Google ships cheaper Gemini Flash models for agent workloads

Google DeepMind released three proprietary Gemini models aimed at cutting the cost and latency of AI agent workloads: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite and Gemini 3.5 Flash Cyber. The commercial point is straightforward for developers running agents at scale: Google says the new Flash models use fewer tokens, and it is pricing 3.6 Flash at $1.50 per million input tokens and $7.50 per million output tokens through its API.

Gemini 3.5 Flash-Lite is cheaper on input, at $0.30 per million input tokens and $2.50 per million output tokens. Google’s older Gemini 3.1 Flash-Lite remains less expensive at $0.25 and $1.50 per million tokens, but Artificial Analysis says the new 3.5 Flash-Lite is about twice as fast, producing 350 output tokens per second.

The pricing puts Google’s new models below some frontier offerings but above several low-cost competitors listed by providers such as Xiaomi, DeepSeek, MiniMax and Alibaba Cloud. Within Google’s own lineup, Gemini 3.6 Flash comes in below Gemini 3.5 Flash on output pricing, which is listed at $9.00 per million output tokens, and below Gemini 3.1 Pro Preview, listed at $12.00 per million output tokens for contexts up to 200,000 tokens.

Token efficiency is the pitch

Google’s claim is that the models reduce the number of tokens needed to complete multi-step work, not just the per-token price. Artificial Analysis reported that Gemini 3.6 Flash uses 17% fewer output tokens than Gemini 3.5 Flash. On DeepSWE, a long-horizon software engineering benchmark, the reported reduction reaches as much as 65%.

Google did not disclose the architectural changes behind those gains. Its materials say the model takes fewer reasoning steps and tool calls and is less verbose. That matters for agent products because token waste compounds quickly when systems call tools, write intermediate plans and revise outputs over long workflows.

Google’s model cards list a 1 million-token input context window and a 64,000-token maximum output limit for both Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. Both models have a March 2026 knowledge cutoff, according to the cards.

Benchmarks and availability

Google and Artificial Analysis reported higher scores for Gemini 3.6 Flash than Gemini 3.5 Flash across several tests. DeepSWE rose to 49% from 37%, while MLE-Bench rose to 63.9% from 49.7%. Google also reported an OSWorld-Verified score of 83.0%, up from 78.4%, and a GDPval-AA v2 score of 1421, up from 1349.

Gemini 3.6 Flash and Gemini 3.5 Flash-Lite are available through the Gemini API in Google AI Studio and Android Studio, and inside the consumer Gemini app and Google Search. Google said the models are intended for coding, document parsing, chart and data analysis, long-form drafting, agentic search and high-volume document processing.

Gemini 3.5 Flash Cyber is more restricted. Google said it is designed for cybersecurity researchers and red teamers, and will be available through CodeMender to governments and trusted partners. Google did not publish pricing for the cyber model, saying only that it is fine-tuned at a lower per-token price than larger models.

The Pro gap remains

The release does not include Gemini 3.5 Pro, the larger model developers have been waiting for after Google previously pointed to a summer release. Google technical staffer Logan Kilpatrick wrote on X that Gemini 3.5 Pro is testing with partners and will be made broadly available when it is ready.

All three new models are closed and distributed through Google-controlled access rather than open model weights. That keeps deployment tied to Google’s APIs and terms, which is routine for Gemini but material for enterprises that want self-hosting or air-gapped systems. Google also said pre-training for Gemini 4 has begun, without giving a release timeline.

This story draws on original reporting from VentureBeat.