Google Research said on 14 October 2026 that it is widening early access to the Geospatial Reasoning agent in Google Earth AI, through an experimental deep-analysis update for Ask Google Earth and a headless agent-to-agent (A2A) service. Google’s technical report estimates the agent’s model cost at $0.199 per expert-level query with Gemini 3.8 Flash, against $0.688 with Gemini 3.1 Pro, while Flash scores higher, 62.7 against 47.6.
Google says the agent pairs Gemini with Earth AI’s foundation models, spatial tools and upfront planning. On the report’s GeoExpertQueries benchmark it gives 62.7 for the full agent with Gemini 3.8 Flash, against 3.3 for the same model with basic tools and 0.0 for the model on its own.
Business Pill · THE MODEL AND THE HARNESS
A one-minute explainer of the model and the harness: why products built on the same model can behave very differently. It teaches the general idea only and says nothing about any organisation in this story.
The key insight: As we read it, Google’s report argues that the gain sits in the layer around the model, not in the model: by its numbers, the same Gemini 3.8 Flash goes from 3.3 to 62.7 on expert-level queries once Earth AI’s skills, tools and planning are added.
Google Research shared the work at its Research@ event and listed three updates: a new technical report with four benchmarks, two of them new and released for researchers (GeoExpertQueries and GeoKTools); results from testing with partner organisations; and wider experimental access on Google Earth and Google Cloud.
Google says the agent “connects Gemini to the capabilities of Earth AI’s foundation models”, including WeatherNext, Flood Forecasting and the Population Dynamics Foundation Model. It says users can bring their own data through the agent API as CSV spreadsheets, custom map boundaries (GeoJSON) or cloud tables.
Access stays narrow. Google says: “We are making these available to a limited group of early users”, with an interest form for teams that want to try it. The Planetary Prediction Engine is also in early preview, Google says.
The report’s tests use Gemini 3.1 Pro and Gemini 3.8 Flash. We covered Gemini 4 Argon pricing on 13 October.
GeoExpertQueries scores with Gemini 3.8 Flash as Google’s technical report gives them: 62.7 for the full Geospatial Reasoning agent, 52.2 with its skills and tools but no upfront planning, 3.3 with basic tools and 0.0 for the model alone. Test set, 0 to 100, mean over three runs.
What Was Posted on X
Shared at our flagship Research@ event today: In a new technical report, we demonstrate how combining domain-specific foundation models, spatial tools, and upfront planning enables a Geospatial Reasoning agent to carry out complex spatial analysis and planetary prediction thatโฆ pic.twitter.com/xT3cFuOszU
Google Research’s own post on 14 October, shared from its Research@ event. It says domain-specific foundation models, spatial tools and upfront planning let the Geospatial Reasoning agent do spatial analysis that standalone LLMs can’t, and links the blog post this piece is built on.
What the Technical Report Tests
The report, arXiv 2610.15750, submitted on 11 October, compares four configurations, each with Gemini 3.1 Pro and Gemini 3.8 Flash: a standalone model, the model with basic tools, the model with the agent’s skills and tools but without its upfront planning, and the full agent.
GeoExpertQueries poses open-ended problems drawn from practitioner use cases. The report says each is paired with “a consensus golden workflow from three independent experts” and scored by an autorater. GeoKTools uses templated queries with deterministic reference solvers. The two external benchmarks are CBGB and UnivEARTH, and the report says it finds UnivEARTH saturated.
The Results Google Reports
With Gemini 3.8 Flash, the report gives GeoExpertQueries scores of 0.0 for the standalone model, 3.3 with basic tools, 52.2 with the agent’s skills and tools, and 62.7 for the full agent.
The report says the skills and tools account for most of the gain. It says planning and execution guidance add 10.5 points with Flash, from 52.2 to 62.7, but not with Pro, where the full agent scores 3.2 points lower, 47.6 against 50.8.
On GeoKTools, the report says the strongest baseline reaches 40.5 against the agent’s 85.9. On CBGB, it says the baselines score below 20 and the tool-equipped configurations 85 to 88.
On UnivEARTH, the report says every configuration scores between 86.9 and 98.8, including the standalone model, and it reports that benchmark as a finding about benchmark durability rather than about system capability.
The Cost per Query
The report estimates the model’s token inference cost per query and says the figures exclude external data and compute services such as Earth Engine.
