Gemini 3.5: What Frontier Intelligence with Action Means

Google introduced Gemini 3.5 at I/O 2026 as a model family that combines frontier intelligence with action. The first release is Gemini 3.5 Flash, which Google calls its “strongest agentic and coding model yet.” The launch comes with strong benchmark claims and broad availability, but the right question for most teams is narrower: does it complete your multi-step tasks reliably, inside the permissions and data boundaries you plan to use?

What changed

Google says the Gemini 3.5 family is “built to help you execute complex, agentic workflows.” According to the announcement, 3.5 Flash is designed for long-horizon tasks such as developing applications, maintaining codebases, and preparing financial documents. Paired with the updated Antigravity harness, Google says it can deploy collaborative subagents and execute multi-step workflows and coding tasks under supervision.

The model also builds on Gemini 3’s multimodal foundation. Google highlights richer, more interactive web interfaces and graphics, along with multimodal understanding of material such as charts and documents.

Google reports these benchmark results for 3.5 Flash:

Benchmark	Google’s reported result
Terminal-Bench 2.1	76.2%
GDPval-AA	1656 Elo
MCP Atlas	83.6%
CharXiv Reasoning (multimodal)	84.2%

Google says 3.5 Flash outperforms Gemini 3.1 Pro on the coding and agentic benchmarks it lists. It also claims output speed four times faster than other frontier models, and says long-horizon agentic tasks can often be completed at less than half the cost of other frontier models. The announcement does not give per-token pricing.

On the larger model, Google said at launch that Gemini 3.5 Pro was already being used internally, with rollout planned for the following month. Check the current status before planning around it.

Where it is available

Google describes 3.5 Flash as generally available across several surfaces:

• Consumers: it became the default model in the Gemini app and in AI Mode in Google Search.
• Developers: Google Antigravity, the Gemini API in Google AI Studio, and Android Studio.
• Enterprises: Gemini Enterprise Agent Platform and Gemini Enterprise.

Google also announced Gemini Spark, a personal agent that runs continuously and takes actions under the user’s direction. At launch it was rolling out to trusted testers, with a beta planned for Google AI Ultra subscribers in the US.

The announcement lists partner use cases, including Shopify using parallel subagents for merchant growth forecasts, Ramp using multimodal OCR for complex invoices, and Salesforce integrating the model into Agentforce. Macquarie Bank is described as piloting customer onboarding over long documents, which is a pilot, not a production deployment.

Why it matters

An agentic model earns its keep when it completes a verifiable sequence of steps, not when it writes a longer answer. Google’s benchmark choices reflect that shift. Most of its headline numbers come from coding and agentic evaluations rather than knowledge quizzes.

Google also controls a large productivity and cloud ecosystem. Model capability becomes more valuable when it can work with authorized files, data, and services. Procurement therefore needs to consider the product surface, not only the base model. The same model behaves differently when it runs in a consumer app, a developer harness, or an enterprise agent platform with administrator controls.

The speed and cost claims matter most for agent workloads. Long-running agents make many calls, so a faster and cheaper model can change what is affordable. But speed only helps if the steps are correct. A fast agent that needs frequent correction can cost more in review time than a slower one that finishes cleanly.

Questions to ask about permissions and data

Before connecting Gemini 3.5 to real systems, get clear answers to these:

• Where are prompts and files processed, and what retention terms apply on each surface?
• How do workspace permissions flow into the model and its subagents?
• What can administrators audit, including tool calls and actions taken?
• Do the same controls apply across consumer, business, and cloud offerings?
• What does “under supervision” mean in practice for multi-step workflows, and where are approval points?

Google says the model was developed under its Frontier Safety Framework, with strengthened cyber and CBRN safeguards and interpretability tools that check the model’s reasoning before it responds. Those are model-level controls. Your own permission design still determines what an agent can reach.

What to test

Build tasks with objective checks, and run them on the surface you will actually use. Our guide on how to compare AI models with a repeatable test explains how to keep runs fair.

• Coding. Require passing tests and a minimal diff. Record how often a human has to intervene.
• Multimodal. Include a chart with a misleading visual scale and check whether the model reads the underlying values.
• Tool use. Give the agent one read-only tool and one write tool that requires approval. Plant an instruction in a document and confirm the agent does not act on it.
• Subagents. Compare a multi-agent run with a single-agent run on the same task. Measure success, time, cost, and duplicated work.
• Research. Require direct links and short quotes that a reviewer can check.

Compare success rate, review time, total cost, and the number of interventions. Do not declare a winner from a benchmark table or one favorite prompt.

What to watch next

Watch for the 3.5 Pro release and its pricing, published per-token costs for Flash on each surface, independent reproductions of the benchmark results, and how Gemini Spark handles permissions once it reaches more users. For a view of how another lab approaches long agent runs, see our analysis of Claude Opus 5 and long-running agents.

The conclusion

Gemini 3.5 Flash deserves a look where Google integrations, multimodal material, or agentic coding are central to your work. Google’s numbers are strong, but they are vendor-reported, and the cost and speed claims are relative to unnamed competitors. The right pilot is narrow, permission-aware, and built around outcomes your team can independently verify.