Gemini Flash · New Tiers
Gemini Flash just split
three ways.
Google DeepMind announced on the DeepMind Blog that Flash is being replaced by three purpose-built models: 3.6 Flash (speed), 3.5 Flash-Lite (cost), and 3.5 Flash Cyber (security). The 3.5 speed tier introduced at I/O on 5/20 is now sub-divided by use case. The "one Flash SKU for everything" era ends today.
What Changed
From "one model, three dials" to "three SKUs"
Flash used to balance speed, cost, and quality in a single SKU. Today it splits into three distinct models, so the trade-offs are picked by SKU rather than by prompt.
Per DeepMind's Gemini page, the new 3.6 Flash improves response latency versus 3.5 Flash and increases parallel token throughput — the next-generation speed tier. 3.5 Flash-Lite is a distilled, cheaper model targeted at tasks "Flash overshoots" — classification, summarization, tagging. 3.5 Flash Cyber is a security-hardened tier trained to resist prompt injection, decompose malicious instructions, and emit provenance-friendly logs.
All three branch from Gemini 3.5's pre-training. DeepMind describes the split as decoupling cost and speed into different axes of the Gemini API. Where developers used to trade one against the other on a single dial, they now pick a SKU that maximizes one without penalizing the other.
Stop asking one Flash to do three jobs.
Pick the SKU that fits the job.
How To Choose
Choosing between them is refreshingly simple
Three SKUs exist to reduce, not increase, decision fatigue. Three scenarios cover most of the cases.
3.6 Flash — chat UI, agents
Where "sub-second first token" is the requirement. DeepMind's model card emphasizes TTFT (Time To First Token), so this is the core for chat, voice, and live-assist. Per-tier benchmarks live in the Gemini API models docs.
3.5 Flash-Lite — batch / high-volume classify
Per-request cost dominates. RAG index generation, pre-embed summarization, internal-doc tagging: the "bulk work where a small quality dip is fine." Complex instruction following degrades — do not use for the primary chat brain.
3.5 Flash Cyber — untrusted-input gate
First layer over user input, email bodies, web pages, PDFs. Trained to decompose prompt-injection patterns and neutralize them, then hand structured, clean output to the downstream 3.6 or Pro model.
Why It Matters
The "one general model for everything" ceiling breaks — from the price side
The SKU split is a milestone in the industry's slow drift from "one GPT for all" toward "workload-tuned models."
Until today, developers on a tight budget only had "Flash." That meant overpaying on simple tasks and under-delivering on complex ones. With the split, per-request cost drops around 40% on Flash-Lite, while quality-critical paths concentrate on 3.6 Flash. Cost and quality now move on independent axes inside the same Gemini wallet.
Second, spinning out a security-hardened SKU is unusual industry practice. Among the frontier labs, Google is the first to sell "resists prompt injection" as a distinct model SKU. It's a statement: this defense should be the model's job, not the prompt engineer's.
Who Feels It
Who this hits — and how
Per-persona notes on picking and switching.
Engineers
Existing 3.5 Flash calls are wire-compatible, but a new branch shows up in your code: which SKU do I invoke? A common design is Cyber for classification of raw input, 3.6 for the reply, Lite for background jobs. Switching is a one-line change — model="gemini-3.6-flash".
PMs / business
Time to revisit the LLM-cost model. Batch-heavy products can realistically cut LLM spend 30–40% by shifting eligible tasks to Flash-Lite. Chat-first products can push 3.6 as a perceived-speed differentiator. Same wallet, less friction to migrate.
Security
Fronting untrusted input with Cyber neutralizes many injection patterns at the model layer. It doesn't remove the need for downstream system prompts and permission scoping — just moves a large class of first-stage attacks out of your prompt-engineering surface.
Caveats
Three points not to over-read
"Three SKUs solve everything" is over-selling it. Watch for these when you migrate.
1) Do not treat Cyber as a silver bullet. Google itself calls it a "detect and decompose" model, not a "stop everything" one. Your existing permission design, input sandboxing, and output review still stand.
2) Flash-Lite's failure mode is subtle. Multi-step reasoning and precise instruction following degrade visibly. Use it for bulk simple work — don't put it on the main chat path or you'll see refusal-shaped drift.
3) 3.6 does not replace Pro. Even as 3.6 Flash edges up on general benchmarks, complex reasoning and long-context still belong to 3.5 Pro. Reasoning-heavy workloads stay on Pro-family models; Flash is the speed-and-volume lane.