模型與開發者平台
Gemini 3.6 Flash Reduces Agent Output Tokens, While 3.5 Flash-Lite Handles High-Volume Subtasks at 350 Tokens/s
Google is moving Gemini 3.6 Flash and 3.5 Flash-Lite to stable APIs, targeting complex primary agent workloads and high-throughput subagents, respectively. Official data shows improvements in token usage, coding tasks, and computer use, but most comparisons come from Google or its evaluation partners and still require validation in real-world workflows.
Google has released stable model IDs for Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, both of which are now available through the Gemini API. 3.6 Flash is priced at $1.50 per million input tokens and $7.50 per million output tokens. Google says it generates 17% fewer output tokens than 3.5 Flash on the Artificial Analysis Index and completes multistep tasks with fewer reasoning steps and tool calls. According to the company’s published results, its DeepSWE score increased from 37% to 49%, MLE-Bench from 49.7% to 63.9%, and OSWorld-Verified from 78.4% to 83.0%. For agent services, eliminating a single unproductive tool loop can reduce model costs, end-to-end latency, and external API expenses simultaneously—potentially delivering more value than simply increasing tokens per second.
Gemini 3.5 Flash-Lite, meanwhile, is positioned for large-scale automation and subagent workloads. Third-party evaluator Artificial Analysis measured throughput of approximately 350 output tokens per second. API pricing is $0.30 per million input tokens and $2.50 per million output tokens. Google reports a Terminal-Bench 2.1 score of 54%, up from 31% for the previous generation. The model also supports multiple thinking levels, allowing systems to constrain the reasoning budget for classification, search, or document-processing tasks and increase it for multistep work. Both models integrate computer-use tools, making it easier for a primary agent to delegate large volumes of short tasks to a less expensive model.
Developers still cannot directly translate vendor benchmarks into production reliability. DeepSWE, OSWorld, and throughput measurements differ in prompts, tool environments, concurrency, and failure-retry configurations. Lower output-token usage may also result from shorter responses rather than genuinely improved planning. The Gemini API is also marking `temperature`, `top_p`, and `top_k` as deprecated, so migrations should involve more than replacing the model name. Teams should retest structured outputs, tool selection, long-context performance, and the tail distribution of costs, while pinning stable model IDs to prevent behavior drift when the `latest` alias is updated.