DEV Community

Induwara Ashinsana
Induwara Ashinsana

Posted on • Originally published at induwara.lk

Gemini 3.6 Flash & 3.5 Flash-Lite: the token math

Gemini 3.6 Flash landed on 21 July 2026 alongside two siblings, and the announcement everyone quoted was the benchmark chart. I want to talk about the boring number underneath it: output tokens dropped 17% for the same work. When you pay per token in dollars and earn or study on a rupee budget, that line matters more than any leaderboard.

Google's post, Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber, covers three models at once. Here's my read on which one you should actually reach for.


🗂️ Three models, one page — what each is for

The launch bundled a workhorse, a speed model, and a locked-down security model. Quick map before the numbers:

Model Job Who can use it
Gemini 3.6 Flash General coding, knowledge work, multimodal Everyone (AI Studio, Gemini app, Enterprise)
Gemini 3.5 Flash-Lite Fastest, cheapest 3.5-class; high-throughput agents Everyone; rolling into Google Search
Gemini 3.5 Flash Cyber Finding and fixing security vulnerabilities Governments + trusted partners only

Key takeaway: Two of these three are for you. The third — the interesting one for security folks — you can't touch yet.


💰 The pricing that actually decides your bill

Here are the published rates, per one million tokens:

Model Input / 1M Output / 1M
Gemini 3.6 Flash $1.50 $7.50
Gemini 3.5 Flash-Lite $0.30 $2.50

Output is where LLM bills quietly balloon, because a chatty agent generates far more tokens than you feed it. Flash-Lite's $2.50 output is a third of 3.6 Flash's $7.50. For a side project or a class assignment, that gap is the difference between a hobby and a habit.

At roughly Rs 300 to the dollar (check the live rate, it moves), a million Flash-Lite output tokens costs around Rs 750. A million on 3.6 Flash is closer to Rs 2,250. If you're not sure how many tokens your prompts actually burn, run them through our free AI Token Counter — it covers GPT, Claude, and Gemini and needs no signup.


⚡ Why "17% fewer tokens" beats a benchmark win

Google says 3.6 Flash cuts output token usage 17% versus 3.5 Flash, and up to 65% on the DeepSWE benchmark. That's not a vanity stat. Fewer tokens for the same answer means every request is cheaper, even before you compare per-token prices.

The benchmarks did move too:

Benchmark 3.5 Flash 3.6 Flash
DeepSWE 37% 49%
MLE Bench 49.7% 63.9%
OSWorld-Verified 78.4% 83.0%

But a coding agent that loops ten times to fix a bug pays ten output bills. Shaving 17% off each one compounds. I'd rather have a model that's slightly less flashy on a chart and consistently cheaper per task, because my monthly cost is the sum of every loop, not the peak score.

A cheaper token on every call scales better than a higher benchmark you hit once.


🚀 Flash-Lite: the throughput model for tight budgets

Flash-Lite is the one I'd start a project on. Google clocks it at 350 output tokens per second and pitches it for high-throughput agentic work and document processing. Its own benchmark jumps are steep:

  • Terminal-Bench 2.1: 31% → 54%
  • SWE-Bench Pro: 49.6% → 54.2%

For the kind of work a small Colombo team or a solo student actually ships — parsing PDFs, tagging support tickets, batch-summarising documents — speed and price beat frontier reasoning. Named customers like Ashler, Palo Alto Networks, and Ramp flagged exactly that: they picked it to scale volume, not to win a hard reasoning puzzle.

  1. Prototype on Flash-Lite to keep costs near zero while you iterate.
  2. Move a route to 3.6 Flash only if Lite visibly struggles on it.
  3. Measure real token counts before you commit to either.

🔒 Flash Cyber: the model you can't have (and why that's notable)

Gemini 3.5 Flash Cyber is a fine-tuned 3.5 Flash aimed at finding and fixing security vulnerabilities, offered only to governments and trusted partners through a limited-access pilot called CodeMender.

Bottom line: A frontier lab is now treating vulnerability-finding as capable enough to gate behind vetting. That's a signal worth watching, even if you're locked out of the tool.

For a student here building a security portfolio, the takeaway isn't the model — it's the direction. Automated vuln-finding is graduating from research demo to controlled release. The public models above are strong enough to practise the same workflow: point one at your own code, ask it to audit for common flaws, and learn to read what it flags.


💡 What this means for you

If you build, study, or freelance from Sri Lanka on a dollar-priced AI bill, the practical order of operations is simple:

  • Default to Flash-Lite. At $0.30/$2.50 per million tokens, it's the cheapest way to keep a project alive while you learn.
  • Reserve 3.6 Flash for the routes that genuinely need more, and lean on its 17% token cut to keep those cheaper than they used to be.
  • Count before you commit. Estimate real usage with a token counter so your first invoice isn't a surprise.
  • Don't chase Flash Cyber. You can't get it, and the public models are enough to build the same security-review muscle.

The headline was three models. The story, for anyone paying per token, is that the floor on running useful AI just dropped again.

Top comments (0)