Gemini 3.5 Flash Lite & Gemini 3.6 Flash
On July 21, 2026, Google added three new models to the Gemini family: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and the limited-access Gemini 3.5 Flash Cyber. This time, the announcement focused not on raw intelligence, but on token efficiency, latency, and reliability in production environments.
This focus is no coincidence. Following the introduction of Gemini 3.5 Flash at Google I/O in May, much of the feedback indicated that the primary cost of scaling agent-based workflows was not model quality, but the number of tokens consumed and steps executed. Gemini 3.6 Flash and Gemini 3.5 Flash-Lite directly address this challenge.
Gemini 3.6 Flash: Fewer Tokens, Lower Cost
Gemini 3.6 Flash is an updated model built directly on Gemini 3.5 Flash. Positioned by Google as a "workhorse," it is designed to complete the same tasks with more concise output.
Key efficiency improvements include:
According to the Artificial Analysis Index, it uses 17% fewer output tokens than Gemini 3.5 Flash. In certain scenarios, such as DataCurve's DeepSWE evaluation, the reduction reaches as much as 65%.
It completes multi-step workflows with fewer reasoning steps and tool calls, and is less prone to becoming stuck in execution loops.
Pricing has also decreased: $1.50 per 1 million input tokens and $7.50 per 1 million output tokens. With Gemini 3.5 Flash, output tokens cost $9.00 per million, while the input price was the same.
Quality comparison with Gemini 3.5 Flash:
Benchmark | 3.6 Flash | 3.5 Flash |
DeepSWE (code-editing accuracy) | 49% | 37% |
MLE Bench (ML research) | 63.9% | 49.7% |
OSWorld-Verified (computer use) | 83.0% | 78.4% |
GDPval-AA v2 (real-world knowledge work) | 1421 | 1349 |
Google's release notes also highlight a notable behavioral change: Gemini 3.6 Flash is more likely to run diagnostic code before making changes. This improves accuracy on complex tasks, although it may add unnecessary exploratory steps to simple front-end work.
There is also a candid caveat: evaluators preferred earlier models for visual layout and styling. The model is better at producing functional code; however, if visual quality is important, explicit design instructions in the prompt are recommended.
Computer Use is now available as a built-in tool through the Gemini API and Gemini Enterprise. Gemini 3.6 Flash is also the new default model for the Antigravity agent.
Gemini 3.5 Flash-Lite: Optimized for Speed and Scale
Gemini 3.5 Flash-Lite is the fastest and most cost-effective model in the Gemini 3.5 family. Its target use cases are clear: high-volume data processing, document parsing, classification, routing, and autonomous subagent execution.
350 output tokens per second, according to Artificial Analysis.
Pricing: $0.30 per 1 million input tokens and $2.50 per 1 million output tokens.
The default reasoning level is minimal, enabling maximum throughput.
Comparison with the previous-generation Gemini 3.1 Flash-Lite:
Benchmark | 3.5 Flash-Lite | 3.1 Flash-Lite |
Terminal-Bench 2.1 (coding/agent tasks) | 54% | 31% |
GDM-MRCR v2 (long context) | 72.2% | 60.1% |
GDPval-AA v2 (real-world tasks) | 1140 | 642 |
HLE (reasoning) | 18.0% | 11.0% |
CharXiv (multimodal) | 74.5% | 63.7% |
Notably, Flash-Lite outperforms the higher-tier Gemini 3 Flash in certain agent and coding benchmarks: 54.2% versus 49.6% on SWE-Bench Pro, and 74.0% versus 65.1% on OSWorld-Verified. This makes it both a faster and more capable option for workloads currently running on Gemini 2.5 Flash or Gemini 3 Flash.
Because the reasoning level is configurable, the model is not limited to a single operating profile. Minimal is recommended for high-volume inference workloads, while medium or high is recommended for subagents that require tool use and multi-step planning.
Gemini 3.5 Flash Cyber: Not Publicly Available
The third model is Gemini 3.5 Flash Cyber, a version of Gemini 3.5 Flash fine-tuned to identify and remediate security vulnerabilities. Within Google's CodeMender code-security agent, multiple Flash Cyber agents work together to produce a unified report, achieving near-frontier performance on the CyberGym benchmark.
Due to the dual-use nature of the technology, the model will not be released publicly. It will be offered through CodeMender only as part of a limited-access pilot program for governments and trusted partners.
Technical Specifications and API Changes
Both models support a 1-million-token context window and up to 64,000 output tokens. They accept text, images, video, audio, and PDF files as input, while output is text only.
Gemini 3.6 Flash | Gemini 3.5 Flash-Lite | |
Model ID | Gemini 3.6 Flash | Gemini 3.5 Flash-Lite |
Default reasoning level | medium | minimal |
Input / output price (per 1M tokens) | $1.50 / $7.50 | $0.30 / $2.50 |
Context window | 1M | 1M |
Computer Use | Available | Available |
Key migration considerations: Two breaking changes take effect with these models and will apply to all subsequent Gemini releases:
The temperature, top_p, and top_k parameters have been deprecated. They are currently ignored and will return errors in future model generations. If deterministic behavior is required, define explicit rules in the system instruction.
Model-turn prefilling is no longer supported. If the final non-empty turn in a request has the model role, the API returns a 400 error. If you use prefilling to enforce prefix text or JSON formatting, use system_instruction or Structured Outputs instead.
In addition, thinking_budget has been replaced by the thinking_level string enum, while candidate_count has been removed entirely in Gemini 3.x.
Where Are the Models Available?
Developers: The Gemini API through Google AI Studio and Android Studio. Gemini 3.6 Flash is also available in Google Antigravity.
Enterprises: Gemini Enterprise Agent Platform. Gemini 3.6 Flash is also available in the Gemini Enterprise application.
End users: The Gemini application. Gemini 3.5 Flash-Lite is also being rolled out to Google Search.
Which Model Should You Choose?
The decision ultimately depends on the nature of your workload:
Code generation, multi-step agent workflows, and multimodal and spatial reasoning -> Gemini 3.6 Flash. It is the natural migration target for users of Gemini 3.5 Flash, Gemini 3 Flash (Preview), or Gemini 3.1 Pro.
High-volume data analysis, document extraction, structured JSON parsing, and autonomous subagents -> Gemini 3.5 Flash-Lite. It is the recommended migration target for users of Gemini 3.1 Flash-Lite or Gemini 2.5 Flash.
In practice, the most efficient architecture is to use the two models together: assign orchestration to Gemini 3.6 Flash and parallel, high-volume subtasks to Gemini 3.5 Flash-Lite. One example shared in Google's announcement illustrates precisely this approach: Gemini 3.6 Flash acts as the primary agent, while Gemini 3.5 Flash-Lite generates dozens of web-design concepts on demand.
Conclusion
The common strength of these two models is that improvements in quality are accompanied by gains in efficiency. Gemini 3.6 Flash completes the same work with fewer tokens and at a lower cost, while Gemini 3.5 Flash-Lite delivers similar benefits for workloads that prioritize speed and cost. For teams building agent-based systems, the practical implication is clear: assigning each task to the right model now affects cost just as directly as it affects model quality.
