Google AI Updates: October 15, 2026

1. Google Research Opens Deep Review, an Agentic Reviewer for Technical Documents

Google. Google Research expanded public access to Deep Review, an agentic system that checks long technical documents and grew out of its Paper Assistant Tool, which gave opt-in pre-submission feedback on more than 16,000 manuscripts at STOC 2026, ICML 2026, NeurIPS 2026, and ICLR 2027. The pipeline splits a document into overlapping sections, assigns each section a compute budget based on information density, verifies sections in parallel with access to the full document, and then runs a synthesis agent that checks cross-section consistency and grounds claims with external tool calls. Deep Review now covers legal, financial, and design documents in addition to CS papers, accepts PDF and plain text, and gives each user two free reviews per week, with higher limits and programmatic access through a Deep Review API using a Gemini API key. Google says uploaded documents are not used for training and are stored for one year for display in the UI. Source

2. Earth AI Geospatial Reasoning Agent Gets Technical Report, Benchmarks, and an Experimental A2A API

Google. Google Research published a technical report on its Geospatial Reasoning agent, which connects Gemini to Earth AI models (WeatherNext, Flood Forecasting, PDFM, AlphaEarth Foundations, Remote Sensing Foundations) plus Google Maps Platform, Earth Engine, Data Commons, and Dataset Search. The report compares a standalone LLM, an LLM with search and maps grounding, an LLM with Earth AI tools, and the full planning agent, finding large gains from the tools on the new GeoExpertQueries and GeoKTools benchmarks (development splits released on GitHub) and on CBGB, while a standalone LLM already scored 93.6% on UnivEARTH. The agent accepts CSV, GeoJSON, and cloud tables, and a Planetary Prediction Engine builds predictive models from user outcome data. Access is experimental and limited: an Ask Google Earth preview, a headless Agent-to-Agent (A2A) Geospatial Reasoning Agent API, and an early preview of the prediction engine. Source

3. Apigee Hybrid Turns Managed APIs Into MCP Tools Inside Customer Clusters

Google. Apigee hybrid 1.17.0 adds optional Model Context Protocol support that converts an OpenAPI 3.0.x spec into an MCP Discovery Proxy, with each operation exposed as a tool reachable through JSON-RPC tools/list and tools/call. The MCP data plane runs in the customer’s own Kubernetes cluster, and a sidecar pulls tool configuration from the Apigee control plane so no inbound connection from Google is required. MCP endpoints inherit Apigee OAuth 2.0, OIDC, API key, and policy enforcement, and API products scope which tools a client sees, with optional per-tool quotas. It is enabled with enableMcpServer: true in overrides.yaml (upgrade apigee-operator before apigee-org), defaults to 2 replicas autoscaling to 10, and forwards W3C trace context from the JSON-RPC _meta field for end-to-end tracing in Cloud Trace. Source

4. BigQuery Adds Daily Token Quotas for Generative AI Functions

Google. BigQuery made generally available four IAM-managed daily token quotas for its generative AI functions: input tokens per project (default 200 billion) and per user (150 billion), and output tokens including thinking tokens per project (20 billion) and per user (15 billion), summed globally across regions. The quotas cover AI.GENERATE, AI.GENERATE_TABLE, AI.GENERATE_TEXT, AI.CLASSIFY, AI.IF, AI.SCORE, and AI.AGG, but not embedding functions, and can be managed through the Cloud Quotas API, gcloud, or Terraform. When a quota runs out mid-query, BigQuery stops issuing LLM calls: functions with max_error_ratio return partial results if within the ratio, and AI.GENERATE leaves skipped rows empty with a quota error in the status field. Google positions the quotas as a preventive complement to Cloud Billing spend caps, which are not enforced instantly, and recommends sizing them with AI.COUNT_TOKENS. Source

5. Research@ 2026 Releases a 2.6M-Event Flood Dataset and Previews Several Research Tools

Google. At its Research@ 2026 event, Google Research announced that its Groundsource dataset of more than 2.6 million flood events is openly available, alongside a model that predicts urban flash-flood likelihood up to 24 hours ahead. It also said TimesFM and TabFM, already available in BigQuery, will be integrated into Gemini for business (no date given), that a low-power edge AI chip co-designed with Synaptics using Google’s open-source IP will be revealed next week, and that MedGemma has passed 10 million downloads. Other items include a multi-agent orchestrator that reproduces prior scientific methods and applies them to new problems, and a Google Labs waitlist for Putty, a shared prototyping tool that generates interfaces from natural-language prompts. Source