Gemini 3.6 Flash and 3.5 Flash-Lite are now generally available
Google's new Gemini 3.6 Flash uses 17% fewer output tokens, scores 83% on computer use, and costs $1.50/$7.50 per million tokens.
On 21 July 2026, Google made two new Gemini models generally available: Gemini 3.6 Flash (gemini-3.6-flash) and Gemini 3.5 Flash-Lite (gemini-3.5-flash-lite). Each model targets a different part of the cost-performance spectrum, and between them they cover most practical production workloads. Here is what each one does, what it costs, and what the changes mean in practice.
Gemini 3.6 Flash: the new everyday workhorse
Google is positioning 3.6 Flash as the default model in the Gemini family, the one you should reach for first when building coding tools, knowledge-work pipelines, or multimodal applications. The headline improvements over the outgoing 3.5 Flash are meaningful rather than incremental.
Token efficiency. On the Artificial Analysis Index, 3.6 Flash produces 17% fewer output tokens than 3.5 Flash for equivalent tasks. It also takes fewer reasoning steps and fewer tool calls on multi-step agentic workflows. That matters for your bill as much as for latency.
Pricing. The rate drops slightly on output: $1.50 per million input tokens and $7.50 per million output tokens, compared with $1.50/$9.00 for 3.5 Flash. Cached input costs $0.15 per million tokens, with a storage fee of $1.00 per million tokens per hour. Put those two changes together, a lower per-token rate on top of fewer tokens consumed, and real-world costs for complex tasks could fall noticeably compared with running 3.5 Flash.
Knowledge cutoff. The training data now runs to March 2026, up from January 2025 on 3.5 Flash. For knowledge-work use cases where recency matters, that extra 14 months of coverage is a practical improvement.
Computer use. Google’s OSWorld-Verified score rises from 78.4% on 3.5 Flash to 83% on 3.6 Flash. Computer use is also now a built-in client-side tool via the Gemini API and Gemini Enterprise, rather than something you need to wire together yourself.
Benchmark highlights. On SWE-Bench Pro (coding), 3.6 Flash scores 58.7% versus 55.1% for 3.5 Flash. On DeepSWE it reaches 49% against 37%, and on MLE-Bench 63.9% against 49.7%. Knowledge work on GDPval-AA v2 improves from 1,349 to 1,421.
Both models share a one-million-token input context window and a 64,000-token output limit. Text, images, audio, and video are all supported as inputs. One thing worth noting: thinking tokens are billed at the output-token rate, so a short visible response can carry more billable tokens than the final answer length implies.
Gemini 3.5 Flash-Lite: fast, cheap, capable
Flash-Lite is Google’s answer to high-throughput, latency-sensitive workloads: agentic search pipelines, document processing at scale, consumer product backends. The Artificial Analysis Index puts it at 350 output tokens per second, making it the fastest model in the 3.5 class.
Pricing is $0.30 per million input tokens and $2.50 per million output tokens. That puts it well below 3.6 Flash for cases where you need volume over depth.
The capability story is stronger than the name might suggest. On SWE-Bench Pro, Flash-Lite scores 54.2%, outperforming 3 Flash’s 49.6%. On OSWorld-Verified it reaches 74.0% against 65.1% for 3 Flash. Terminal-Bench 2.1 shows 54% versus 31%, and long-context GDM-MRCR v2 hits 72.2% versus 60.1%. For workloads that previously ran on 2.5 or 3 Flash, Flash-Lite is worth testing before assuming you need the heavier model.
Google confirmed on 21 July 2026 that Flash-Lite is rolling out inside Google Search and the Gemini app, so the performance numbers are being validated at consumer scale in parallel with the API launch.
What does this mean for you?
If you are running production workloads on 3.5 Flash, the case for migrating to 3.6 Flash is straightforward. You pay a lower output rate and, on typical tasks, generate fewer tokens to get the same result. The knowledge cutoff improvement also removes a common friction point for applications that need reasonably current information without live search integration.
If your workloads are high-volume and latency-sensitive, think document classification, search re-ranking, or high-frequency agentic loops, Flash-Lite deserves a serious look. The benchmark scores suggest it is not simply a stripped-down version of the fuller model; on several evals it outpaces the previous generation’s mid-tier option.
For developers building computer-use agents, the OSWorld-Verified score of 83% and the new built-in API tool are both meaningful steps forward. Earlier, wiring up computer use required more custom scaffolding; integrating it as a first-class API tool lowers that barrier.
Both models are available from 21 July 2026 in Google AI Studio, the Gemini API, Vertex AI, Gemini Enterprise, Android Studio, and the Gemini app. Model IDs are gemini-3.6-flash and gemini-3.5-flash-lite. Note that these are separate tiers, not a typo, read the model ID carefully when setting up integrations. Batch, Flex, and Priority inference tiers are all available.
What comes next
Google also noted that Gemini 3.5 Pro is in testing with select partners and will be made broadly available once it clears the company’s internal quality bar. The company has started what it describes as its most ambitious pre-training run to date for Gemini 4. Neither has a public release date.
For now, 3.6 Flash and 3.5 Flash-Lite cover the practical middle ground well, and the pricing and efficiency improvements give developers real reasons to move rather than stay put.