OpenAI’s GPT Models and Codex Power a Public AI Platform Used by 1,050 Japanese Municipalities

OpenAI’s technology has quietly become part of the daily administrative work of more than a thousand local governments in Japan — not through a direct deal with the national government, but through a small Tokyo startup called Polimill.

According to a customer story OpenAI published on August 31, 2026, Polimill built a platform called QommonsAI using OpenAI’s GPT models and its Codex coding tools, and now counts roughly 1,050 municipalities and about 550,000 public employees among its users. The figures come from OpenAI and Polimill themselves — there’s no independent audit confirming them — but they describe a live, operating system inside Japanese local government, not a pilot or a proof of concept.

The story is notable less for being new — QommonsAI has been running since October 2024 — and more for what it reveals about how OpenAI is expanding into public administration in one of the world’s most bureaucratically cautious markets, through a startup rather than a government contract.

What Happened

Polimill, a Tokyo-based startup, has built its municipal AI platform, QommonsAI, on top of OpenAI’s GPT models and Codex, OpenAI’s coding-agent product. Per OpenAI’s account, the platform now serves about 1,050 municipalities and roughly 550,000 public employees across Japan, spanning tasks like drafting assembly responses, searching legal and welfare regulations, and supporting general public-service work.

These numbers should be read for what they are: claims made jointly by OpenAI and its customer, published in a marketing-style customer story. That doesn’t make them false, but it does mean they haven’t been verified by a government registry, procurement filing, or independent journalist. No named list of participating municipalities has been published.

Who Is Polimill, and What Is QommonsAI

Polimill didn’t start out building government software. The company’s first product, Surfvote, is a citizen-opinion platform where people can weigh in on political and social issues. While working with local governments to build Surfvote, Polimill’s team noticed a structural problem: public-sector staff were so consumed by routine operations that they had little time left to actually incorporate citizen input into policy.

That observation led Polimill to shift focus. Before it could help governments listen to citizens more, it decided, it needed to make government work itself more efficient. The company built QommonsAI, a generative AI platform for public-sector workflows, and released it in October 2024. Today, according to OpenAI, QommonsAI provides specialized AI support across assembly response, public services, social welfare, and legal search.

Masahiro Wakabayashi, Polimill’s Chief AI Officer, described the ambition in blunt terms in OpenAI’s story: with Japan facing a worsening labor shortage, he said, using AI to make government work more efficient “is essential.” His stated goal is for QommonsAI to become “a common foundation that supports every municipality equally” — evolving, in his words, into “the public OS that supports Japan’s government.”

Financial details about Polimill — funding raised, revenue, valuation — are largely undisclosed in public sources.

Timeline

Why It Matters

Japan’s public sector is often characterized by strict regulatory requirements and legacy, fragmented bureaucracy — each municipality tends to have its own workflows and document formats, with historical records scattered across systems that don’t talk to each other. Getting AI tools adopted at scale inside that environment is a genuinely harder problem than selling into a single enterprise customer.

This deployment also fits into a broader pattern of OpenAI investment in Japan. The company opened its first Asian office in Tokyo in April 2024, appointing former AWS Japan president Tadao Nagasaki to lead the market, and released a custom GPT-4 model optimized for Japanese-language tasks. OpenAI has since described local-government AI adoption “from Saitama to Fukuoka” as part of its broader economic vision for the country.

How the Technology Is Used

Two separate OpenAI products sit at the core of QommonsAI: GPT models power the day-to-day interaction layer, and Codex accelerates how Polimill builds the software itself.

On the user-facing side, OpenAI’s account states that GPT models are the most frequently selected option within QommonsAI’s general conversation feature — used for tasks ranging from reading files to responding to routine work requests. Wakabayashi pointed to the broad familiarity of ChatGPT as a practical advantage: public employees rolling out a new tool are more likely to trust something built on technology they’ve already heard of, even if they don’t understand the underlying architecture. QommonsAI also includes administrative controls that let organizations review usage history and restrict which models are available, which OpenAI frames as important for information management and audit readiness in government settings.

It’s worth noting what wasn’t disclosed: neither OpenAI’s customer story nor Polimill’s own materials name a specific GPT model version. There’s no confirmation of whether QommonsAI runs on GPT-4o, GPT-5-series models, or something else. One third-party company-database listing claims the platform uses “GPT-5.4 via Azure OpenAI’s Japan region,” but that listing is AI-generated aggregator content, not a primary source, and it could not be independently corroborated. It shouldn’t be treated as fact.

On the development side, Polimill says it adopted Codex across its engineering workflow — from defining requirements to checking consistency with existing code, implementation, and testing. Engineers now spend more time reviewing AI-generated plans and making high-level decisions, while Codex handles more of the implementation work. Combined with what OpenAI describes as “hands-on support” — sharing engineering practices and helping design where AI should operate versus where humans should review — Polimill says its overall development speed increased by a factor of 3 to 5. That figure, again, is company-reported and hasn’t been independently benchmarked.

Clearing Up the Codex Confusion

If the name “Codex” rings a bell from a few years ago, that’s not a coincidence — but it’s also not the same product.

OpenAI originally launched Codex in 2021 as a code-generating model that powered early versions of GitHub Copilot. The company deprecated that standalone Codex API in March 2023, and its underlying GPT-3-era models were fully retired by January 2024.

