Moonshot AI’s Kimi K3 has gained an enterprise route into OpenAI’s Codex environment through U.S. inference provider Baseten, a development reported by TechNode on September 30. TechNode reported that customers can access the model through Codex while usage counts against existing OpenAI commitments. Baseten’s own announcement says the route also extends through the Responses API. The route combines a Chinese company’s model, a U.S. inference layer, and OpenAI’s developer-facing products in one procurement path.
Baseten’s September 29 partnership announcement says the provider is in an early cohort of OpenAI B2B Marketplace inference providers serving open models. It describes native use within Codex and availability through the Responses API. Baseten, rather than OpenAI, is the company making the integration and hosting assertions in the available materials. That allocation of responsibility is central to understanding the arrangement: Baseten serves model inference, while Codex is the developer interface.
Kimi K3 Joins Codex Through a Separate Inference Layer
The integration puts a clear division of labor around Kimi K3. TechNode says Baseten operates inference and Codex supplies the interface used by developers. That means Codex is the surface on which a customer works, while Baseten is the provider serving the model behind that workflow. Baseten says its connection also works through OpenAI’s Responses API, giving the arrangement an API route alongside the Codex experience.
OpenAI’s Codex launch page provides the relevant product context, although it predates this integration. The company introduced Codex in May 2025 as a cloud-based software engineering agent able to write features, answer questions about codebases, fix bugs, and propose pull requests. The launch page said Codex was available at that time to ChatGPT Pro, Business, and Enterprise users. It does not document the later Baseten partnership or present Kimi K3 as an OpenAI-announced model.
The new channel follows EastFrontier coverage of reported preliminary talks on a possible U.S. cloud path for Kimi K3, for which no agreement had been confirmed. The current development is specifically about access inside Codex and the Responses API through Baseten. The Sept. 30 TechNode account frames the enterprise change around a customer’s existing OpenAI relationship. Baseten’s marketplace position supplies the inference connection that makes this route possible, rather than placing Moonshot AI in the role of the direct enterprise infrastructure provider.
OpenAI Commitments Become the Procurement Mechanism
For enterprise buyers, the key operational detail is billing and procurement. Baseten says customers with OpenAI commitments can apply those existing commitments to open models served by Baseten within Codex or through the Responses API. TechNode likewise reports that Kimi K3 usage can count against existing OpenAI spending commitments, so customers do not need to establish a separate supplier contract for that usage. The fact pattern is about how a model is purchased and reached, not an independent assessment of its capability or adoption.
Shanghai Metals Market reports that Kimi K3 fees can be charged against existing OpenAI procurement commitment quotas, removing a new supplier-procurement step. Its page says its information draws on public information, market communication, and its internal database model. TechNode also relays Sina Finance’s characterization of the arrangement as a debut for a Chinese-origin open model in OpenAI’s enterprise billing channel. That first-status language remains a reported characterization, rather than settled industry history established independently here.
Access should not be treated as universal across OpenAI enterprise customers. Baseten says customers should get on a list to access open models natively in Codex through an OpenAI commitment. The materials provide no figures for customer uptake, model pricing, revenue, or the number of enterprises using Kimi K3. They also do not establish that every enterprise customer has access now. The procurement path is concrete; the commercial scale of its use remains unspecified.
This purchasing arrangement is separate from Moonshot’s effort to connect Kimi to financial data for China’s institutional users. That earlier institutional-use context involved a different Kimi application and should not be read as evidence about Codex access, Baseten hosting, or the terms of this OpenAI commitment channel. The common thread is Kimi’s expansion into enterprise-oriented routes, with distinct products and operating settings.
Baseten’s U.S. Hosting and Kimi K3 Infrastructure Claims
Baseten says the inference for this setup runs on U.S.-based infrastructure and uses zero data retention for all prompts. It also says customers can pin deployments to particular regions, apply fine-grained authentication and authorization, and view usage by models, users, or keys. These are Baseten’s platform and governance claims. The available documentation does not establish that OpenAI independently audits those practices, applies the same terms across its broader services, or guarantees that all enterprise data is U.S.-resident.
Baseten describes its infrastructure footprint as more than 90 clusters across more than 20 clouds, supported by relationships with more than 200 compute providers and active-active deployments. Those scale figures explain why Baseten can position itself as an inference provider for open models in the marketplace, but they remain company claims. They do not identify the number of Kimi K3 deployments, customers, or clusters dedicated to this partnership.
A separate Baseten technical post from July 27 describes Kimi K3 as a 2.8-trillion-parameter open frontier model with vision input and a one-million-token context window. Baseten says Moonshot AI shared the model’s weights for early access. The post also says the weights exceed 1.4 TB in MXFP4 and that fitting them in VRAM requires eight NVIDIA GB300 GPUs. Baseten says it used NVIDIA GB300 NVL72 systems for the deployment. These details describe Baseten’s implementation, rather than OpenAI’s infrastructure or an independent measurement of Kimi K3 performance.
The immediate significance lies in the route’s components: Baseten describes Kimi K3 as an open model, Baseten is the inference and marketplace provider, and Codex plus the Responses API are the OpenAI product channels. The published materials specify enterprise commitments, access-list language, and Baseten’s stated hosting controls. They leave open questions about pricing, uptake, and commercial outcomes, while drawing a more precise boundary around what this integration actually makes available.