Model / inference API

Databricks

Databricks sells model serving inside its data platform, so it suits teams that already keep data and pipelines on Databricks and want hosted open models next to them.

Website (opens in a new tab)

Checked 2026-11-05

Databricks sells model serving inside its data platform, so it suits teams that already keep data and pipelines on Databricks and want hosted open models next to them. Its foundation model serving page, as listed on 5 November 2026, bills in Databricks Units, called DBUs. Pay Per Token is DBUs per million tokens, split into input, output and cache read. Provisioned Throughput is DBUs per hour per 50 model units, and Batch Inference is DBUs per hour. The dollar rate per DBU did not load, only an example of $0.07 per DBU on AWS in US East. At that rate Kimi K3 input costs $3.00 per million tokens, and a one-month GLM 5.3 reservation is $10.00 per hour. The catch is the unit: you cannot read a price without the DBU rate for your cloud and region, and regional processing adds 10% on models marked for it. The privacy notice says it does not cover data customers upload, which falls under the customer agreement and a data processing addendum we did not read, so we state nothing on training or retention. Figures come from Databricks' pricing page. We have not run it ourselves. We would use it for governed access to open models on existing data.

Mentioned here

  • Microsoft Azure AI Foundry

    Microsoft Foundry is Azure's platform for deploying models, including OpenAI models sold by Azure, and for building agents on them.

    • Cloud AI platform (hyperscaler)
    • OpenAI compatible

  • AWS Bedrock

    Amazon Bedrock is AWS's managed service for foundation models from several vendors, including Amazon, AI21 Labs, Stability AI, Writer, xAI and others, through one API.

    • Cloud AI platform (hyperscaler)
    • OpenAI compatible

  • Google Cloud Vertex AI

    Vertex AI is Google Cloud's platform for building with Gemini and other hosted models, billed through a Google Cloud account.

    • Cloud AI platform (hyperscaler)
    • OpenAI compatible

  • Kimi K3

    Kimi K3 is an open-weights model from Moonshot AI, released 2026-07-16, with a 1,048,576-token context and text, image and video input.

    • Open weights

    Moonshot AI, 1M context, $0.19 in / 1M

  • GLM 5.3

    GLM 5.3 is an open-weights model from Z.ai, released 2026-08-18, with a 1,048,576-token context and text-only input.

    • Open weights

    Zhipu AI (Z.ai), 1M context, $0.07 in / 1M

Questions

How much does Databricks cost?

Foundation Model Serving is priced in DBUs: the SKU price per DBU multiplied by the model's DBU rate per million tokens (pay per token) or per hour (provisioned throughput). For example, at $0.07 per DBU, Kimi K3 input works out to $3.00 per million tokens. A priority pay-per-token tier costs more (checked 5 November 2026).

Is Databricks OpenAI compatible?

Yes, it offers an OpenAI-compatible API, so most agents and SDKs can point at it by changing the base URL.

Reviews

Loading reviews…

Write a review

Your rating

Up to 1000 characters. Stars alone are enough.

Similar

  • Moonshot AI

    Moonshot AI sells the Kimi API through its developer platform, now at platform.kimi.ai, and it suits teams who want Kimi K3 and K3.1 Pro with a 1,048,576-token context window straight from the maker.

    • Model / inference API
    • OpenAI compatible

  • Featherless

    Featherless is a flat-rate subscription for open models, not a per-token meter, and it suits hobbyists and small teams who want to switch between many open models for one monthly price.

    • Model / inference API
    • OpenAI compatible

  • Z.AI

    Z.ai sells the GLM model family through its open platform API, and it suits developers who want GLM 5.3 straight from the maker, with free small models for testing.

    • Model / inference API
    • OpenAI compatible

  • Minimax

    MiniMax sells its model API through the MiniMax Open Platform, and it suits developers who want M3 or M3.1 Flash directly from the maker at a low token rate.

    • Model / inference API
    • OpenAI compatible

  • AI21

    AI21 sells the Jamba model family through an API, and it suits teams that want a long-context model from a vendor that says it does not train on customer inputs.

    • Model / inference API

  • Nvidia

    NVIDIA's API platform at build.nvidia.com hosts models as NIM microservices on its DGX Cloud for trying out, and it suits developers who want to test Nemotron and other models before deciding where to run them.

    • Model / inference API