Google has made Gemini 3.6 Flash and Gemini 3.5 Flash-Lite generally available, giving developers two new production paths for speed-sensitive AI workloads.
The release is not a simple "new smartest model" story. It is a cost-and-latency story. Google is splitting Flash into two clearer lanes: 3.6 Flash for more capable agentic, coding, multimodal, and knowledge-work tasks, and 3.5 Flash-Lite for high-volume automation where throughput and price matter most.
The timing also matters. Google's July 21 release notes mark both models as stable, production-ready releases in the Gemini API. Google says the models are available through Google AI Studio, Android Studio, Gemini Enterprise Agent Platform, the Gemini app, and, for Flash-Lite, Google Search rollout.
Official demo video
Google published several official demo clips with the launch. The most useful one for builders is the side-by-side showing Gemini 3.6 Flash using fewer tokens than Gemini 3.5 Flash on a multi-step OSWorld-style task.
Google also published a Flash-Lite latency comparison, showing Gemini 3.5 Flash-Lite against Gemini 3.5 Flash for high-volume tasks.
The benchmark snapshot
Google's official numbers position 3.6 Flash as a quality-and-efficiency upgrade over 3.5 Flash, not just a faster label. The headline claim is that 3.6 Flash uses 17% fewer output tokens than 3.5 Flash on the Artificial Analysis Index, with Google also pointing to larger token reductions on some workloads such as DeepSWE.
| Benchmark or metric | Gemini 3.6 Flash | Comparison model | Why it matters |
|---|---|---|---|
| Output token usage, Artificial Analysis Index | 17% fewer than 3.5 Flash | 3.5 Flash baseline | Lower verbosity can reduce agent cost per task. |
| DeepSWE | 49% | 37% on 3.5 Flash | Coding agents need fewer unwanted edits and debug loops. |
| MLE Bench | 63.9% | 49.7% on 3.5 Flash | Better fit for ML research and technical workflows. |
| OSWorld-Verified | 83.0% | 78.4% on 3.5 Flash | Computer-use and UI automation are central to agent products. |
| GDPval-AA v2 | 1421 | 1349 on 3.5 Flash | Google is pitching stronger knowledge-work performance. |
For Flash-Lite, the pitch is different: high throughput, lower latency, and a better price-to-performance ratio than prior Lite models. Google says 3.5 Flash-Lite reaches 350 output tokens per second according to Artificial Analysis.
| Benchmark or metric | Gemini 3.5 Flash-Lite | Comparison model | Why it matters |
|---|---|---|---|
| Output speed, Artificial Analysis | 350 output tokens/s | n/a | The model is built for latency-sensitive production traffic. |
| Terminal-Bench 2.1 | 54% | 31% on 3.1 Flash-Lite | Lite is no longer only for simple classification and extraction. |
| GDM-MRCR v2 | 72.2% | 60.1% on 3.1 Flash-Lite | Better long-context handling makes it more useful for document workflows. |
| GDPval-AA v2 | 1140 | 642 on 3.1 Flash-Lite | A big jump for business and knowledge-work tasks. |
| SWE-Bench Pro | 54.2% | 49.6% on Gemini 3 Flash | Google says Lite beats older Flash on some agentic coding evals. |
| OSWorld-Verified | 74.0% | 65.1% on Gemini 3 Flash | A faster model can still be viable for computer-use tasks. |
As always, benchmark comparisons need context. These are Google-selected results from the launch materials, not a full independent evaluation suite. The practical takeaway is narrower: Google is trying to make Flash cheaper to run in agent loops, while making Lite strong enough for subagents and background automation.
Pricing and model IDs
For developers, the model names are straightforward:
- Gemini 3.6 Flash:
gemini-3.6-flash - Gemini 3.5 Flash-Lite:
gemini-3.5-flash-lite
Google lists 3.6 Flash at $1.50 per 1 million input tokens and $7.50 per 1 million output tokens. That keeps input pricing aligned with 3.5 Flash while lowering the output-token cost from the previous Flash price point.
Gemini 3.5 Flash-Lite is much cheaper: $0.30 per 1 million input tokens and $2.50 per 1 million output tokens. That makes it the obvious candidate for routing, extraction, document processing, search agents, classification, translation, and any workflow where many smaller calls are more important than one maximal answer.
What changed for developers
The migration notes are worth reading before swapping model IDs in production. Google says new Gemini releases starting with 3.6 Flash and 3.5 Flash-Lite deprecate the classic sampling parameters temperature, top_p, and top_k. The current API ignores them, and future models may return errors if they are supplied.
Google also warns against prefilled model turns. If a request ends with a non-empty model role turn, the API can return a 400 error. The preferred path is to move output constraints into system instructions or structured-output settings.
The default thinking levels are different too. Gemini 3.6 Flash defaults to medium thinking. Gemini 3.5 Flash-Lite defaults to minimal thinking, which is sensible for throughput but not always enough for autonomous tool use. For subagents that need code execution, tool calls, or multi-step reasoning, Google recommends raising Flash-Lite to medium or high.
Why this matters
The release says a lot about where model competition is heading. Google is not only chasing a flagship score. It is trying to make its high-volume model layer more economical.
That matters for agent products because the cost of an agent is usually not one prompt. It is many model calls: plan, search, inspect, call tools, revise, check, and summarize. If 3.6 Flash really completes those loops with fewer tokens and fewer tool calls, it can lower the cost per finished task even when the model looks only modestly cheaper on the pricing page.
Flash-Lite is the other half of that strategy. A fast, low-cost model can sit under larger agents as a worker layer for extraction, routing, cleanup, translation, data normalization, and simple UI automation. If the Lite model is reliable enough, developers can reserve heavier models for planning and synthesis while sending routine subtasks to the cheaper lane.
Our take
Gemini 3.6 Flash looks like Google's new practical default for builders who care about coding, document work, multimodal reasoning, and agents but still need Flash economics. Gemini 3.5 Flash-Lite looks like the volume model: less glamorous, but potentially more important for real products that run thousands or millions of calls.
The caveat is independent validation. The benchmark deltas are promising, especially on token efficiency and agentic coding tasks, but teams should test on their own traces before migrating production traffic.
For now, this is a meaningful release because it pushes the Gemini stack toward a clearer model hierarchy: 3.6 Flash for capable workhorse agents, 3.5 Flash-Lite for cheap worker calls, and 3.5 Pro still waiting in the wings.