Databricks Mosaic AI

Databricks Mosaic AI is a collection of generative AI and machine learning capabilities within the Databricks Data Intelligence Platform. It is designed to help organizations build, customize, evaluate, deploy, govern, and monitor AI applications while keeping their models and data connected to the Databricks lakehouse environment. Rather than functioning as a single chatbot or standalone AI model, Mosaic AI brings together model development, model serving, retrieval, AI agents, evaluation, governance, and application development.

Mosaic AI grew out of Databricks’ acquisition of MosaicML in 2023. MosaicML had developed technologies for efficiently training and fine-tuning large language models, and Databricks incorporated those capabilities into its broader data and AI platform. The resulting Mosaic AI portfolio has expanded beyond model training to cover the lifecycle of production AI applications.

One of the platform’s main advantages is that organizations can work with both Databricks-hosted and externally hosted models. Foundation Model APIs provide access to supported models through Databricks-managed endpoints, while Model Serving can also manage external models such as OpenAI and Anthropic models. Databricks provides both pay-per-token and provisioned-throughput options, allowing teams to choose between flexible consumption and more predictable production capacity.

Mosaic AI also provides AI Functions that allow teams to apply generative AI directly to data-processing and analytics workflows. Data teams can use AI functions for tasks such as classification, extraction, summarization, translation, sentiment analysis, embeddings, and other forms of AI-assisted processing. Because these capabilities are integrated into Databricks, they can be incorporated into existing SQL and data pipelines rather than requiring a completely separate AI infrastructure.

For organizations building applications around their proprietary information, Mosaic AI provides retrieval and search capabilities. Databricks Vector Search, now documented as AI Search in current materials, can create vector indexes from enterprise data and retrieve relevant information for retrieval-augmented generation applications. AI Search supports vector indexes and endpoints and provides cost-management capabilities for monitoring its usage.

The platform also supports AI agents and compound AI systems. Databricks introduced its Mosaic AI Agent Framework to help developers build, deploy, and evaluate production RAG applications and agent-based systems. The framework included capabilities for logging agents and chains, evaluating retrieval and response quality, measuring cost and latency, deploying AI applications, and collecting user feedback.

Mosaic AI’s model-serving capabilities are another important component. Model Serving provides a unified interface for deploying and querying AI and machine-learning models for real-time and batch inference. Served models are exposed through REST APIs, allowing developers to connect AI models to websites, applications, business systems, and other services. Databricks can automatically scale serving infrastructure according to demand, while Unity Catalog can be used to govern models and AI assets.

The Foundation Model APIs provide access to a changing selection of models. Current Databricks documentation lists models from providers and families including OpenAI, Anthropic, Meta, Google, Mistral, DeepSeek, and Databricks-hosted models. The exact models available depend on the cloud, region, service, and current model availability. Recent releases, for example, have added Databricks-hosted versions of OpenAI GPT-5.3 Codex, Google Gemini 3.1 Pro Preview, and Anthropic Claude Opus 4.6.

Mosaic AI also gives organizations the option to fine-tune models for specialized workloads. Teams can adapt supported foundation models to their own datasets and then deploy those models through Databricks Model Serving. This can be useful when a general-purpose model does not provide the required accuracy, terminology, or behavior for a particular business domain.

AI Gateway provides another layer of control over model usage. Databricks describes it as a centralized way to govern and monitor model APIs. It can apply permissions and rate limits, track usage and costs, log payloads where configured, and apply AI guardrails. Current documentation states that requests to Databricks-hosted and external foundation models are routed through Unity AI Gateway, allowing organizations to apply controls around model access and spending.

Governance is closely tied to Unity Catalog. Databricks uses Unity Catalog to manage access to data and AI assets, including governed model APIs. Model APIs can be represented as Unity Catalog securable objects, allowing organizations to define and manage access centrally. This is particularly important for enterprises that need consistent permissions and oversight across multiple teams and AI applications.

Mosaic AI is also connected to Databricks’ broader developer ecosystem. Developers can work with Python, SQL, REST APIs, MLflow, the Databricks SDK, and OpenAI-compatible interfaces. Foundation Model APIs can be accessed through the Databricks interface, Python SDK, MLflow Deployments SDK, REST API, or an OpenAI client.

