Google quietly released Gemini 3.6 Flash, 3.5 Flash Light, and 3.5 Cyber, prioritizing token efficiency over raw benchmark chasing. Here's the breakdown.
What did Google actually release? #
Google released three Gemini models on the same day: Gemini 3.6 Flash, Gemini 3.5 Flash Light, and Gemini 3.5 Cyber. None of them are frontier flagship models. They sit in Google’s fast, cost-efficient tier, built for developers running agentic workflows, document processing, and specialized security tasks rather than people chasing the top spot on general reasoning leaderboards. The headline isn’t raw intelligence gains. It’s that Google is now optimizing for token efficiency alongside quality, which matters a lot if you’re paying per token in production.
TL;DR #
Gemini 3.6 Flash replaces 3.5 Flash as Google’s workhorse fast model, cutting output token usage by 17% according to the Artificial Analysis Index while improving coding and knowledge work quality.Token efficiency is now a first-class design goal alongside raw benchmark scores, because reasoning-heavy models tend to burn far more output tokens to get to an answer.Gemini 3.5 Flash Light is the speed-optimized member of the family, running at 350 output tokens per second at $0.3 per million input tokens and $2.5 per million output tokens, aimed at agentic search and high-throughput document workflows.Gemini 3.5 Cyber is a specialized cybersecurity model that performs competitively with dedicated security models from other labs on vulnerability-detection benchmarks like CyberGym.Gemini 3.6 Flash reportedly beats Gemini 3.1 Pro on several benchmarks, which says less about Flash being extraordinary and more about how fast the lower tier is catching up to last generation’s flagship.Gemini 3.5 Pro is already in partner testing, and Google says it plans wider availability once it’s ready, with no firm release date confirmed.** Gemini 4 pretraining has started**, described by Google as its most ambitious pretraining run yet, though no timeline or benchmark details have been shared.
Remy doesn't build the plumbing. It inherits it. #
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
Why release three models at once instead of one flagship? #
Frontier labs increasingly split their model lineups into tiers rather than shipping a single do-everything model. Google’s structure here mirrors what OpenAI and Anthropic already do: a workhorse tier for general tasks, a lightweight tier for high-volume or latency-sensitive work, and increasingly, specialized tiers for narrow domains like security.
Releasing three at once signals that Google is filling out its product surface rather than chasing a single benchmark win. Gemini 3.6 Flash covers coding and knowledge work. Gemini 3.5 Flash Light covers cheap, fast, high-volume agentic tasks. Gemini 3.5 Cyber covers a specific enterprise use case: finding and patching software vulnerabilities. None of them compete with a hypothetical Gemini 4 or even fully with Gemini 3.5 Pro, which is still in partner testing. They’re infrastructure moves, not headline moves.
How is Gemini 3.6 Flash different from 3.5 Flash? #
Gemini 3.6 Flash is positioned as a direct upgrade to 3.5 Flash, built from developer and customer feedback. The two changes that matter most:
Token efficiency. According to the Artificial Analysis Index, Gemini 3.6 Flash uses 17% fewer output tokens than 3.5 Flash to complete comparable tasks. On specific benchmarks like SWE-bench (via DataCurve), Google reports up to 65% lower cost per resolved task. That’s a meaningful shift in framing: instead of just asking “is it smarter,” the pitch is “is it smarter per dollar.”
Coding and multimodal performance. Google’s own examples show the model handling tasks like locating a file on Wikipedia, converting it with Python, and resizing an image, or running multi-agent code migrations with lower latency than 3.5 Flash. These aren’t reasoning showcases. They’re the kind of grunt-work automation that agentic pipelines rely on constantly.
The tradeoff is behavioral. Gemini 3.6 Flash is tuned to act quickly rather than reason through every step upfront. In practice, that means it benefits from more explicit instructions. Reasoning-heavy models will often plan out steps before acting; Flash models tend to jump straight into execution. If you’re building agents on top of it, front- the instructions rather than relying on the model to infer intermediate steps produces better results.
