Three Flash variants, one agent loop: picking between Gemini 3.6, 3.5-Lite, and Flash Cyber

What it is

Three Gemini Flash variants shipped together on July 21, 2026. Gemini 3.6 Flash is positioned as the new default workhorse model and the successor to Gemini 3.5 Flash. 3.5 Flash-Lite is the cheap-and-fast tier. And Gemini 3.5 Flash Cyber is a lightweight cybersecurity model based on Gemini 3.5 Flash, with additional training to improve its ability to discover, verify, and fix software vulnerabilities.

The pitch to developers: pick per node in your agent graph, not per app.

The next-work-session test

Concrete scenario: you have a code-review agent loop — planner, patcher, verifier, security check — currently all running on 3.5 Flash. In one session you can route the planner and patcher to 3.6 Flash, drop the verifier's tool-picking node to 3.5 Flash-Lite, and hand the security-check subagent to Flash Cyber. Then re-run your eval suite.

What changes: your per-run token bill and your security recall, in opposite directions. Whether that math works is what your evals are for — don't take it on trust.

Pricing

Verified via multiple third-party trackers, all consistent with Google's published tier:

• Gemini 3.6 Flash: $1.50 per million input tokens, $7.50 per million output tokens, on a 1M-token context window with 65,536 max output.
• Gemini 3.5 Flash-Lite: $0.30 per million input tokens, $2.50 per million output tokens, same context window.
• Gemini 3.5 Flash Cyber: pricing not separately published in the sources we checked — treat as unverified until Google lists it on the API pricing page.

Note the price move worth knowing: the input price remains unchanged, while the output price falls from $9.00 to $7.50 versus 3.5 Flash. Google also claims 3.6 Flash produces ~17% fewer output tokens on representative workloads — that's a vendor claim, not a benchmark you should trust before running your own.

What we'd actually use it for

Honestly, most of us aren't rebuilding the whole stack. The realistic move is narrower:

Limits

Flash Cyber is not a general assistant. On the V8 JavaScript engine, it found 55 unique confirmed issues, including 10 that Gemini 3.5 Flash and Claude Opus 4.6 did not detect. Google uses the model to identify and fix vulnerabilities across internal codebases — impressive, but note it was evaluated on CyberGym, an evaluation developed by its Big Sleep team, i.e. Google evaluating Google. Independent evals still pending.

Other limits:

• The "17% fewer tokens" number is a Google claim on their workload distribution. Your traffic isn't their traffic.
• 3.6 Flash output pricing is lower but still well above Flash-Lite — don't route your entire agent graph to it by default.
• Flash Cyber's pricing and access model aren't clearly published; assume enterprise gating until you see a public API page.
• Migration isn't free: prompts tuned for 3.5 Flash may need re-tuning. Third-party guides like CometAPI's migration writeup exist, but you'll want your own regression suite.

Try it if

• You run an agent loop on 3.5 Flash today and have an eval harness that can A/B two models cleanly.
• You have a high-volume classification or routing node and pricing dominates your bill.
• You maintain a codebase where a security-focused subagent doing patch suggestions would meaningfully offload human review.
• You care about output-token economics on long-horizon tasks (the VentureBeat coverage has the vendor framing).

Skip it if

• You're on Gemini 3.1 Pro or a competitor and your quality bar is above what any Flash tier delivers — wait for 3.5 Pro.
• You don't have evals. Swapping models blind on a production agent is how you learn what "silent regression" means.
• Your security workflow already uses a dedicated SAST/DAST stack — Flash Cyber is a complement, not a replacement, and the Help Net Security writeup is honest about the confirmed-vs-reported gap.
• You need transparent, published pricing before you can commit — Flash Cyber isn't there yet.