Baseten says enterprise teams can now use open models such as Z.ai's GLM-5.3 Flash and Moonshot AI's Kimi K3 natively in Codex, OpenAI's coding agent, and count the spend against their OpenAI commit. The company announced the partnership on September 29, saying it is one of the first open-model inference providers in the OpenAI B2B Marketplace, with a native integration inside Codex. Baseten's Philip Kiely summarized it in an announcement thread, which has drawn about 884,000 views.
How it works
The OpenAI Marketplace launched at OpenAI DevDay on September 29 as a beta program. Eligible enterprise customers can apply part of their OpenAI commitment to qualifying partner products, and eligibility is set per product. The launch cohort has 32 partners, including Adobe, Figma, Salesforce, ServiceNow, HubSpot, Sierra, Decagon, Harvey, Palo Alto Networks, CrowdStrike, Runway, ElevenLabs and DevRev. Baseten is the provider of open-source models.
For Baseten's customers, the practical change is on the billing side. Baseten says they can use existing OpenAI commitments for open models it serves, either within Codex or through the Responses API. The models named are GLM-5.3 Flash, a 320B-parameter, MIT-licensed model with 18B active parameters, and Kimi K3, Moonshot AI's 2.8-trillion-parameter open model. Baseten hosts Kimi K3 in US data centers, according to The New Stack, which names Baseten, Fireworks and Azure as US hosts for Kimi K3 and GLM-5.2.
Chinese open models inside OpenAI's enterprise channel
Both models come from Chinese labs, and the listing arrives amid a debate over exactly that kind of model. On the same day, Anthropic published an analysis of GLM-5.3 finding that it builds working exploits at close to the rate of Claude Mythos Preview and that its safeguards are easily bypassed. NIST's CAISI called GLM-5.3 on September 17 "the most cyber-capable open-weight model released to date," about four months behind the US frontier. The Marketplace listing names GLM-5.3 Flash, the smaller sibling, and Anthropic's tests also ran Flash: it reached a working exploit chain on a Chrome flaw, and its refusals fell from 95% to 14% after abliteration.
OpenAI's own leadership has flagged the risk. President Greg Brockman said on August 17 that cyber-capable open-weight models would likely "significantly accelerate the threat landscape," according to The New Stack. The policy side is unsettled too: in July, Axios and TechCrunch reported that the administration weighed curbs on Chinese open models, including an Entity List designation and procurement limits, an episode this blog covered in The Week Washington Flirted With Banning Open Weights. No ban has been announced. Meanwhile, Z.ai is courting hosts through its Sovereign Partner Program, and Kimi K3 was the subject of an evaluation-environment episode in August.
The move also fits a demand trend. Vercel's gateway data showed open-weight models moving 62% of its tokens by late August, while Vercel's July index put them at under 4% of spend. A route that lets large buyers use those models through an existing OpenAI contract lowers the procurement friction for the same demand.
Baseten
Baseten has grown quickly. It raised a Series E in January 2026, $300 million at a $5 billion valuation according to Bloomberg, followed by a $1.5 billion Series F in June, reported by BusinessWire and TechCrunch. Bloomberg reported on September 23 that the company was in talks at about a $26 billion valuation.
The OpenAI relationship is not new. Baseten was a day-zero launch partner for OpenAI's gpt-oss models in August 2025, and the Marketplace integration extends that relationship from OpenAI's own open-weight models to open models from other labs.