Notably, Gemini 3.6 Flash reportedly outperforms Gemini 3.1 Pro, last generation’s higher tier model, on multiple benchmarks. That’s not unusual in this industry. Efficiency gains at the flash tier routinely erase the gap with the previous generation’s flagship within a release cycle or two.
What is Gemini 3.5 Flash Light built for? #
Gemini 3.5 Flash Light is the speed specialist in the lineup. It’s built for scaling agentic workflows where latency and throughput matter more than depth of reasoning: agentic search, document processing, and other tasks that require many fast, “good enough” completions rather than one slow, highly reasoned answer.
The numbers Google shared: 350 output tokens per second, priced at $0.3 per million input tokens and $2.5 per million output tokens. That pricing puts it squarely in the “cheap enough to run at scale” category. Google also claims significantly better quality than the prior 3.1 Flash Light on coding and other benchmarks.
Remy is new. The platform isn't. #
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
The use case here isn’t a chatbot. It’s the layer underneath agent systems that need to spin up many lightweight workers doing narrow tasks in parallel, where model intelligence matters less than consistent speed and low cost per call.
Is Gemini 3.5 Cyber worth paying attention to? #
Gemini 3.5 Cyber is the most narrowly scoped of the three, and arguably the most interesting from a benchmarking standpoint. It’s built specifically for cybersecurity work: finding and fixing vulnerabilities in code.
On CyberGym, a benchmark for vulnerability detection and patching, Gemini 3.5 Cyber performs on par with specialized security models from competing labs, including GPT-5.5 Cyber and Mythos-series models. For a model built by a general-purpose lab entering a niche dominated by security-focused releases, that’s a stronger showing than expected. It suggests Google can compete in vertical, specialized domains without needing a separate architecture or training pipeline built from scratch.
Whether it’s “worth it” depends entirely on whether you have a security workflow to plug it into. It’s not a general assistant, and comparing it to Gemini 3.6 Flash or a Pro-tier model misses the point of what it’s designed to do.
What does this signal about Gemini 3.5 Pro and Gemini 4? #
Google has confirmed that Gemini 3.5 Pro is currently in testing with partners, with plans to widen availability once it’s ready, but no firm date has been given. That’s a notably vague timeline given how many competing model releases have shipped in the interim from other labs.
More significant: Google says it has begun what it describes as its most ambitious pretraining run yet, for Gemini 4. No benchmarks, no release window, and no architecture details were shared. The company only stated that early progress has been encouraging.
Reading between the lines, the three-model Flash release looks like Google keeping its developer ecosystem current and competitive while the bigger pretraining and testing work happens out of public view. Shipping efficient, cheaper, tier-appropriate models buys time and keeps the platform useful without needing to reveal where Gemini 3.5 Pro or Gemini 4 actually land on the frontier.
Frequently Asked Questions #
What is Gemini 3.6 Flash used for?
It’s designed as Google’s general workhorse fast model, handling coding, knowledge work, and multimodal tasks like image processing and code migrations, with a focus on lower token usage per task than its predecessor.
How much does Gemini 3.5 Flash Light cost?
Google prices it at $0.3 per million input tokens and $2.5 per million output tokens, with output speeds around 350 tokens per second, aimed at high-throughput agentic and document-processing tasks.
Is Gemini 3.5 Cyber a general-purpose model?
No. It’s a specialized model built for cybersecurity tasks like vulnerability detection and patching, and it’s benchmarked specifically against other security-focused models rather than general assistants.
Does Gemini 3.6 Flash outperform Gemini 3.1 Pro?
According to the reporting on Google’s benchmarks, yes, on several tasks. This reflects how quickly efficiency-tier models can catch up to the previous generation’s higher tier rather than an unusual leap in capability.
When is Gemini 3.5 Pro or Gemini 4 coming out?
Remy doesn't write the code. It manages the agents who do. #
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
Google hasn’t given a firm date. Gemini 3.5 Pro is in partner testing with plans for broader release when ready, and Gemini 4 is in an early, large-scale pretraining phase with no announced timeline.