7/22/2026

Gemini 3.6 Flash and 3.5 Flash-Lite: two models for an agent router, not one winner

Realistic AI engineer routing business documents between fast and capable production models

Google released two generally available production models on July 21: Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. It also announced Gemini 3.5 Flash Cyber, but that is not a public Gemini API model. It is planned only for a limited CodeMender pilot with governments and trusted partners.

The first two matter to enterprises. They are not competitors where one must win. They are separate layers in an agent router: 3.6 Flash for harder reasoning and multi-tool work, Flash-Lite for low-cost, fast, repeatable work at volume.

A practical map

Model API price per 1M tokens Best role Typical workflow
Gemini 3.6 Flash $1.50 input / $7.50 output main worker agent email + attachment + RAG + internal API, multimodal analysis, harder coding task
Gemini 3.5 Flash-Lite $0.30 input / $2.50 output low-cost worker or subagent ticket classification, field extraction, JSON normalisation, agentic search, bulk document processing
Gemini 3.5 Flash Cyber not published for ordinary API access security pilot in CodeMender do not plan it as an available component today

Both public models have 1M context, 64k maximum output, thinking, and built-in tools including Computer Use. That does not mean every ticket should receive a million tokens. Context remains cost, latency, and error surface; the router should send only what the case needs.

Gemini 3.6 Flash: when the agent must connect multiple steps

Google positions 3.6 Flash as the workhorse for coding, knowledge work, and multimodal tasks. It reports 17% fewer output tokens than 3.5 Flash and a lower output price: $7.50 rather than the earlier $9 per million tokens.

The reported benchmarks are mainly comparisons with its predecessor:

  • DeepSWE: 49% vs. 37% — agentic software tasks. Signal: a better candidate for fixes, migrations, and coding agents than the older Flash, not proof it can merge production PRs without review.
  • MLE-bench: 63.9% vs. 49.7% — machine-learning research tasks. Signal: stronger multi-step and tool-assisted work, not a replacement for research sign-off.
  • OSWorld-Verified: 83.0% vs. 78.4% — verified computer-use tasks. Signal: better for UI automation, while actions involving money, permissions, or publishing still need an explicit control point.
  • GDPval-AA v2: 1421 vs. 1349 — knowledge-work tasks. Signal: a stronger candidate for document analysis, draft reports, and work across multiple sources.

Use it where an agent reads an email and attachment, retrieves company procedure, calls a CRM or ERP in read-only mode, builds a structured proposal, and only then lets validation decide whether to proceed. Let 3.6 Flash choose a workflow route, not make an irreversible write without rules.

Gemini 3.5 Flash-Lite: when throughput matters more than peak capability

Flash-Lite is materially cheaper. Google calls it the fastest 3.5 model and cites 350 output tokens per second from Artificial Analysis. Speed only matters when a narrow operation repeats thousands of times.

Google reports these comparisons against older Flash-Lite / Flash models:

  • Terminal-Bench 2.1: 54% vs. 31% over 3.1 Flash-Lite. Signal: a stronger subagent for bounded terminal and coding tasks, not an architect replacement.
  • GDM-MRCR v2: 72.2% vs. 60.1% — long-context work. Signal: useful for search and sorting across large document sets when retrieval still reduces the prompt to relevant chunks.
  • GDPval-AA v2: 1140 vs. 642 over 3.1 Flash-Lite. Signal: the lower-cost model can now handle more knowledge work, while 3.6 Flash remains the stronger escalation layer.
  • SWE-Bench Pro: 54.2% vs. 49.6% over Gemini 3 Flash and OSWorld-Verified: 74.0% vs. 65.1%. Signal: Lite is more than a classifier, but its economics shine in repeated bounded work.

Use Flash-Lite for first-pass support routing, invoice field extraction, contact deduplication, fixed JSON schemas, content pipelines, candidate-document search, or autonomous subagents with a limited tool-call budget.

Use a router, not a bulk switch

In n8n or Make: first apply a cheap deterministic filter for input type, PII, and mandatory fields. Let Flash-Lite extract and classify ordinary cases, with a JSON validator rejecting invalid results. Escalate only ambiguous or multi-source work to 3.6 Flash. Keep financial, legal, and security-impacting actions for a human; the model can prepare evidence but should not decide.

This does more than reduce cost. It gives you a readable audit trail: why a case escalated, which model ran, which prompt version, what it cost, and whether the output passed validation.

How to read the benchmarks without hype

The numbers above are Google-reported results, not an independent all-model leaderboard. Each test measures a different slice: DeepSWE and Terminal-Bench are closer to software agents, OSWorld to UI use, GDM-MRCR to long context, and GDPval to knowledge work. No high score says how the model handles your invoices, internal shorthand, or messy CRM.

Before migration, freeze 50–100 anonymised real cases. Measure accuracy, valid JSON, tool calls, latency, cost per approved result, and human corrections on the same set. Switch one workflow at a time, not the entire stack.

Sources: Google: Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber and Google AI for Developers: model IDs, pricing, capabilities, and migration notes.