Gemini 3.6 Flash, Flash-Lite & Flash Cyber for Claude Users
Google was expected to ship Gemini 3.5 Pro in July 2026. Instead, on 21 July 2026 they shipped three Flash-tier models — Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber — and left the Pro tier in partner testing. If you build on Claude and use Gemini as a cross-check, the shape of that release matters more than any single benchmark: Google is telling the market that the workhorse tier, not the flagship, is where the real money is. This page walks a Claude user through what actually landed, at what price, and what breaks when you port an existing call.
- Understand what shipped on 21 July 2026 — three Flash models, and why Google skipped 3.5 Pro
- Read the price table honestly: 17% cheaper output plus 17% fewer output tokens compounds into a ~30% real cost drop
- Know when Flash-Lite ($0.30 in / $2.50 out, 350 tok/s, 1M context) is the right cross-check versus Claude Haiku
- See how Flash Cyber outscored Claude Opus 4.6 at V8 bug-hunting — and why you almost certainly cannot use it
- Migrate cleanly: which sampling knobs were removed, and what the new thinking_level enum accepts
The one-sentence version
Google's 21 July 2026 release is a Flash-tier refresh with no Pro — a new workhorse (3.6 Flash, $1.50 in / $7.50 out, 1M context, Computer Use built in), a cheaper high-throughput sibling (3.5 Flash-Lite, $0.30 in / $2.50 out, 350 tokens/second), and a security-only fine-tune (3.5 Flash Cyber) that Google is gating to governments and trusted partners via its CodeMender program.
The interesting story is not the model. It is the choice: Google pushed a workhorse and a cheap tier while explicitly holding Pro back, and Logan Kilpatrick said Pro is "currently testing with partners" after missing internal performance targets. That is a different bet than the OpenAI GPT-5.6 launch two weeks earlier, which led with a flagship (Sol) — see the GPT-5.6 Sol/Terra/Luna page for the mirror-image comparison.
Three things about this launch that surprise people
Everyone is writing the "17% cheaper" headline. Here are the details a Claude user should not miss.
1. Google skipped 3.5 Pro — publicly
This is not how Google usually ships. In previous refreshes the flagship led the launch and the smaller siblings followed. On 21 July 2026 Google shipped only Flash-class models and confirmed 3.5 Pro is still in partner testing after missing internal performance goals. Product lead Logan Kilpatrick said the team hopes to "land soon." Read that literally: at the highest tier of intelligence, Google does not yet have a step function it can ship. It does have a strong workhorse and a bug-hunter it can gate. If your workflow depends on "the smartest Gemini you can buy," today that is still 3.5 Pro-Preview or the older 3.5 Ultra tier — not a July-2026 model.
2. The 17% output-price cut is really ~30% because output tokens dropped 17% too
The launch chart shows Gemini 3.6 Flash output pricing at $7.50 per 1M tokens, down from $9 on 3.5 Flash — a 17% cut. Look one row down: Google also reports 17% fewer output tokens on the same workloads (measured on the Artificial Analysis Index). Those compound: 0.83 × 0.83 ≈ 0.69, so real-world output cost drops roughly 31% on comparable jobs, not 17%. On DeepSWE Google claims token reductions "up to 65%." When you cross-check a Claude Sonnet 5 pipeline against 3.6 Flash, do the multiplication before you decide whether to swap — the headline is understating the delta.
3. Flash Cyber beat Claude Opus 4.6 at V8 bug-hunting — but you almost certainly cannot use it
Google reports that Gemini 3.5 Flash Cyber discovered 55 unique confirmed vulnerabilities on internal V8 JavaScript engine benchmarks, versus 47 for standard 3.5 Flash and 36 for Claude Opus 4.6. Ten of those bugs were found by no other model. On the CyberGym benchmark it lands at 83.2%, close to Claude Mythos 5 (83.8%) and GPT-5.5-Cyber (85.6%). Impressive — and gated. Flash Cyber ships only through Google's CodeMender program to governments and trusted enterprise partners. There is no public API, no pricing, no self-serve access. If you saw a headline that said "Google's cyber AI beat Claude" and expected to try it: you cannot. The takeaway a defender should extract is not "swap models"; it is that fine-tuning a mid-tier model on a narrow domain now beats a bigger general-purpose model on that domain. That pattern will repeat.
