On October 14, 2026, Google Research reported benchmark results and partner testing for Google Earth AI’s experimental Geospatial Reasoning agent (GRA), alongside expanded early access for a limited group. The headline GeoExpertQueries result—62.7 ± 2.5 for full GRA with Gemini 3.8 Flash—measures how closely a system follows expert workflows, not whether its final answers are correct. Google Research announced the update, and the technical report details the evaluation.
What the geospatial agent does
GRA uses Gemini to plan multi-step geospatial analyses and call specialized models, datasets and tools, including Google Earth Engine, Data Commons and Google Maps data. It presents its plan and data sources, and Google says its experimental API can also work with user-provided CSV files, GeoJSON boundaries and cloud tables. Intermediate spatial data passes between tools by reference rather than occupying the model’s context window, according to the technical report.
That makes GRA a geospatial research system, distinct from Google Cloud’s Gemini agent for businesses. Google says expanded early access covers deeper analysis in Ask Google Earth, an experimental agent-to-agent API and a preview of the Planetary Prediction Engine; access is limited to a group of early users.
Google also reported a public-health use case: during an Ebola outbreak in the Democratic Republic of the Congo, WHO AFRO used GRA to map 48 exposed settlements and more than 45,500 people at risk. Google said the analysis took minutes, compared with weeks for a similar process. The company’s account of the partner work also describes People in Need using the agent to analyze 14 historical cyclone paths and identify villages with high exposure.
What GeoExpertQueries measures
The technical report tested four system configurations with Gemini 3.1 Pro and Gemini 3.8 Flash across four benchmarks. In GeoExpertQueries, the full GRA configuration scored 62.7 ± 2.5 with Gemini 3.8 Flash. The same model scored 52.2 ± 5.4 with GRA tools but no upfront planning, and 3.3 ± 0.8 with basic tools. Each figure is the mean and standard deviation across three runs. The report compares outputs with consensus expert workflows; the scores indicate workflow quality, not a general rate of correct answers.
| Gemini 3.8 Flash configuration | GeoExpertQueries score (0–100; mean ± standard deviation across three runs) |
| Full GRA | 62.7 ± 2.5 |
| GRA tools, without upfront planning | 52.2 ± 5.4 |
| Basic tools | 3.3 ± 0.8 |
The gap between basic tools and full GRA is large in this test; the comparison between the two GRA configurations also shows a higher score when upfront planning is included. Those results apply to this benchmark and setup, not to every geographic question or real-world deployment. Another benchmark, UnivEARTH, was already saturated: even the standalone model scored 93.6, so its high scores offer little distinction between configurations.
What the cost estimates cover
For a GeoExpertQueries query, the paper estimates model-inference costs of $0.199 with GRA and Gemini 3.8 Flash, versus $0.688 with GRA and Gemini 3.1 Pro. The Flash estimate is about 71% lower than the Pro estimate, using the Pro figure as the baseline. These are estimates for model inference under benchmark conditions, not customer charges or a complete operating budget: external data and compute services, including Earth Engine, are excluded. The technical report provides the estimates and their scope.
A separate route for developers
On June 18, 2026, Google Earth described open-source Agent Development Kit (ADK) reference agents for sustainable-sourcing due diligence and forest monitoring, built with Google Earth Engine and Gemini. They are separate developer examples, not the core GRA system or a route into its limited early-access group. Google Earth’s announcement describes the examples.
