OpenAI Puts Singapore's Official Statistics Inside ChatGPT
OpenAI has turned Singapore's official statistics into something you can simply ask for. The company surfaced the SingStat plugin this week, so a question inside ChatGPT about population, consumer prices or household income now comes back from the government's own Table Builder rather than from a model's memory, which is a meaningful difference for anyone who needs a figure they can defend.
The pipeline behind the plugin is public, and that is the part worth understanding. SingStat publishes the connection address, requires no authentication, and documents the same setup for other assistants, with Claude used as the worked example. Three tools sit behind it, one to find the right table, one to read its metadata and locate the specific series, and one to pull the figures for a chosen period.
And a user asking for the monthly inflation rate over the past year does not get a guess. The assistant searches the Table Builder for candidate tables, reads the structure of the one it selects, then filters the requested series and time range before presenting the result. The numbers arrive from the published dataset, which is the difference between an answer you can check and an answer you have to trust.
And the scale behind it is larger than the feature suggests. The Table Builder holds about 2,800 tables from 70 public sector agencies, covering close to 180,000 data series across population, income, prices, national accounts, trade, tourism, transport, education, health and housing. OpenAI's own note mentions the popular queries, but the protocol reaches the whole catalogue.
There is a caveat that matters for anyone planning to quote the output. The department states that the presentation, write ups and analysis are entirely the chatbot's, and that neither the government nor any statutory board is liable for inaccuracy, misinterpretation or misrepresentation generated by an AI assistant. The data is official, in other words, while the sentence around it is not.
But the technical limits are worth knowing too. Responses are capped at the first 3,000 rows, and the documentation advises always reading table metadata first so that row numbers and valid date ranges are mapped before any data is requested. Skip that step and a model can pull a series next to the one you wanted and present it with complete confidence.
For newsrooms the practical gain is verification speed. A claim about resident population or household income can be checked against the same table the government publishes, and the resource identifier travels with the query, so the trail from sentence to source is short. That is a different proposition from asking a model to remember a figure it saw during training.
So the plugin is the front door rather than the building. OpenAI moved first with a product ordinary people can use, and that is what turns a government data service into something a journalist or a student will actually reach for. But the endpoint underneath is open to any assistant that speaks the protocol, and other agencies in the region can publish one the same way.
So the commercial side moved too. OpenAI said this week that more than 100,000 advertisers have now run campaigns in ChatGPT, and that ads have expanded to 26 more countries for a total of 89, among them Cambodia, Bangladesh and Pakistan. Ads appear only on the Free and Go plans, are labelled and kept separate from answers, and the company says they do not influence responses.
Google Cloud, meanwhile, will carry OpenAI's frontier models and Codex agentic workflows in its Model Garden, with general availability planned in the coming weeks. Buyers can purchase through the Google Cloud Marketplace and draw down existing cloud spending commitments, and Google says it now offers all three major closed model families. Microsoft remains OpenAI's primary cloud partner.
SingStat runs an open MCP server that any compatible AI assistant can connect to, with no authentication required.
The server bridges to the Table Builder, which holds about 2,800 tables and 180,000 series from 70 agencies.
Responses are capped at 3,000 rows, and metadata should be read before a data request to map the correct series.
The department says analysis and presentation are the chatbot's output and accepts no liability for misrepresentation.
OpenAI's ads now cover 89 countries, and its models arrive in Google Cloud's Model Garden in the coming weeks.
Source: SingStat MCP
