Verdict: For most small businesses shipping production AI agents, Gemini 3.5 Flash-Lite is the one you should run first — it's 5× cheaper per token than Gemini 3.6 Flash at $0.30/1M input vs $1.50/1M, and it's the fastest model in the 3.5 series at 350 output tokens/s. Reach for Gemini 3.6 Flash when you genuinely need the step-up in coding, knowledge work, or multi-tool reasoning — Google's own benchmarks show 3.6 Flash pulls ahead on every shared eval, but in a cost-routed agent setup, Flash-Lite handles most daily tasks for a fraction of the budget.
On July 21, 2026, Google shipped three new Gemini models in a single drop: Gemini 3.6 Flash (the new mid-tier workhorse), Gemini 3.5 Flash-Lite (the low-latency, high-volume model), and Gemini 3.5 Flash Cyber (a security-trained variant behind a gated pilot). Every headline you read about this launch will fixate on benchmark scores. The real story is simpler: Google is making AI cheaper to run at scale. When running AI gets cheaper, people run more of it — and that's what unlocks agent teams, not one better chatbot.
What's the difference between Gemini 3.6 Flash and 3.5 Flash-Lite?
Gemini 3.6 Flash is Google's new mid-tier workhorse — better at coding, knowledge work, and multi-step agentic workflows than the 3.5 Flash it replaces. Gemini 3.5 Flash-Lite is the stripped-down, fastest, cheapest model in the 3.5 series, explicitly tuned for high-throughput, low-latency tasks like classification, document processing, and translation.
The two share a 1M-token context window and multimodal inputs (text, image, audio, video). They differ on price, speed, and reasoning depth:
| Spec | Gemini 3.6 Flash | Gemini 3.5 Flash-Lite |
|---|---|---|
| Input price (per 1M tokens) | $1.50 | $0.30 |
| Output price (per 1M tokens) | $7.50 | $2.50 |
| Context window | 1M tokens | 1M tokens |
| Output speed (per provider reports) | rumored ~280 tok/s | ~350 tok/s (Artificial Analysis) |
| Intelligence Index (Artificial Analysis) | higher on every shared benchmark | scores 36 on the AA Index (mid-tier) |
| Thinking levels | fixed | configurable (minimal/low/high) |
| Computer use | built-in | built-in |
| Knowledge cutoff | March 2026 | March 2026 |
Sources: Google Gemini API pricing, Artificial Analysis – Gemini 3.5 Flash-Lite, Google blog announcement.
How much does each model actually cost to run?
Gemini 3.5 Flash-Lite costs $0.30 per 1M input tokens and $2.50 per 1M output tokens. Gemini 3.6 Flash costs $1.50 per 1M input tokens and $7.50 per 1M output tokens — 5× more on input and 3× more on output, per Google's official API pricing page.
The lever that matters more than the absolute price is the token-efficiency jump in 3.6 Flash. On the Artificial Analysis Index, 3.6 Flash consumes 17% fewer output tokens than 3.5 Flash for the same task, and on the DeepSWE agentic-coding benchmark (by Datacurve) Google reports up to a 65% reduction in output tokens, per Google's announcement.
In plain English: 3.6 Flash costs more per token, but it uses fewer tokens to do the same job. 3.5 Flash-Lite costs less per token and is built to fire lots of small jobs quickly. They're solving different problems.
A worked cost example
Say you run a small-business support operation that processes 100,000 customer messages a day. Each message needs ~500 input tokens (the message + context) and produces ~200 output tokens (a triage tag, a draft reply, or a routing decision).
| Quantity | Gemini 3.6 Flash cost/day | Gemini 3.5 Flash-Lite cost/day |
|---|---|---|
| 100k messages × 500 input tokens = 50M input | $75.00 | $15.00 |
| 100k messages × 200 output tokens = 20M output | $150.00 | $50.00 |
| Daily total | $225.00 | $65.00 |
| 30-day monthly total | $6,750 | $1,950 |
At high volume, routing to Flash-Lite saves ~$4,800/month — and Flash-Lite is faster, so your customers also wait less. That's the case for routing routine work to the smaller model. We've covered this routing logic in more depth in Which AI model for which task? and in Should You Switch to Gemini 3.6 Flash?.
What is Gemini 3.6 Flash best at?
