Snowflake released new capabilities at the June 2026 Summit to redefine interoperability for the AI era. Their Horizon Catalog uses Apache Polaris to enable bi-directional read and write access to Iceberg data using external engines. This capability allows organizations to work on a single, live, governed copy of data across Snowflake and external lakes. The Snowflake Horizon Catalog connects to external catalogs like Unity Catalog via Catalog Linked Databases and directly calls the catalog’s Iceberg REST Catalog endpoint to get access from the catalog for that specific query. In contrast, Databricks Unity Catalog only supports inbound interoperability. External engines connect to Unity Catalog’s Iceberg REST Catalog endpoints for inbound access, but outbound access requires the Snowflake Catalog SDK which uses JDBC to fetch the metadata.json path and then accesses the underlying storage directly using preconfigured IAM credentials. This outbound access is read only for Iceberg tables managed by external catalogs. Unity Catalog requires users to migrate all tables or limit operations to read only, which delays time to value.
The DBRX model profile
Databricks provides the DBRX 132B Instruct model as part of its Foundation Model API. This model is an open-source mixture-of-experts model. It has a context window of 33k tokens. While Databricks announced it was stopping efforts around the DBRX model because costs ran too high, the model remains available for enterprise use. It is a strong candidate for organizations that need to integrate AI with existing Databricks data engineering workflows. You should evaluate DBRX on task completion, format reliability, and latency tolerance rather than generic leaderboard scores. The model works well for enterprise data assistants, RAG evaluation, and internal automation. A good test set for DBRX measures if the response survives validation and can move to the next step of the workflow.
Comparing DBRX costs with frontier models
The cost of deploying DBRX depends on the request shape. A short classification call has a different cost profile than a long document analysis or an agentic coding session. Users should model DBRX pricing around the amount of input and output tokens rather than just the provider price card. For input-heavy workflows, you must monitor prompt size and retrieved chunks. For output-heavy tasks, you must track output cost separately.
| Model | Input /1M | Output /1M | Context |
|---|---|---|---|
| DBRX 132B Instruct | Free | Free | 33k |
| GPT-5.5 | $6.50 | $39.00 | 1M |
| Claude Sonnet 5 | $2.60 | $13.00 | 1M |
| Kimi K3 | $3.90 | $19.50 | 1M |
| Gemini 3.5 Flash | $0.65 | $3.90 | 1M |
The DBRX 132B Instruct model is free for both input and output tokens. This makes it a distinct option when compared to GPT-5.5, which costs $6.50 per 1M input tokens and $39.00 per 1M output tokens. Claude Sonnet 5 costs $2.60 per 1M input tokens and $13.00 per 1M output tokens. Kimi K3 costs $3.90 per 1M input tokens and $19.50 per 1M output tokens. Gemini 3.5 Flash is the cheapest of these options at $0.65 per 1M input tokens and $3.90 per 1M output tokens.
Managing Foundation Model API limits
Databricks Foundation Model APIs enforce rate limits to ensure fair resource allocation. These limits vary based on the workspace platform tier and the foundation model type. They include input tokens per minute (ITPM), output tokens per minute (OTPM), and queries per hour (QPH). ITPM limits control the input token throughput of an endpoint. OTPM limits control the output token throughput of an endpoint. QPH limits define the maximum number of requests within a 60 minute window. If you reach an ITPM or OTPM limit, the API returns a 429 error. This error includes the limit type, the configured limit, your current usage, and a retry after value. To manage these limits, you should monitor token usage and implement retry logic with exponential backoff. You can also optimize usage by minimizing prompt length and using the max_tokens parameter to control output size.
Deploying models for enterprise data assistants
DBRX is a practical choice for enterprise data assistants and RAG evaluation. For these use cases, you should benchmark the model on real inputs and define an accepted output metric before scaling. The value of DBRX is highest when the team knows what a successful output looks like, such as a valid JSON object or a reviewed code patch. If you plan to deploy these models within an existing Databricks environment, you should prioritize models with low latency for real-time tasks. Using Eden AI allows you to connect DBRX as one route inside a broader model stack. This allows you to test Databricks and add fallback without rebuilding your integration. You can log latency, cost, and the accepted output rate to determine if the model meets your requirements.
Eliminating data movement in Databricks workflows
Databricks released the Excel Add-in on September 10, 2026. This tool brings governed enterprise data directly into Excel. It helps reduce manual exports and stale spreadsheets. Databricks also made OpenSharing capabilities generally available on September 11, 2026. This allows organizations to share foreign Iceberg tables federated from external Iceberg catalogs. You can also share foreign schemas and tables from external sources through Lakehouse Federation without copying those datasets into Databricks. This reduces the need for duplicate datasets, synchronization jobs, and unnecessary data movement. If you already manage your data using Databricks, you will find the DBRX integration exists within your current environment.
Governance challenges in the Iceberg ecosystem
The open data landscape requires portable governance. Snowflake Horizon Catalog provides an end-to-end implementation of the Iceberg REST Catalog protocol. It allows users to define access policies in Horizon Catalog and have them enforced in preferred engines. This capability is in public preview and requires Snowflake Enterprise Edition or higher. It uses the Iceberg Scan Plan API to allow row and column masking to apply when external engines query Iceberg tables. Databricks Unity Catalog does not use the Iceberg REST Catalog for outbound federation. Instead, it relies on the Snowflake Catalog SDK and JDBC. This creates a dependency on long-lived credentials and IAM roles. Will Databricks match the bidirectional capability of Snowflake’s Horizon Catalog for all Iceberg engines?
Strategy for selecting enterprise models
Selecting an AI model requires looking at your specific business needs. Organizations with high-volume workloads might choose Llama 4 or Mistral models to avoid high API costs. Llama 4 offers a context window up to 10M tokens and allows for free weights if you host the model. Organizations needing strong reasoning and governance might choose Claude. For cost-sensitive workloads, DeepSeek provides competitive pricing at $0.28 per 1M input tokens. You should use DBRX when its profile matches a real product constraint like internal automation or RAG evaluation. If you need to maintain data or process sovereignty, a self-hosted model might be the better choice. Use DBRX when it produces more usable outputs than Mixtral 8x7B or Llama 4 Maverick in your specific tests.