For the full agent, it says Gemini 3.8 Flash is cheaper than Gemini 3.1 Pro on all four benchmarks, by 38 to 71%, at $0.199 against $0.688 on GeoExpertQueries. It says the full agent with Flash costs $0.199 a GeoExpertQueries query against $0.167 without upfront planning, so the 10.5-point gain comes “at about three cents per query”.
The report also says that on CBGB with Gemini 3.1 Pro the full agent costs $0.151 a query against $0.514 for the model with basic tools, while scoring 84.8 against 13.6.
What the Partners Say
Google says the World Health Organization’s Regional Office for Africa used the agent at the start of the 2026 Ebola outbreak in the Democratic Republic of the Congo to map remote mining corridors with high exposure risk and human mobility.
People in Need is quoted by Google saying the agent “helped us analyze 14 historical cyclone paths to identify high-exposure villages in minutes rather than days of manual GIS work”.
Google says GiveDirectly cut crisis scoping from days to minutes, that WPP connected Population Dynamics Insights with its own data for marketing campaigns, and that Milken Institute teams screened their own list of 1,500+ candidate buildings with Google Maps Platform’s Solar API.
The Structural Read
As we read it, the cheaper model winning is the commercial point: the report puts Flash ahead of Pro inside the agent on all four benchmarks at 38 to 71% lower model cost, which shifts the buying decision from model tier to the tools wrapped around it.
As we read it, the data and the foundation models are the part a rival cannot copy quickly: the skills call WeatherNext, flood forecasting and population models that Google says it built.
As we read it, the partner list (WHO AFRO, GiveDirectly, WPP, Milken Institute) spans public health, aid and marketing, which is the range Google is testing before it sets any price.
People in Need, as quoted by Google Research, October 2026
“The experimental Geospatial Reasoning agent helped us analyze 14 historical cyclone paths to identify high-exposure villages in minutes rather than days of manual GIS work”
Three Implications
TEAMS BUYING AI FOR SPATIAL WORK The report says the model with basic tools reaches at most 3.3 on expert-level queries, against 62.7 for the full agent with the same model.
DEVELOPERS ON GOOGLE CLOUD Google says the agent is offered as an experimental headless agent-to-agent (A2A) service, with bring-your-own data as CSV, GeoJSON or cloud tables.
ANYONE PRICING MODEL TIERS The report says that on GeoExpertQueries Gemini 3.8 Flash costs $0.199 a query inside the agent against $0.688 for Gemini 3.1 Pro, while scoring higher.
The framework puts it this way: “Value migrates model → harness → skill marketplace. Pricing power leaves the model and lands on whoever controls the surface and its reusable workflows.”
As we read it, Google’s report is a measurement of that migration in one vertical: the modular Earth AI skills, not the choice of Gemini model, carry most of the score, and the cheaper model does better inside them.
Google’s technical report estimates the full agent’s model token cost at $0.199 per GeoExpertQueries query with Gemini 3.8 Flash, against $0.688 with Gemini 3.1 Pro. The estimates exclude Earth Engine and other external services.
What Is Not Established
The scores and costs come from Google’s own testing of its own system. The report says absolute GeoExpertQueries scores depend on the rating method and that “a stricter autorater would likely assign lower scores across the board”. Google’s blog says it treats the agent’s output as a first draft for experts to review.
The partner results are Google’s account and the quotes it publishes. The cost figures cover model tokens only. Google gave no price for the agent API in the material we read.
Business Engineer Framework
The Agentic Harness War
A Business Engineer framework on why orchestration, memory, tool use and workflow execution are becoming the place where AI value is captured.
Google Research is widening early access to its Earth AI Geospatial Reasoning agent, and its own report estimates $0.199 of model cost per expert-level query with Gemini 3.8 Flash, which scored 62.7 against 47.6 for the costlier Gemini 3.1 Pro.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. On 14 October 2026 we read the Google Research blog post on its Earth AI Geospatial Reasoning agent in full, the abstract and results section of Google’s technical report (arXiv 2610.15750), and the post by Google Research on X that shared them. Every score, cost and partner result here is Google’s own, from its own testing; we ran no tests. The cost figures are the report’s estimates of model token cost only.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur, HackerNoon and Search Engine People. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