The platform supports a broad range of data sources and third-party technologies. Databricks can connect with cloud storage, databases, business applications, BI platforms, ETL tools, orchestration systems, and AI providers. Its Partner Connect ecosystem provides validated integrations, while Databricks Marketplace provides access to datasets, AI models, applications, notebooks, and MCP servers.

Mosaic AI can therefore be used in many different stages of an AI project. A data team can prepare enterprise information in the lakehouse, retrieve relevant information through AI Search, call a foundation model through Model Serving or Foundation Model APIs, evaluate the resulting application, deploy it through a governed endpoint, and monitor its usage through Databricks’ AI governance and observability capabilities.

The platform is particularly suited to organizations that already have substantial data infrastructure in Databricks or want an integrated environment for data engineering, analytics, machine learning, and generative AI. It can support use cases ranging from document and text processing to RAG applications, AI assistants, predictive machine learning, natural-language data analysis, recommendation systems, and production AI agents.

Mosaic AI is more technically oriented than consumer-focused AI applications. Users working with the platform may need knowledge of data engineering, SQL, Python, machine learning, APIs, cloud infrastructure, or model deployment depending on the project. Databricks does provide increasingly accessible interfaces, including AI/BI Genie and AI-assisted development tools, but the broader Mosaic AI environment is fundamentally designed for professional data and AI teams.

Another important consideration is that Mosaic AI is not one independently priced product with one universal subscription. Its capabilities are part of the wider Databricks platform and can involve different forms of consumption depending on the service. Model APIs can be billed per token, while provisioned model serving, AI Search, compute, storage, and other platform services have their own charges.

For enterprises that want to develop production-grade AI applications around their own data, Mosaic AI offers a broad set of capabilities without requiring every component to be sourced from a separate vendor. Its combination of models, fine-tuning, retrieval, agents, serving, evaluation, governance, and data integration makes it one of Databricks’ central technologies for bringing generative AI into enterprise data workflows.

Pricing

Databricks Mosaic AI uses a consumption-based pricing model, but there is not one universal Mosaic AI price. Costs depend on the specific services and deployment options used.

Foundation Model APIs support pay-per-token pricing for supported models. Databricks also provides priority pay-per-token for latency-sensitive production workloads and provisioned throughput for workloads requiring more predictable performance.

Other Mosaic AI capabilities have their own usage-based charges. For example, AI Search is priced according to its vector-search infrastructure and endpoint usage, while Model Serving and other Databricks compute services have separate consumption charges.

Databricks therefore does not have one simple flat monthly price for all Mosaic AI capabilities.

Pricing Model

Consumption-based

Pricing Note

Mosaic AI is primarily consumption-based. Foundation Model APIs can be billed per token, while provisioned throughput, Model Serving, AI Search, compute, storage, and other services may have separate charges. Exact rates depend on the model, cloud provider, region, service, and configuration.

Review

Databricks Mosaic AI is a powerful platform for organizations that want to build production AI applications around their own enterprise data. Its strongest advantage is the way AI capabilities are integrated with the Databricks data platform rather than operating as an isolated AI service.

The platform offers considerable flexibility because organizations can choose among different model providers, use Databricks-hosted models, connect external models, fine-tune supported models, or deploy their own models. Foundation Model APIs also make it easier to experiment with different models without maintaining the underlying model infrastructure.

Its governance capabilities are another major strength. Unity Catalog and Unity AI Gateway provide mechanisms for controlling access, monitoring model usage, applying rate limits, managing costs, and governing AI assets. These features are particularly useful for enterprises that need stronger controls around production AI.

The main weakness is complexity. Mosaic AI is a broad enterprise platform rather than a simple AI application, and organizations may need experienced data scientists, ML engineers, data engineers, and developers to take full advantage of it.

Pricing can also be difficult to estimate because multiple services use different consumption measurements. Teams need to account for model tokens, serving capacity, vector-search infrastructure, compute, storage, and other applicable charges.

For companies already using Databricks, however, the integration between enterprise data, machine learning, generative AI, governance, and application development is a significant advantage.