Gemini 3.6 Flash is best at coding, knowledge work, and multi-step agentic workflows where it takes fewer reasoning steps and tool calls to finish — the step-up over 3.5 Flash is meaningful and measurable. Per Google's blog post, 3.6 Flash delivers:
- DeepSWE (agentic coding): 49% vs 37% on 3.5 Flash.
- MLE-Bench (ML research): 63.9% vs 49.7%.
- OSWorld-Verified (computer use): 83.0% vs 78.4%.
- GDPval-AA v2 (knowledge work): 1421 vs 1349.
It also ships with enhanced Frontier Safety safeguards in Chemical, Biological, Radiological, and Nuclear (CBRN) and cyber-offense domains, and is "substantially more resistant to jailbreaks while trained to minimize refusals for beneficial uses" per its DeepMind model card.
Use 3.6 Flash when:
- You're building agents that call multiple tools in a row and you need fewer dead-end tool calls.
- You're writing or reviewing code and need the coding/ML-research capability.
- You're parsing long documents or charts and synthesizing a report.
- You explicitly need the better quality of answer and can afford ~3-5× the per-token cost.
What is Gemini 3.5 Flash-Lite best at?
Gemini 3.5 Flash-Lite is best at high-volume, low-latency tasks where reasoning depth isn't the bottleneck — classification, routing, summarization, tagging, document extraction, and translation. Per Google's announcement, it's the fastest model in the 3.5 series at 350 output tokens/s per Artificial Analysis, and it ships with configurable thinking levels: prioritize low-latency execution for high-volume tasks, or engage higher thinking for multi-step subagent workloads.
Its benchmark jumps over prior Flash-Lite generations are dramatic:
- Terminal-Bench 2.1: 54% vs 31% on the prior Flash-Lite.
- GDM-MRCR v2 (long context): 72.2% vs 60.1%.
- GDPval-AA v2: 1140 vs 642.
- SWE-Bench Pro: 54.2% — even beats the bigger Gemini 3 Flash (49.6%).
- OSWorld-Verified: 74.0% vs 65.1% on Gemini 3 Flash.
Use 3.5 Flash-Lite when:
- You're running thousands of small tasks a day (sorting emails, reading invoices, tagging messages, pulling data out of PDFs).
- Latency matters — the model's faster throughput means agents finish sooner.
- Cost matters — 5× cheaper input, 3× cheaper output means you can afford 3-5× more agents on the same budget.
- You want to scale a multi-agent team where most agents are narrow, fast specialists.
For a worked example of how to do this kind of multi-agent setup, see How to Run a Small Business on AI Agents: The Delegation Framework.
What is Gemini 3.5 Flash Cyber for?
Gemini 3.5 Flash Cyber is Google's first model fine-tuned specifically for cybersecurity — finding, validating, and patching software vulnerabilities in code. Per the DeepMind introduction blog, it's built on top of 3.5 Flash and is specifically trained to outperform mainline Flash models at vulnerability discovery, often used together with Google's CodeMender agent.
Key facts from the DeepMind blog:
- CodeMender invokes 3.5 Flash Cyber multiple times (up to 5) per report, allowing agents to analyze hugely more code paths at a lower price per token than a single expensive call to a massive LLM.
- On the CyberGym benchmark it achieves "competitive performance against significantly larger models" thanks to the low per-call cost.
- On the Big Sleep evaluation (focused on vulnerabilities in Chrome and Safari), 3.5 Flash Cyber significantly surpassed both 3.5 Flash and 3.6 Flash.
- In a real case study, Google's Cloud Vulnerability Research team used it to find RCE vulnerabilities in public APIs and a memory-corruption bug in a production service — all in 2 hours — and generated a 100% reliable RCE exploit that bypassed ASLR and W^X mitigations.
Who gets access to Flash Cyber — and who doesn't
This is the model most beginners will not get to touch directly. Given its dual-use nature, 3.5 Flash Cyber is exclusively available to governments and trusted partners via CodeMender as part of a limited-access pilot, expanding over time (DeepMind blog). CodeMender's foundational capabilities are also available to enterprise customers through the mainline Gemini models on the Gemini Enterprise Agent Platform.
