Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber

TL;DR Highlight

Gemini 3.6 Flash cuts output tokens 17% and price; 3.5 Flash-Lite hits 350 tok/s.

Who Should Read

Backend developers running agents on the Gemini API who need to choose models and control token costs.

Core Mechanics

3.6 Flash uses fewer tokens
- Per the Artificial Analysis Index, output tokens are 17% lower than 3.5 Flash.
- On some benchmarks like DeepSWE, the reduction reaches up to 65%.
- It also takes fewer reasoning steps and tool calls in multi-step workflows.
3.6 Flash is cheaper
- Pricing is $1.50/1M input and $7.50/1M output tokens.
- Google says this is lower than 3.5 Flash.
3.6 Flash beats 3.5 Flash on benchmarks
- DeepSWE 49% vs 37%, MLE Bench 63.9% vs 49.7%.
- OSWorld-Verified 83.0% vs 78.4%, GDPval-AA v2 1421 vs 1349.
- Computer use is now a built-in client-side tool in the Gemini API.
3.5 Flash-Lite is the fastest 3.5 model
- It runs at 350 output tokens/s per Artificial Analysis.
- Pricing is $0.3/1M input and $2.5/1M output tokens.
- Minimal and low thinking levels favor latency and cost, while higher levels suit multi-step subagent work.

Community Reactions

Complaints about rising prices
- A developer who uses 2.5 Flash-Lite for bulk document processing noted 3.5 Flash-Lite costs 6.25x more than 2.5.
- Several users worried about older versions being deprecated.
Criticism of missing comparisons
- One view said it costs more than GLM-5.2 and looks worse.
- Others noted the post compares only against previous Gemini versions.
Opinions on speed and frontend strength
- One commenter said it is mid-pack on Artificial Analysis but very fast.
- Another found 3.5 Flash better than GPT 5.5 at frontend work in Antigravity and good for iteration.
A positive take on 3.5 Flash-Lite
- One commenter tested it as comparable to Opus 4.8 on a few tests, 6x faster and 19x cheaper.
Complaints about Google's product handling
- One company left Antigravity after the AI Ultra subscription was phased out.

How to Apply

A/B test Flash-Lite thinking levels
- Pick 50 prompts from your bulk workload and run them on 3.5 Flash-Lite at minimal and low.
- Compare quality, latency, and token cost against your current model in a table.
Split agents into orchestrator and workers
- Use 3.6 Flash for planning and 3.5 Flash-Lite for search and document-processing subagents.
- Google's own demo used 3.6 Flash as master agent with Flash-Lite below it.
Re-measure cost moving to 3.6 Flash
- Compare output tokens and tool-call counts on the same agent task logs.
- Compute cost per task and decide whether to switch.
Check deprecation dates and price changes
- Look up the deprecation schedule for the Flash-Lite version you use.
- Calculate how the price increase affects your budget.

Try It

copy & run

# Paste into Claude/ChatGPT: model switch cost calculation prompt

I run agents on the Gemini API. Using the prices below, calculate cost per task for each model.
- 3.5 Flash: input $1.5/1M, output $9/1M
- 3.6 Flash: input $1.5/1M, output $7.5/1M
- 3.5 Flash-Lite: input $0.3/1M, output $2.5/1M

My average task: {{input tokens}} input, {{output tokens}} output, {{count}} tasks per day.
Calculate both with and without assuming 3.6 Flash uses 17% fewer output tokens than 3.5 Flash.

Next: summarize as a monthly cost table and suggest criteria for moving simple tasks to Flash-Lite.

Terminology

Flash-LiteThe lightweight line in the Flash series, tuned for speed and price. Think of it as the cheap tier for high-volume work.

thinking levelA setting for how much the model thinks before answering. Higher can be more accurate but slower and costlier.

CodeMenderGoogle's code security agent. It finds, validates, and patches vulnerabilities.

computer useA capability where the model views the screen and operates a computer by clicking and typing.

Artificial Analysis IndexAn external benchmark site's composite metric comparing model intelligence, speed, and price.