Reference total cost of ownership
Model usage is identical on every option and is measured separately. Agent Runtime is the current name of Vertex AI Agent Engine.
The workload
Agent sessions a year
That is about 274,000 sessions a day and 5.0 trillion tokens a year.
One session
Model usage
The same models on every option
Akka on your Google Cloud calls the same models as the two Google options, so model usage is the same on all three. It is handled separately from the platform.
Akka Optimize
Akka Optimize searches production traffic for avoidable token spend. Effective optimizations lower token use by up to 70%.
Platform costs
Choose a cost to see how it is worked out
Akka. 205 service cores and 40 platform cores on GKE at $562 a core-year. A session needs 25% of 90 CPU-seconds, sized for a peak of 2.8 times the average, plus 5 for development.
Agent Runtime. 100 million sessions, each billed 90 seconds on 2 vCPU at $0.085 a vCPU-hour and 300 seconds on 4 GiB at $0.009 a GiB-hour.
GKE. 805 cores on GKE at $562 a core-year. A session needs 90 CPU-seconds, sized for a peak of 2.8 times the average, plus 5 for development.
cloud.google.com/products/gemini-enterprise-agent-platform/pricing; cloud.google.com/kubernetes-engine/pricing
Included services
These functions run on the service cores the tier already covers.
Runtime
Google ADK on GKE: 805 cores
Akka: 205 service cores
Each dot is 5 cores, sized for a peak of 2.8 times the average load.
Akka runs each session as a small actor, so a session waiting on a model holds no thread. Manulife reported up to 300% more concurrency after porting agents to Akka. Fox reports 150,000 cores reduced to 22,000.
Akka compute on GKE nodes
Agent compute on GKE nodes
Agent Runtime, at 1.9 times the node price
People
Akka site reliability engineers run the platform around the clock, and Akka support is part of the tier.
Crossover
Akka is below Google ADK on Agent Runtime from 49 million sessions a year, about 134,000 a day.
Akka is below Google ADK on GKE from 141 million sessions a year, about 386,000 a day.
The Akka fee is fixed for the year and steps up with each core pack. Drag the chart to move the volume.
Next step
Platform cost a year at 100 million sessions, before model usage.
Outside the figures
Sources
Assumptions
The agent profile and the workload mix are the Base case of the Akka cost calculator. A session's history is 200 kilobytes. Agents on GKE need the same CPU-seconds a session as agents on Agent Runtime. A core on GKE is one vCPU and 4 GiB at the Autopilot general-purpose Pod price, $562 a year. The Akka Day 2 Ops tier is $175K a year with 25 service cores, one development and one production region, and 40 platform cores on your nodes; service cores beyond those are licensed in packs of 25 and 100. 5 service cores serve development. One semantic governance check runs for each tool call, and its evaluation-model tokens are excluded. Model Armor screens 15% of tokens: the model's output and the user's input. Logs are the session history, and a session writes 1 KiB of metric data. A token is four characters. The support plan is charged on platform charges alone. A full-time engineer costs $300K a year. Prices are list prices, on demand, with one production region.
Sources, read on 21 November 2026
Agent Runtime, Agent Gateway, Sessions and Memory Bank: cloud.google.com/products/gemini-enterprise-agent-platform/pricing. GKE cluster fee and Autopilot general-purpose Pod prices: cloud.google.com/kubernetes-engine/pricing. Cloud SQL SSD storage, high availability: cloud.google.com/sql/pricing. Cloud Logging, Cloud Trace and Cloud Monitoring: cloud.google.com/stackdriver/pricing. Model Armor: cloud.google.com/security-command-center/pricing. Premium Support: cloud.google.com/support/docs/premium. Agent Runtime and the service level agreement: cloud.google.com/terms/gemini-enterprise/sla. Akka customer figures: akka.io/blog/go-slow-to-go-fast and akka.io/customer-stories/tubi.