Google DeepMind Launches Gemini 3.5 Flash AI Model for Agents and Coding Tasks

A new chapter in Google’s AI ambitions opened this week as Google DeepMind announced Gemini 3.5 — a fresh family of multimodal AI models built around what the company calls “frontier intelligence with action.” The first model out of the gate is Gemini 3.5 Flash, and it’s available now.

This isn’t a quiet internal update. Gemini 3.5 Flash is already live across a mix of Google products and developer tools, including the Gemini app, AI Mode in Google Search, Google Antigravity, the Gemini API in Google AI Studio, Android Studio, the Gemini Enterprise Agent Platform, and Gemini Enterprise. That’s a broad rollout by any measure.

What Makes Gemini 3.5 Flash Different

The Flash line within the Gemini family has always been Google’s faster, lower-latency option — built for everyday use rather than the heavy reasoning workloads that Pro and other variants are designed to handle. Gemini 3.5 Flash continues that tradition, but Google says this version pushes the boundaries of what a Flash model can do.

For their part, the focus is on agentic execution. That means the model is designed to carry out multi-step tasks on your behalf, using tools like function calling, code execution, search grounding, and structured output. It can preserve reasoning context across long, multi-turn conversations — which matters a great deal when you’re asking an AI to work through a complex problem over many exchanges rather than a single prompt.

The technical specs back that up. Google’s enterprise documentation lists a maximum input limit of 1,048,576 tokens — that’s roughly the equivalent of several full-length novels fed into a single context window. The maximum output limit sits at 65,535 tokens. For developers building applications that need to process large documents, codebases, or lengthy conversation histories, those numbers are meaningful.

On benchmarks — and it’s worth being clear that these are Google’s own reported figures, not independently verified results — Gemini 3.5 Flash scores 76.2% on Terminal-Bench 2.1, 83.6% on MCP Atlas, and 84.2% on CharXiv Reasoning. It also posts an Elo score of 1,656 on GDPval-AA. Google also claims the model runs at around four times the output speed of other frontier models in tokens per second, though that figure hasn’t been independently confirmed.

Built for the Age of AI Agents

The word “agentic” keeps coming up in Google’s language around this launch, and it’s not accidental. The broader AI industry is moving rapidly toward models that don’t just answer questions but actively complete tasks — booking things, writing and running code, searching the web, filling in forms, calling APIs. Google wants Gemini 3.5 Flash to be the engine powering that kind of work.

So the model supports what Google describes as “long-horizon tasks” — jobs that unfold over many steps and require the AI to keep track of what it’s done, what it still needs to do, and what tools it has available. That’s a different kind of capability from generating a paragraph of text or summarising an email.

The release puts Google squarely in competition with other frontier AI models targeting coding and automation, including Anthropic’s Claude family and OpenAI’s GPT-4o and o-series models. The AI model market is moving fast, and each company is pushing hard on the agentic angle right now.

Questions Around Data and Reliability

Faster and more capable models bring real productivity gains. But they also raise questions that don’t disappear just because the technology is impressive.

Hallucinations — where AI models confidently produce incorrect information — remain a known issue across all large language models, including Gemini. Any organisation relying on Gemini 3.5 Flash for important decisions will still need human review built into the process. That’s not a criticism unique to this model; it applies across the board.

Data handling is another consideration. Businesses using the Gemini API or enterprise products need to understand what data is sent to Google’s servers, how it’s stored, and whether that’s compatible with their obligations under UK GDPR and guidance from the Information Commissioner’s Office on AI and automated decision-making. Google does offer enterprise-grade data controls, but organisations need to check the specifics of their own plans and agreements.

Liz Reid, Google’s Vice President and Head of Search, said of the broader Gemini push: “We’re moving from a model that answers questions to one that gets things done.” That framing captures what Google is aiming for — though how reliably it delivers in practice will depend on the use case and the safeguards organisations put in place around it.

What Comes Next

Google has framed Gemini 3.5 as a family, not a single model. Gemini 3.5 Flash is the first release, which suggests more variants — likely including a Pro-tier model — are in the pipeline. The company hasn’t confirmed a timeline for those, but the naming convention makes additional releases probable.

For developers, the model is accessible now through Google AI Studio and the Gemini API. Consumer-facing features powered by Gemini 3.5 Flash are already rolling out through Google Search and the Gemini app, though availability of specific features may vary by region.

What This Means for Kent Residents

Kent residents using Google Search or the Gemini app may already be interacting with Gemini 3.5 Flash, depending on which features have been enabled in the UK — it’s worth checking your app settings if you want to know what’s running under the bonnet. Local businesses, developers, and public-sector suppliers considering the model for coding, automation, or customer-service workflows should review their Google plan terms and confirm UK availability before building anything around it. Organisations such as councils, NHS bodies, and universities in Kent will need to assess their data protection obligations under UK GDPR and ICO guidance before deploying AI tools in public-facing services. If you’re an individual user, the practical advice is straightforward: don’t input personal or sensitive information without understanding how it’s handled.