Bottom line for small businesses: If you're not running a security team auditing codebases, Flash Cyber isn't for you. But the direction it signals — specialist models fine-tuned for one job at a lower cost than a frontier model — is the same direction Google is taking 3.6 Flash and 3.5 Flash-Lite. The whole release is built on the same idea: smaller, faster, job-specific beats bigger and general.
How do you pick between the three for a real workflow?
For a typical small business or solo builder, here is the decision framework:
- Start with Gemini 3.5 Flash-Lite for the boring, high-volume stuff. Email triage, document extraction, routing, tagging, classification, simple Q&A against your own docs. Most of the work you need AI to do is boring. That's where Flash-Lite wins on both cost and speed.
- Add Gemini 3.6 Flash as the "brain" for complex tasks. Multi-step agent reasoning, code generation, deep document analysis, anything that benefits from fewer tool calls and better reasoning steps. Use it sparingly — like a senior reviewer on the high-leverage tasks the small models route up to it.
- Use Gemini 3.5 Flash Cyber only if you have a code-security problem and you're in the gated pilot / CodeMender through the Enterprise Agent Platform. For everyone else, it's a signal about where the market is going, not a tool you'll deploy today.
This routing pattern — small, fast, cheap models do the high-volume handling, and an "expert" model is invoked only when small model confidence drops or the task demands it — is the cheapest and fastest architecture for production agents right now. We've covered the broader stack for this in How to Build a Customer Onboarding Dashboard With No Code and How to Use Google AI Studio With Gemini 3.6 Flash.
What this means for you
If you're a small business owner or a builder shipping AI things, the July 2026 Gemini drop is the moment to stop thinking "which AI model is best" and start thinking "which AI model for which job." Token efficiency is the unlock for running agent teams — not one better chatbot, but five specialists running in parallel, each doing the boring small jobs they're best suited for at the lowest cost per call.
The single highest-leverage move right now: pick one boring, repetitive task in your business. Something you do every week. Give it to Gemini 3.5 Flash-Lite. Get it working. Then add another. Then add Gemini 3.6 Flash as the senior reviewer that handles the harder cases the small model escalates. That's how you build a real AI system — not all at once, one boring job at a time.
FAQ
Q: How cheap is Gemini 3.5 Flash-Lite compared to Gemini 3.6 Flash? A: Flash-Lite is $0.30 per 1M input tokens and $2.50 per 1M output tokens. 3.6 Flash is $1.50 per 1M input and $7.50 per 1M output — 5× pricier on input and 3× pricier on output. At high volume, routing routine work to Flash-Lite can cut your monthly invoice by 3-5×. Source: Google Gemini API pricing.
Q: Is Gemini 3.6 Flash worth the extra cost? A: For tasks that need it — coding, multi-step agentic reasoning, complex document synthesis — yes. 3.6 Flash beats 3.5 Flash on every shared benchmark Google publishes (DeepSWE 49% vs 37%, MLE-Bench 63.9% vs 49.7%, OSWorld-Verified 83.0% vs 78.4%). For routine triage, classification, and drafting, Flash-Lite is the cheaper-and-faster pick. Source: Google blog announcement.
Q: Can I try Gemini 3.6 Flash and 3.5 Flash-Lite for free? A: Yes — both are available in Google AI Studio and on the free tier of the Gemini Developer API. 3.5 Flash-Lite has free-tier input/output and limited grounding-with-Google-Search (5,000 prompts/month free), per the pricing page.
Q: Do I need Gemini 3.5 Flash Cyber? A: Almost certainly not, unless you're vetting software vulnerabilities at scale. Flash Cyber is restricted to governments and trusted partners via CodeMender as part of a limited-access pilot, per the DeepMind blog. For readers of this site, the takeaway is that Google is investing in specialist, cost-efficient models — the same thesis behind 3.6 Flash and 3.5 Flash-Lite.
Q: What's the context window on these models? A: Both Gemini 3.6 Flash and Gemini 3.5 Flash-Lite have a 1M-token context window, per their Artificial Analysis profiles. 3.6 Flash specifies 1,048,576 input tokens and 65,536 output tokens.
Q: Is Gemini 3.5 Pro out yet? A: No. In the July 21, 2026 announcement, Google said Gemini 3.5 Pro is "currently testing with partners" with broad availability planned "as soon as it's ready" — and disclosed that they've started their "most ambitious pre-training run yet" for Gemini 4. Source: Google blog announcement.
Discussion
0 comments