Baseten and Postman Team Up to Help Enterprises Run Agents in Production

What each company brings

According to Baseten, Postman is used by over 40 million developers and 98% of the Fortune 500. The two companies describe their offerings as complementary: Postman handles building and operating agent workflows, and Baseten handles running the models behind them. Baseten's tagline for the deal is "Build with Postman. Run on Baseten."

Baseten is named as Postman's recommended inference partner for enterprise teams. In that role it runs the models behind agents, offering fast inference, multi-cloud capacity and the post-training infrastructure needed to specialize models.

Postman's Agent Stack

The partnership centers on Postman's Agent Stack, which helps enterprises prepare APIs for agents, control which systems agents can access, and govern their actions. Baseten's post puts the motivation bluntly: "If you do not control your agents, your agents control your company." The stack includes four products:

• Passport: secures API access for humans and agents using scoped credential references, without exposing the underlying API keys.
• Fabric: an AI-native gateway that routes and governs AI, MCP and API traffic.
• Astropods: an operating system with runtime, observability and guardrails for AI agents in production.
• Fern: generates documentation, SDKs and CLIs that make APIs discoverable, typed and callable by developers and AI agents.

Speed, reliability and cost on the Baseten side

Baseten says its Model APIs deliver high throughput and low latency, and that seven of its 2026 launches rank among the top two providers on Artificial Analysis for both throughput and latency. It lists Beam 1.0 (#1 in throughput and latency), Nemotron 3.5 Super VL (#2 in throughput, #1 in latency), Muse Spark Open (#2 in both), DeepSeek V4.1 Flash 1017 (#2 in both), DeepSeek V4.1 Flash (#1 in both), DeepSeek V4 Flash (#1 in both) and GLM-5.2 Fast (#2 in both).

On reliability, Baseten points to Multi-cloud Capacity Management (MCM), which pools GPU capacity across clouds and has built-in resilience to hardware failures. The company says it offers zero data retention (ZDR) by default, and customers can restrict or customize the clouds, regions and clusters they run on, including single-tenant and active-active deployments.

Post-training to cut costs

Baseten argues that for the majority of economically valuable tasks, the gap between a general-purpose model used via prompting and a specialized, post-trained model is large. Baseten Training offers infrastructure and research support to post-train open models for an agent's specific tasks. The company claims this can match or exceed the quality of large frontier models at a fraction of the cost.

As an example, Baseten says Harvey post-trained a 27B open model to match the closed-source frontier on its legal agent benchmark, where top closed models cost roughly $50 per task.

The process Baseten describes has four parts:

• Train for specific tasks using supervised fine-tuning and reinforcement learning.
• Deploy trained models on Baseten and use usage data to guide further training.
• Work directly with Baseten's post-training team.
• Keep ownership of model weights, evaluations and training scripts.

What it means for enterprise teams

For enterprises already building with Postman, the arrangement offers a defined path for hosting the models behind their agents, along with an option to specialize those models for lower cost and better response quality. A statement attributed to Postman in the post says Baseten provides "the infrastructure and post-training expertise to put the right model on every job in Agent Mode." Baseten invites teams preparing to deploy to production to contact its engineers about reliability, cost efficiency and performance across agent workloads.