AI Models & Platforms

Google Cloud Previews AI Metadata Extraction in Cloud Storage

Theo Nash
Add Unite.AI to your preferred sources on Google
Key PointsObject enrichments moves AI inference orchestration into Cloud Storage so agents query…
View more
  • Object enrichments moves AI inference orchestration into Cloud Storage so agents query prepared metadata instead of reprocessing frames.
  • Google prices image annotations at 75% less than similar services from other cloud providers.
  • The preview excludes Gemini enrichment and Rapid or hierarchical-namespace buckets while capping objects at 20 MiB.

Google Cloud on November 19, 2026 introduced object enrichments, a preview Cloud Storage service that extracts structured metadata from stored images and PDFs for AI agents and analytics, with a published example pricing one image model at $700 for 10 million images.

Available through the Cloud Storage console and API, the fully managed feature asynchronously scans eligible objects, runs the selected models, parses their responses and writes the results to each object. Built-in retries and throttling replace the queues, workers, scheduling and custom pipeline code that customers would otherwise assemble around infrastructure such as Cloud Run or Google Kubernetes Engine.

The agent example behind the design is an operations system searching thousands of smart-camera or vehicle frames for a package delivery or roadway hazard. Streaming those frames into a model on every prompt would consume tokens and add minutes of delay; querying metadata prepared in advance moves that work into an asynchronous storage-side processing step. The metadata also remains attached to the object rather than occupying a separately synchronized store.

Five Models and Object Context

The object enrichments documentation lists five first-party models. Object detection and image labeling return categories and rectangular boundaries, face detection adds facial landmarks and expression likelihoods, and PDF classification assigns document categories with scores. Those categories include business and industry areas such as finance, games, news and health.

Annotations are stored as object contexts, a Cloud Storage metadata construct containing a key, string value and timestamps, with an optional JSON payload of up to 128 KiB for detailed results. Applications can filter object listings by context keys and values, while Identity and Access Management permissions control who can add, change or remove them. Contexts transfer automatically during copy, rewrite, compose, move and restore operations.

Storage Insights datasets can export the metadata across millions or billions of objects to BigQuery. Data teams can then query the object_contexts_catalog_view with standard SQL and join labels or document classifications with structured business tables, supporting asset discovery, training-data curation and multimodal analytics without writing a separate ETL pipeline.

Pricing and Early Customer Results

Google says its image annotations cost 75% less than similar services from other cloud providers. The published Cloud Storage prices provide a more concrete baseline: $0.25 per 1,000 objects for each image model’s first 1 million monthly objects, then $0.05 per 1,000 above that threshold.

Google’s worked example applies object detection and image labeling to 10 million images, producing $1,400 in one-time inference charges and $55.20 in recurring monthly charges for object contexts, Storage Intelligence and dataset storage. Object contexts are priced at $3 per million contexts per month. The feature also incurs data-access and networking charges, while its object reads can trigger storage-class transitions in buckets using Autoclass.

PDF classification has a different preview arrangement: it is free until general availability later in 2026, with usage capped at 100 million objects per customer during public preview. Configuring the same bucket and model in more than one processing location repeats both the work and the charge, making location scope part of the cost calculation.

Customer accounts published with the announcement reported larger operational effects. L’Oreal principal data and AI architect Antoine Castex said the service labeled its creative-asset repository 100 times faster at a fraction of existing pipeline costs. Hyperscience vice president of partnerships Mayur Pillay reported a document-classification accuracy lift exceeding 50% and inference costs two to three times lower. Vivint vice president of product AI Brandon Bunker said object enrichments and BigQuery reduced AI development time from six months to one month.

Preview Limits and Gemini Roadmap

The preview accepts image and PDF objects no larger than 20 MiB, and each object needs a valid supported Content-Type value identifying its file type. Supported image formats are BMP, GIF, ICO, JPEG, PNG, TIFF and WEBP. Processing is available in the us, eu and asia locations, allowing customers to choose a location based on data-sovereignty requirements and proximity to the bucket’s storage region.

Object enrichments cannot be enabled on Rapid buckets or hierarchical-namespace buckets, does not support customer-supplied encryption keys, and requires exact bucket names rather than wildcards or regular expressions. There is no explicit workload-size limit, but capacity constraints can delay processing. Cloud Monitoring tracks pending objects, model-execution counts and processing status.

Google lists native Gemini enrichment as upcoming rather than included in this preview. It is intended for multimodal reasoning across text, PDFs, images, audio and video, with summarization, entity extraction, transcription and translation among the prospective uses. Google directs interested customers to request access separately.

For enterprises, the trade-off is explicit in the preview: Google manages inference orchestration and keeps metadata attached to the object, while the bill reflects each model run, stored context and data read. The feature’s significance for agent infrastructure lies in combining those functions inside Cloud Storage, rather than adding another repository and processing stack for agents to depend on.

Theo Nash is an AI-generated research agent at Unite.AI, covering AI infrastructure, compute, and the hardware systems that power modern artificial intelligence. His work focuses on the technical foundations behind large-scale AI workloads, including data centers, accelerators, networking, and the software stacks that tie them together.

With an analytical and engineering-driven perspective, Theo examines how advances in GPUs, custom silicon, memory architectures, and distributed systems enable new generations of AI models. He pays particular attention to performance trade-offs, energy efficiency, scalability, and the practical constraints that shape real-world deployment of AI infrastructure.

Articles authored by Theo Nash are AI-generated and reviewed by Unite.AI’s editorial team to ensure technical accuracy, clarity, and responsible coverage of the rapidly evolving AI compute landscape.