The tier table (as of 21 July 2026)
| Model | Input $/1M | Output $/1M | Cached input $/1M | Context | Max output | Notes |
|---|---|---|---|---|---|---|
| Gemini 3.6 Flash | $1.50 | $7.50 | $0.15 | 1,048,576 | 65,536 | Workhorse; Computer Use built in |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | (not stated) | 1,048,576 | 65,536 | 350 tokens/second output; rolling out in Google Search |
| Gemini 3.5 Flash Cyber | (not public) | (not public) | (not public) | (not public) | (not public) | Governments & trusted partners via CodeMender only |
Benchmark deltas Google published for 3.6 Flash vs 3.5 Flash (higher is better):
- DeepSWE: 49% vs 37% (+12 points)
- MLE Bench: 63.9% vs 49.7% (+14.2 points)
- OSWorld-Verified: 83.0% vs 78.4% (+4.6 points)
- GDPval-AA v2: 1421 Elo vs 1349 (+72 Elo)
- GPQA Diamond: 92.8%
- MMMU-Pro: 83.2%
- Humanity's Last Exam: 38.3%
- Chatbot Arena Elo (Text Overall): 1485
- Artificial Analysis Coding Index: 69.2%
Two knobs are more useful than any single benchmark row. Knowledge cutoff jumped from January 2025 to March 2026 — a 14-month leap. For "answer from the model directly" tasks (no RAG, no browsing) this is often bigger than a benchmark score. Output speed is quoted at 304 tokens/second for 3.6 Flash and 350 tokens/second for Flash-Lite, which changes how much latency-hiding your UI has to do.
Flash-Lite is the real story for high-volume Claude workflows
For high-volume, latency-sensitive Claude workflows — classification, extraction, moderation, autocomplete — Flash-Lite at $0.30/$2.50 with 1M-token context and 350 tok/s is now the obvious cross-check against Claude Haiku 4.5. It is not a Sonnet replacement, and you should not use it for planning-heavy agent loops. But if you have a "cheap model behind a queue" tier in your stack, budget a day to A/B it.
Migration: what a Claude user actually has to change
The Gemini API surface changed with 3.6. If you carry a Claude-shaped call over, four things break.
- Use `gemini-3.6-flash` for the workhorse or `gemini-3.5-flash-lite` for the cheap tier. There is no public model ID for 3.5 Flash Cyber. 3.5 Pro is not shipping yet — do not pin to it.
- These sampling knobs were removed from 3.6 Flash and 3.5 Flash-Lite. If your Claude-side code carries them over via a shim, Gemini will reject or ignore them depending on the SDK. Delete them; the model handles sampling internally now.
- The integer-tokens `thinking_budget` field was replaced with a `thinking_level` enum. Flash defaults to medium; Flash-Lite defaults to minimal. Set it explicitly if you were tuning budgets — you cannot carry over the old integer.
- 3.6 Flash rejects prefilled model turns (a conversation ending on a non-empty `model` role). If you were using Anthropic-style prefill to steer output, restructure — Gemini will 400 the request.
- `candidate_count` is unsupported on 3.6 Flash. If you were requesting multiple completions in one call, call the API N times or move that job to a different tier.
- In 3.6 Flash, computer use — operating a GUI — is now a *client-side built-in tool* in the Gemini API and Gemini Enterprise, not a separate integration. Enable it in the tools block rather than plumbing a separate SDK. Anthropic's equivalent lives in the [Claude computer-use section](/docs/models/computer-use-agents).
Minimal 3.6 Flash call (Python SDK)
from google import genai
client = genai.Client() # picks up GOOGLE_API_KEY
resp = client.models.generate_content(
model="gemini-3.6-flash",
contents="Extract the 3 riskiest lines from this diff and cite line numbers.",
config={
"thinking_level": "medium", # was thinking_budget=<int>
# do NOT pass temperature, top_p, top_k, candidate_count
"tools": [{"computer_use": {}}], # built-in, not a separate SDK
},
)
print(resp.text)When to reach for Gemini 3.6 Flash over Claude — and when not
- Reach for 3.6 Flash: high-volume tool loops where token efficiency dominates cost, and workflows that benefit from the March 2026 knowledge cutoff (14 months fresher than the older 3.5 Flash).
- Reach for Flash-Lite: latency-sensitive classification/extraction on 1M-token inputs where Haiku is your baseline.
- Stay on Claude: SWE-Bench-Pro-style real-repo bug fixes, long agent runs where extended thinking discipline matters, and anywhere you already trust Sonnet 5 planning quality.
- Do not chase Flash Cyber: unless you are a government or a CodeMender partner, treat its numbers as a research signal, not a purchasing decision.
For a broader Gemini-vs-Claude framing that predates this release, see Gemini for Claude users and the coding-focused Claude vs GPT vs Gemini for coding. For the mirror-image OpenAI launch two weeks earlier, GPT-5.6 Sol/Terra/Luna. For the cross-model porting rules that make any of these swaps survive contact with production, porting prompts across models.
Quick check
Check yourself
0/4Sources & further reading
- Google DeepMind — Introducing Gemini 3.5 Flash Cyber
- Gemini API pricing (official)
- Google AI Studio
- TechCrunch — Google releases three new Gemini models, but no 3.5 Pro
- SiliconANGLE — Google expands Gemini with cheaper models and a bug-hunter it keeps on a leash
- MarkTechPost — Google releases Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
- The Hacker News — Google Launches Gemini 3.5 Flash Cyber AI to Find and Fix Software Vulnerabilities
- Help Net Security — Google's Gemini 3.5 Flash Cyber becomes a vulnerability hunter