Gemini 3.6 Flash, 3.5 Flash-Lite & Flash Cyber: Google's New Agent-Tier Models and What They Cost

Google just dropped three new models onto OpenRouter — Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini Flash Cyber — each designed from the ground up for agent workloads. With throughput exceeding 150 tokens per second and pricing positioned below the existing Gemini 2.5 Flash tier, Google is making an aggressive play for the multi-turn agent market. Let's break down what "agent-tier" means for your budget.

What Are Agent-Tier Models?

Agent-tier models are optimized for a specific usage pattern: high-throughput, multi-turn loops where an AI agent repeatedly calls tools, evaluates results, and decides next steps. Unlike single-shot chat models, agent workloads generate 10–50x more API calls per task, making per-token cost and latency the two critical variables.

Google's new trio targets exactly this: minimal latency at extreme concurrency, with pricing structured to make long agent loops economically viable. The 150+ tok/s throughput means each turn completes faster, reducing wall-clock time and — crucially — the cost of keeping orchestration infrastructure waiting.

The Three New Models at a Glance

Pricing Context: Where They Fit in Google's Lineup

These models aren't in our pricing database yet, but we can anchor them against Google's existing tiers. Currently, the Gemini family spans a wide cost range:

Google's Flash tier has historically sat at $0.30 input / $2.50 output. For agent-optimized variants, we expect pricing to sit 20–50% below the standard Flash tier — likely in the $0.10–$0.25 input and $0.80–$2.00 output range. Flash-Lite will be the cheapest (it sacrifices some reasoning depth for raw speed), while Flash Cyber may carry a slight premium for its security-hardened inference pipeline.

Why Agent-Tier Pricing Changes the Math

Consider a typical coding agent session. An agent using Gemini 2.5 Pro might execute 40 tool calls across a complex refactoring task, consuming roughly 200K input tokens and 80K output tokens per run. At current Pro pricing, that's:

200K × $1.25/M + 80K × $10/M = $0.25 + $0.80 = $1.05 per task

If an agent-tier model handles the same workload at $0.15 input / $1.50 output:

200K × $0.15/M + 80K × $1.50/M = $0.03 + $0.12 = $0.15 per task

That's a 7x cost reduction — but only if quality holds. The real question is whether these agent-tier models can handle complex tool orchestration without degrading accuracy enough to require retries (which negate the savings).

When to Use Each Model

Gemini 3.6 Flash — Your default agent backbone. Use it for coding agents, research loops, and multi-step workflows where you need both speed and reasonable quality. It replaces 2.5 Flash in most agent architectures.

Gemini 3.5 Flash-Lite — Sub-agent tasks: intent classification, routing decisions, tool-call validation, and summarization within a larger pipeline. Don't use it for primary reasoning — use it for the 80% of agent turns that are mechanical.

Gemini Flash Cyber — Security-specific workloads: SAST/DAST analysis, vulnerability triage, code review for injection patterns, and compliance checking. The hardened pipeline means it's less likely to be jailbroken by adversarial code inputs.

Competitive Implications

Google is clearly responding to Anthropic's Claude agent ecosystem and OpenAI's agent infrastructure push. By releasing purpose-built agent models at sub-Flash pricing, Google is betting that developers will route their agent traffic to the cheapest viable option — and that volume will compensate for thinner margins.

For teams currently running agents on Gemini 2.5 Flash at $0.30/$2.50, the new models represent an immediate cost reduction opportunity. For teams on GPT-5.6 Sol ($5/$30), switching to an agent-tier Gemini model for the mechanical sub-tasks while keeping Sol for complex reasoning could cut total agent costs by 60–80%.

Bottom Line

Google's agent-tier models signal a new pricing paradigm: models designed not for the best single-turn answer, but for the fastest adequate answer in a multi-turn loop. If your agent architecture can tolerate slightly lower per-turn quality in exchange for dramatically lower cost and latency, these models are worth immediate evaluation.

Want to estimate how much these new models could save you? Use our AI cost calculator to compare Google's pricing tiers against your current agent workload and see the difference per task.