Google announced three new Gemini models in the Flash series. 3.6 Flash is the main model, with increased efficiency: it consumes 17% fewer output tokens compared to 3.5 Flash on the Artificial Analysis Index, with token costs dropping to $1.50/1M input and $7.50/1M output. 3.5 Flash-Lite focuses on speed (350 tokens/second) and minimal cost. 3.5 Flash Cyber is a specialized model for code security, paired with the CodeMender agent.

The choice among frontier models is shifting from abstract benchmarks to use-case specialization and verified cost per output. We saw this with the convergence on Bedrock and Opus 5 at half price: the question now is “what model solves a specific task at the right cost”. Google’s move with Flash Cyber makes this shift in perspective explicit.

The new models are Google’s response to pressure from Anthropic and OpenAI, and the reduction in verbosity (fewer reasoning steps, fewer tool calls) translates to agents that cost less per completed task. For those building agentic workflows, token savings per agentic workflow is the parameter to measure before migrating.

To understand if the token reduction is real for your work, the method remains hands-on testing: measure the cost per verified output on your tasks. The playbook for comparing two models in fifteen minutes gives you the structure to do it without getting swept away by marketing numbers.

In detail

Google’s announcement describes 3.6 Flash as an update built on feedback from those who used 3.5 Flash in production. The leap is in the efficiency of the agentic cycle: the model takes fewer reasoning steps and fewer tool calls to complete a multi-step task. On Datacurve’s DeepSWE benchmark, Google reports an output token reduction of up to 65%. This is a number to read with caution: agentic coding benchmarks are sensitive to harness setup, and DeepSWE measures a specific case. A 65% reduction on a benchmark doesn’t guarantee the same savings on your workflow, because the number of steps depends on task type, connected tools, and how the system prompt instructs the agent. The 17% on the Artificial Analysis Index is the most solid data because it comes from a third-party standardized reference, and that’s what to use as a conservative estimate when doing the math.

The pricing of 3.6 Flash ($1.50/1M input, $7.50/1M output) is lower than 3.5 Flash. The model seeks the balance between efficiency and quality, which is exactly the market where OpenAI and Anthropic are pushing: compact models that cost little per completed task, rather than maximum frontier models that burn through budget. For those building agents that run in continuous cycles, the difference in price per token multiplies quickly: an agent that takes twenty steps per task, each with 500 output tokens, goes from a cost per task of $0.10 to $0.075. Over a thousand tasks per day, that’s $25 in daily savings—numbers to verify on your actual load but showing why price per token has become an operational parameter.

3.5 Flash Cyber is the most interesting innovation from a pattern perspective. Google presents it as “a specialized model for cybersecurity paired with the CodeMender agent”. The message is that code security is solved with a model fine-tuned for that domain, orchestrated by an agent that knows how to use it. It’s the same approach Anthropic followed with Claude Code for general coding: the model is one thing, the agentic infrastructure is another. Specialization in security means the weights were trained on examples of vulnerabilities, attack patterns, and defense practices, while the CodeMender agent manages the cycle of detection, correction, and verification. For those working in development teams, this pattern suggests code security can become a continuous control integrated into the development cycle, instead of a manual step at the end of the pipeline. How effective CodeMender is on real vulnerabilities, compared to traditional SAST tools or a generic model instructed with the right prompt, remains to be verified with testing on your code.

Google also announces that Gemini 3.5 Pro is testing with partners and that pre-training of Gemini 4 has begun. The signal is that competition on top-tier frontier models continues, while the daily battle for those using AI is fought on more compact, efficient models specialized for specific tasks. Those choosing a model today look first at cost per verified output and specialization for their domain, then general quality on benchmarks.