Startups & Business

Databricks Agent Bricks pricing risks

Databricks Agent Bricks uses an always-on instance model costing $270 per month, creating scaling risks for low-volume agents. Recent October 2026 price hikes for GTE embedding and Qwen 3 models further complicate budget forecasting for Mosaic AI users.

Databricks Agent Bricks pricing risks

The Scaling Risk

I find Databricks Agent Bricks a poor fit for spiky, low-volume agents because the platform uses an always-on instance model that costs an inferred $270 per month regardless of how many sessions arrive. Databricks publishes no price for the Agent Runtime. If the runtime bills like a default Medium Databricks App, one instance costs 0.5 DBU per hour at $0.75 per DBU. The platform scales reactively to traffic bursts. These additional compute nodes stay active for a minimum duration after the workload drops. I reject the platform for predictable budget planning because the lack of a documented per-session price makes forecasting difficult. For 100,000 customer-service sessions, model tokens cost $6,000, which equals 12 to 19 times the platform bill at 1 vCPU sizing. Users must pay their cloud provider separately for the underlying compute. Managed agent memory is in Beta. This memory store in Lakebase costs $0.069 per capacity-unit hour as the always-on minimum. Autoscaling usage for Lakebase costs $0.092 per CU-hour. This differs from Amazon Bedrock AgentCore, which reclaims idle memory after 120 seconds but maintains a 128 MB floor.

The Infrastructure Lock-in

The Mosaic AI Gateway creates heavy ecosystem dependencies. It relies on Unity Catalog for governance and stores inference logs in Delta Tables. You should prepare for complex cost attribution because DBU consumption aggregates at the workspace level. Isolating Mosaic AI Gateway costs from general data engineering workloads requires custom tagging. While the platform unifies AI and data spend, it locks users into the Databricks control plane. TrueFoundry offers a Kubernetes-native alternative that avoids the management DBU premium. Model training and fine-tuning costs $0.65 per DBU. Foundation Model APIs use a pay-per-token model in DBU per 1M tokens. Model serving includes both pay-per-token and provisioned throughput modes. Provisioned throughput for the newest models requires a minimum concurrency commitment. Vector Search endpoints also carry a charge, with Standard versions costing $0.28 per DBU-hour plus $0.230 per GB-month for storage. For high-concurrency environments, users may need provisioned capacity that prevents full scale-to-zero. Does the convenience of the Databricks control plane justify the management DBU premium?

The October Price Shifts

October 2026 brought price hikes for several provisioned models. Databricks raised the GTE embedding rate from 20 to 100 DBU/hour. The Qwen 3 0.6B Embedding rate rose from 25 to 50 DBU/hour. At the same time, the price for Qwen 3 80B Instruct increased from 78.571 to 104.762 DBU/hour. DeepSeek V4.1 Flash joined the reservation-only group during this period. Foundation Model API rates for Kimi K3 include 75.000 DBU for 1M input tokens and 375.000 DBU for 1M output tokens. GLM-5.2 rates include 35.000 DBU for 1M input tokens and 110.000 DBU for 1M output tokens. Qwen 3.5 122B costs 5.500 DBU for 1M input tokens and 55.000 DBU for 1M output tokens. GLM-5.3 Flash costs 3.750 DBU for 1M input tokens and 12.500 DBU for 1M output tokens. Kimi K3 input at $0.07/DBU equals $3.00 per 1M input tokens.

Service/Model Metric Rate (Oct 2026)
AI/Mosaic AI DBU (US East) $0.07
Kimi K3 1M Input Tokens 75.000 DBU
Kimi K3 1M Output Tokens 375.000 DBU
GLM-5.2 1M Input Tokens 35.000 DBU
GLM-5.2 1M Output Tokens 110.000 DBU
Qwen 3.5 122B 1M Input Tokens 5.500 DBU
Qwen 3.5 122B 1M Output Tokens 55.000 DBU
GTE Embeddings Provisioned 100 DBU/hour