.79—a gap of about 5.2x. This figure provides a rough cost coordinate system, but for any specific model pair, the difference can be much larger.

The Two Ends of the Price Spectrum

The span from cheapest to most expensive is striking:

• Tencent Hunyuan A13B is the cheapest option, at ¥0.50 input and ¥2.00 output per million tokens
• DeepSeek^Ad V4.1 Flash is priced at $0.30 input and $1.20 output
• Google Gemini 4 Argon is $2.00 input and $10.00 output
• Anthropic Claude Opus 5.5 is $4.00 input and $20.00 output
• OpenAI GPT-5.5 Pro / GPT-5.4 Pro sit at the very top, at $30.00 input and $180.00 output

DeepSeek V4.1 Flash drew a lot of discussion on X over the weekend. At list prices, its input token cost is about 1/13 that of Claude Opus 5.5, and its output token cost about 1/17. A price gap of this magnitude is enough to make many teams reconsider their model routing strategy—routing simple tasks to cheap models and reserving complex reasoning for high-end models has become a pragmatic choice.

AI Funding: Over 150 Records in One Week

TANDIU's funding tracker records deals by round, presenting amounts, valuations, and investors exactly as reported by sources, with links to the original articles. For the week of October 13–19, it recorded over 150 deals. A few notable cases:

• Ramp: $1.85 billion Series G at a $60 billion valuation (per Reuters and others)
• ICON: $200 million Series D
• Suanmiao Technology (AI chips): nearly RMB 2 billion combined across Pre-A2 and A rounds within a year (per Caixin)

One detail of the data is commendable: approximations like "nearly RMB 2 billion" remain approximations in the database and are not rounded into precise figures. For those doing funding analysis, this restraint is more valuable than false precision.

Data Interfaces Designed for Agents

TANDIU's core proposition is that AI agents shouldn't have to scrape web pages built for humans. It offers several directly usable entry points:

• https://www.tandiu.cn/data/latest.json: Same-day LLM API prices, with source URL and verification date attached to each row
• https://www.tandiu.cn/data/ai_funding.json: Funding rounds with source links
• https://www.tandiu.cn/data/gpu_prices.json: GPU prices and cloud rental rates
• https://www.tandiu.cn/llms.txt: An index for crawlers and LLMs
• https://www.tandiu.cn/mcp: Remote MCP server, no authentication, no API key required

Quick command-line verification is straightforward:

curl -s https://www.tandiu.cn/data/latest.json | jq '.summary'

One update this week: TANDIU's MCP server has been added to the awesome-remote-mcp-servers list. Keyless MCP servers remain a minority today—most MCP implementations either require local deployment or need authentication configured. This design lowers the barrier for agent integration, but it also means the data provider must bear the load pressure that comes with public access.

MCP (Model Context Protocol) is an open-source protocol proposed by Anthropic, designed to provide a standardized way for AI models to communicate with external tools and data sources. It defines how models discover tools, call tools, and obtain contextual data. Traditionally, connecting an AI application to external data often required writing separate adapter code for each data source; MCP instead uses a unified client-server architecture that lets models dynamically discover and use tools, greatly reducing integration costs. Remote MCP servers allow clients to connect via HTTP and other means without local deployment, and the authentication-free design further simplifies the access process, making it suitable for public data scenarios.

llms.txt is a website index file format optimized for large language models, similar to sitemap.xml for search engines but designed specifically for LLM crawling and parsing. It is typically located at the site root and lists key pages, data endpoints, and resource links in plain text with brief descriptions, helping models quickly understand site structure. For data platforms, providing llms.txt can guide AI agents to efficiently obtain the content they need, avoid crawling irrelevant pages, and reduce server load. This practice is gradually becoming a standard configuration for AI-friendly websites.

Transparency in Data Methodology

The note this week deliberately excludes GPU rental prices. TANDIU explains that the site is switching its weekly change data to a median-based method and will publish once a full week of data is in place. This "rather absent than padded" approach is rare in data products.

Another point worth noting is how exchange rates are handled: USD prices for Chinese models move only with exchange rates, not with vendor price changes. Only by distinguishing these two types of movement can one avoid misreading exchange-rate noise as a change in pricing strategy.

For teams that rely on LLM APIs for cost planning, the real value of this kind of weekly data is not in watching single-week ups and downs, but in establishing a traceable price baseline. When a vendor suddenly changes prices, you have historical data to compare against; when exchange rates fluctuate, you know which "changes" aren't really changes.

The median statistical method refers to sorting a set of data by magnitude and taking the middle value as the representative. Compared with the average, it is insensitive to extreme values (abnormally high or low prices) and better reflects the typical level. In volatile data such as GPU rental prices, a few extremely high or low quotes can skew the average, while the median provides a more robust weekly trend. TANDIU's switch of weekly change data to the median method means its price index will be closer to mainstream market transaction levels and less disturbed by anomalous quotes. However, the median method requires a sufficient volume of data to be stable, so waiting until a full week of data is in place before publishing is a reasonable approach.

Distinguishing exchange-rate noise from pricing strategy is crucial in cross-border model cost comparisons. Because Chinese models are denominated in RMB, converting them to USD causes them to fluctuate with the USD/CNY exchange rate; this fluctuation is not a vendor price adjustment but a change in the conversion result. Misreading exchange-rate fluctuations as pricing changes could lead to incorrect judgments about cost trends. TANDIU explicitly distinguishes the two types of movement: vendor price changes are active行为, while exchange-rate fluctuations are passive results. For cost planners, the focus should be on official list prices in local currency, and when comparing across currencies, use the exchange rate at a single point in time or track local-currency prices directly to avoid exchange-rate noise interfering with decisions.

One-line Summary

TANDIU's weekly report shows a 5.2x price gap between Chinese and US models, over 150 AI funding deals in one week, and a new keyless MCP interface.

Paragraph Summary

TANDIU released its first weekly data note, covering October 13–19. This week, none of the 69 LLM APIs changed prices; the blended average price per million tokens differs by about 5.2x between Chinese and US models, and DeepSeek V4.1 Flash's input cost is only 1/13 that of Claude Opus 5.5. Over 150 AI funding deals were recorded, with Ramp completing a $1.85 billion Series G at a $60 billion valuation. TANDIU emphasizes data interfaces designed for AI agents, offering JSON, llms.txt, and an authentication-free remote MCP server, and makes its data methodology transparent—for example, switching GPU prices to median statistics and delaying publication, and presenting exchange-rate fluctuations separately from vendor price changes.