Google Forces Vertex AI Migration as Imagen 4 Models Face July Sunset

Google has given developers exactly zero breathing room this week, announcing the immediate deprecation of older Imagen 4 models with a hard deadline that's already passed. Meanwhile, the search giant is simultaneously rolling out compelling new features like flex-start VMs and memory-enabled agents, creating a classic carrot-and-stick scenario for Vertex AI users.

What's changing with Vertex AI endpoints?

Google's deprecation of older Imagen 4 models represents one of the most aggressive API sunset timelines we've seen this year. The effective date of 7 July 2025 means any applications still running these deprecated models are now experiencing service interruptions. This isn't a gentle nudge towards migration; it's a forced march.

The migration path leads to the gemini-2.5-flash-image models, which Google positions as the recommended replacement. However, the abrupt timeline raises questions about Google's communication strategy with enterprise customers. Large organisations typically require months of planning for model migrations, particularly when image generation capabilities are embedded in customer-facing applications.

This deprecation sits alongside Google's introduction of flex-start VMs, a new compute option designed for cost-effective inference jobs. These VMs leverage Dynamic Workload Scheduler technology to deliver significant cost savings for intermittent or bursty workloads. The timing is interesting: whilst forcing users off older models, Google is simultaneously offering new cost optimisation tools. It's a clear signal that Google is restructuring its AI infrastructure around efficiency and newer model architectures.

The flex-start VM capability addresses a genuine pain point for many organisations running inference workloads with unpredictable demand patterns. Traditional always-on instances can be expensive for sporadic use cases, whilst cold starts often introduce unacceptable latency. Google's approach appears to balance these concerns, though users will need to evaluate whether their specific workloads align with the flex-start model's constraints.

Vertex AI Agent Engine gains persistent memory

Google's release of Memory Bank functionality for Vertex AI Agent Engine marks a significant step towards more sophisticated conversational AI applications. This preview feature allows agents to dynamically generate and maintain long-term memories from user conversations, addressing one of the key limitations of stateless AI interactions.

The capability enables agents to retain context across extended conversations, moving beyond simple session-based memory to persistent, evolving understanding of user preferences and interaction history. This represents a fundamental shift in how agents can be deployed, particularly for customer service, personal assistance, and complex workflow automation scenarios.

For organisations building agent-based applications, this memory capability opens new possibilities for personalisation and continuity. However, it also introduces new considerations around data governance, privacy, and memory management. Users will need to carefully consider how long-term memory aligns with their data retention policies and user privacy commitments.

Elastic achieves FedRAMP High milestone

Elastic's achievement of FedRAMP High "In Process" status for Elastic Cloud Hosted on AWS GovCloud represents a significant milestone for the observability provider. This designation demonstrates Elastic's commitment to meeting the rigorous security requirements demanded by US federal agencies handling highly sensitive data.

The "In Process" status means Elastic is actively working through the comprehensive FedRAMP assessment process, which evaluates everything from technical controls to operational procedures. For federal agencies, this progress unlocks new opportunities to leverage Elastic's platform for SIEM, logging compliance, and generative AI applications that were previously off-limits due to security requirements.

Elastic's simultaneous recognition as a Leader in Gartner's Magic Quadrant for Observability Platforms validates the company's continued investment in AI-driven features like anomaly detection and root cause analysis. The combination of compliance progress and industry recognition positions Elastic strongly in the competitive observability market, particularly for organisations with stringent security requirements.

Worth watching: new models and expanded capabilities

Mistral AI has expanded its Devstral model family with the release of Devstral Small 1.1 and Devstral Medium. These development-focused models provide new options for coding assistance and software development workflows, though developers should evaluate their specific capabilities against existing alternatives before integration.

Pinecone's expansion of its sparse English embedding model's context window to 2048 tokens represents a meaningful capability enhancement for applications requiring longer-range semantic understanding. The accompanying SDK updates, particularly new admin operations and migration capabilities, should streamline integration for teams building knowledge management and document retrieval systems.

Together AI's launch of Whisper Speech-to-Text APIs delivers impressive performance gains, claiming 15x faster processing compared to OpenAI's implementation whilst adding features like speaker diarisation and large file support. This represents a direct challenge to OpenAI's dominance in speech-to-text capabilities.

Quick hits

• AI21 Labs launched Maestro, an enterprise AI system for knowledge agents with RAG and real-time validation capabilities
• Together AI achieved SOC 2 Type II compliance, strengthening its position for enterprise customers in regulated industries
• Amazon Bedrock added API key authentication support, simplifying integration for new applications
• AI21 Labs integrated Google Drive with File Library, enabling seamless file synchronisation
• Vertex AI Workbench released M131 with updated Dataproc JupyterLab plugin (v0.1.89)

The week ahead: migration deadlines and new previews

The immediate priority for any organisation using older Imagen 4 models is completing migration to supported alternatives. Google's hard deadline has passed, making this a critical operational issue rather than a planning exercise.

For teams evaluating new capabilities, Vertex AI Agent Engine's Memory Bank preview warrants close attention, particularly for organisations building conversational AI applications. The persistent memory functionality could fundamentally change how agents interact with users, but early adopters should carefully consider the data governance implications.

Elastic's FedRAMP progress bears monitoring, especially for federal agencies and contractors evaluating observability platforms. The "In Process" designation suggests full certification may arrive within months, potentially opening new procurement opportunities.

Next week's focus should be on understanding the practical implications of Google's infrastructure changes and evaluating whether the new flex-start VM offering aligns with existing inference workloads. The combination of forced migrations and new cost optimisation tools suggests Google is preparing for a significant shift in how Vertex AI services are delivered and priced.