In 2025, OpenAI revived the Codex name for something functionally different: an agentic coding tool, not a simple code-completion API. The current Codex reads entire repositories, writes and edits code across multiple files, runs tests, and can operate somewhat autonomously within sandboxed environments, available through a terminal-based CLI, a cloud version, and IDE integrations. It launched as Codex CLI in April 2025, followed by a Codex Cloud research preview the following month, and has been updated with new underlying models since.

Polimill’s use of Codex refers to this current, relaunched product — not the discontinued 2021-era tool.

What’s Verified vs. Unverified

Given how many of this story’s headline numbers come from a single company-produced source, it’s worth laying out clearly what’s confirmed and what isn’t.

How This Fits OpenAI’s Broader Japan Strategy

It’s easy to conflate this story with another, higher-profile OpenAI initiative in Japan: its collaboration with Japan’s Digital Agency, announced in October 2025. That’s a distinct undertaking. The Digital Agency partnership centers on a tool called Gennai, made available directly to government employees, with plans for OpenAI to pursue ISMAP certification — the Japanese government’s security assessment framework for cloud services — before wider rollout across ministries, expected as early as fiscal year 2026.

Polimill and QommonsAI represent a different track entirely: a private startup building a commercial product on top of OpenAI’s API and Codex, then selling it into municipal governments. There’s no evidence that QommonsAI has pursued or obtained ISMAP certification. Polimill’s own site states that it offers a version of its platform for Japan’s LGWAN government intranet, suggesting at least some accommodation of government network requirements — though this detail comes from Polimill directly and hasn’t been independently confirmed. The two initiatives — one government-to-company, one startup-to-government — shouldn’t be treated as the same effort.

Competitive Landscape

QommonsAI doesn’t have a well-documented direct competitor in the specific niche of municipal administrative AI. The most relevant adjacent player is Sakana AI, a Tokyo-based AI research company that has positioned itself around implementing AI for Japan’s enterprise, public sector, finance, and defense needs — including a March 2026 contract with Japan’s Acquisition, Technology & Logistics Agency to develop vision-language models for defense applications. Sakana AI’s public-sector work, however, skews toward national defense and enterprise partnerships rather than local-government back-office administration, so it isn’t a confirmed head-to-head rival to QommonsAI.

Cloud providers like Microsoft Azure and Google Cloud are relevant mainly as potential underlying infrastructure — Microsoft Azure specifically, given the unverified claim about QommonsAI’s hosting — rather than as competing application-layer products.

Limitations and Open Questions

Beyond the unverified figures already noted, several gaps remain. No contract terms or financial arrangement between Polimill and OpenAI have been disclosed. No independent source has evaluated QommonsAI’s accuracy or reliability for legal or administrative search tasks. And while Codex as a product category has had documented security issues — researchers at Check Point disclosed in 2025 that project files could be turned into an execution vector on the Codex CLI, breaking an expected trust boundary — there’s no indication this specifically affected Polimill’s deployment.

Polimill’s own marketing materials also describe newer initiatives, including disaster-response coordination tools between municipalities and a joint proof-of-concept involving humanoid robots, alongside a stated goal of reaching about 1,200 municipalities by the end of 2026. Neither claim appears in OpenAI’s account and neither could be independently verified.

Conclusion

What’s solidly established here is that a Japanese startup has built a generative AI platform for local government using OpenAI’s technology, that it has been operating since October 2024, and that OpenAI is showcasing it as evidence of AI’s utility in public administration. What remains open is how independently verifiable the scale of that adoption really is, which specific models are involved, where the data lives, and whether the platform meets the kind of formal security certification that Japan’s national government is requiring of OpenAI’s other initiatives in the country.

Worth watching: whether Polimill discloses more technical detail as it scales toward its stated 2026 targets, whether independent Japanese reporting or government sources corroborate the municipality and user counts, and how this startup-led track compares in pace and scrutiny to OpenAI’s direct, certification-bound collaboration with Japan’s Digital Agency.

FAQ

What is QommonsAI?
QommonsAI is a generative AI platform built by the Japanese startup Polimill using OpenAI’s technology. It’s designed for public-sector workflows, including assembly responses, public services, social welfare, and legal search.

How many Japanese municipalities use OpenAI-powered tools through Polimill?
OpenAI and Polimill report that about 1,050 municipalities and roughly 550,000 public employees use QommonsAI. This figure is company-reported and hasn’t been independently audited.

Is this the same as OpenAI’s deal with Japan’s Digital Agency?
No. Polimill/QommonsAI is a separate, startup-led commercial platform. OpenAI’s collaboration with Japan’s Digital Agency, announced in October 2025, is a direct government partnership centered on a different tool called Gennai.

Wasn’t OpenAI’s Codex discontinued in 2023?
The original 2021-era Codex API was deprecated in March 2023. OpenAI relaunched Codex in 2025 as a different, agentic coding product. Polimill uses this newer version.

What OpenAI GPT model does QommonsAI use?
This hasn’t been publicly disclosed. Neither OpenAI’s nor Polimill’s official materials name a specific model version.

Is Polimill’s data hosted in Japan?
Not publicly disclosed. No primary source confirms whether QommonsAI’s data is hosted on Japan-based servers or elsewhere.

Does OpenAI have official government security certification in Japan?
OpenAI has said it plans to pursue ISMAP certification as part of its separate collaboration with Japan’s Digital Agency. There’s no public information indicating that Polimill’s QommonsAI platform holds or requires the same certification.

Sources

• Polimill, company and services pages
• OpenAI, strategic collaboration announcement with Japan’s Digital Agency