Introduction
On July 21, 2026, Google released Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, both generally available the same day, plus a restricted security model called Gemini 3.5 Flash Cyber. The headline is unusual for a model launch: Gemini 3.6 Flash scores higher than the model it replaces and costs less to run.
Output pricing falls from $9.00 to $7.50 per million tokens while input stays at $1.50. On top of that, Google says the model spends 17% fewer output tokens on the same work, measured across the Artificial Analysis benchmark suite at the time of publication. The two effects compound: for a typical agentic workload, the bill drops by more than the price cut alone suggests.
The release also closes the most-complained-about gap in the Gemini line. Gemini 3.5 Flash shipped in May 2026 with a January 2025 knowledge cutoff, which was old on arrival. Gemini 3.6 Flash moves it to March 2026.
What changed
- Model ID:
gemini-3.6-flash, stable — not a preview label. - Context: 1,048,576 input tokens, 65,536 output tokens. Unchanged from 3.5 Flash. See context window for what those two numbers mean in practice.
- Knowledge cutoff: March 2026, up from January 2025.
- Pricing: $1.50 in / $7.50 out per million tokens.
- Modalities: text, image, video, audio and PDF in; text out. No image or audio generation, no Live API.
- Computer use is now a built-in client-side tool in the Gemini API and Gemini Enterprise rather than a separate model path.
Google frames the improvement as precision rather than raw capability: fewer unwanted code edits, fewer execution loops, and fewer tool calls to finish a multi-step task. That is a token-efficiency story as much as an intelligence one.
Benchmarks
Google-reported scores, comparing the new model to its two predecessors:
| Benchmark | Gemini 3.6 Flash | Gemini 3.5 Flash | Gemini 3.1 Pro |
|---|---|---|---|
| SWE-Bench Pro | 58.7% | 55.1% | 54.2% |
| DeepSWE v1.1 | 49% | 37% | 12% |
| Terminal-Bench 2.1 | 78.0% | 76.2% | 73.8% |
| MLE-Bench | 63.9% | 49.7% | 42.6% |
| GDPval-AA v2 | 1421 Elo | 1349 Elo | 965 Elo |
| OSWorld-Verified | 83.0% | 78.4% | — |
| GDM-MRCR v2 (1M tokens) | 54.0% | under 27% | under 27% |
Two results stand out. MLE-Bench — machine-learning research tasks — jumps 14 points, and the 1M-token long-context score roughly doubles, which matters because a large context window is only useful if retrieval accuracy holds across it. On OSWorld-Verified, a computer-use benchmark, 83.0% is the best score in Google's whole comparison set.
The comparison set is also where the honest caveat lives. Within Google's own table, GPT-5.6 Luna tops DeepSWE (67%) and Terminal-Bench 2.1 (84.7%), Grok 4.5 tops SWE-Bench Pro (64.7%), and Claude Sonnet 5 tops MLE-Bench (66.9%) and GDPval-AA v2 (1607 Elo). Gemini 3.6 Flash is not the best model on most of these — it is a fast, cheap one that lands close to models costing several times more.
Read "tops" narrowly: it means top of the set Google chose to publish. Claude Fable 5, which is absent from that table, scores 80.3% on SWE-Bench Pro — well above every model in it. Vendor comparison charts select their own opposition, and these are vendor-run numbers throughout; independent evaluations may differ.
Gemini 3.5 Flash-Lite
The second model targets volume, not intelligence. Gemini 3.5 Flash-Lite (gemini-3.5-flash-lite) runs at roughly 350 output tokens per second and costs $0.30 in / $2.50 out per million tokens — a fifth of 3.6 Flash's input price.
Its point is that cheap models have caught up: Flash-Lite scores 54.2% on SWE-Bench Pro against 49.6% for Gemini 3 Flash, a full-size model from December 2025 — two Flash releases and seven months back. Google positions it for document processing, high-volume automation, and as a subagent inside larger AI agent systems. It is also rolling out to AI Mode in Google Search.
Gemini 3.5 Flash Cyber
The third model is not on the API. Gemini 3.5 Flash Cyber is tuned to find and fix software vulnerabilities, and Google is releasing it only through CodeMender, in a limited-access pilot for governments and trusted partners. The gating is explicit about the reason: a model good at finding exploitable bugs is useful to both defenders and attackers.
What Google said about what's next
Two forward-looking notes accompanied the release. Gemini 3.5 Pro, announced at Google I/O in May 2026 and still unshipped, is "currently testing with partners" and will launch when ready — a second slip past the original June timeline. And Google says it has "already started our most ambitious pre-training run yet, for Gemini 4."
Conclusion
Gemini 3.6 Flash is an incremental release that does the unglamorous things well: a fresher cutoff, a lower output price, measurably less token waste, and a long-context score that finally justifies the 1M-token window. It does not take the benchmark crown from GPT-5.6 Luna, Grok 4.5 or Claude Sonnet 5, and Google's own charts say so.
If you are already running Gemini 3.5 Flash, the migration is a model-ID change that costs less and returns fresher answers. If you are choosing a model for high-volume inference, Flash-Lite at $0.30 per million input tokens is the more interesting half of this announcement. And Gemini 3.5 Pro remains, two months after its announcement, a model you cannot call.
Sources
- Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber — Google's announcement
- Gemini 3.6 Flash — Google DeepMind model page
- Gemini 3.6 Flash model docs — limits, modalities, supported features
- Gemini API pricing
- Gemini API changelog — July 21, 2026 entries for both models
- Gemini 3.5 — our model page, updated for 3.6 Flash
- Google Gemini — the consumer app and developer platform