Google Releases Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber: A Cheaper, More Token-Efficient Flash Tier Built for Agentic Workloads

By

Developers building production agents need higher token efficiency, lower latency, and more reliable performance. Today, Google has released three new Gemini models. The lineup is Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber. All three sit in the Flash tier, which Google tunes for speed, cost, and high-volume agentic work rather than maximum reasoning depth.

Gemini 3.6 Flash: better quality, fewer tokens, lower price

Gemini 3.6 Flash is the new default workhorse. It builds on 3.5 Flash and targets coding, knowledge work, and multimodal tasks. The main point is efficiency. On the Artificial Analysis Index, 3.6 Flash uses 17% fewer output tokens than 3.5 Flash. On the DeepSWE benchmark by Datacurve, Google reports up to a 65% reduction. The model also takes fewer reasoning steps and tool calls per multi-step workflow.

Pricing moves down alongside efficiency. Gemini 3.6 Flash is priced at $1.50 per 1M input tokens and $7.50 per 1M output tokens. The output rate drops from the previous $9.00 on 3.5 Flash. Lower verbosity and a lower output price reduces the total cost per agentic task.

Quality gains accompany the efficiency gains. On DeepSWE, 3.6 Flash scores 49% versus 37% for 3.5 Flash. On MLE Bench, it reaches 63.9% versus 49.7%. On OSWorld-Verified, it hits 83.0% versus 78.4%. On GDPval-AA v2, a knowledge-work benchmark, it scores 1421 versus 1349. Computer use is now a built-in client-side tool through the Gemini API and Gemini Enterprise. Early customers including Hebbia and Harvey cite gains in document parsing, chart and data analysis, and report drafting.

Google is shipping 3.6 Flash with enhanced Frontier Safety safeguards. These cover Chemical, Biological, Radiological, and Nuclear (CBRN) and cyber-offense misuse. Full details are in the 3.6 Flash model card.

The interactive explainer below lets you compare each model against its predecessor and estimate token cost at your own volume.

Gemini 3.5 Flash-Lite: the fastest model in the 3.5 line

Gemini 3.5 Flash-Lite is highlighted for low-latency and high-throughput jobs. Target use cases include agentic search and document processing. As measured by Artificial Analysis, it runs at 350 output tokens per second. Pricing is $0.30 per 1M input tokens and $2.50 per 1M output tokens.

The model clears the prior 3.1 Flash-Lite by wide margins. On Terminal-Bench 2.1, it scores 54% versus 31%. On GDM-MRCR v2, a long-context benchmark, it reaches 72.2% versus 60.1%. On GDPval-AA v2, it scores 1140 versus 642. Notably, Flash-Lite also beats the older 3 Flash on some evals. It leads on SWE-Bench Pro at 54.2% versus 49.6% and on OSWorld-Verified at 74.0% versus 65.1%.

Flash-Lite exposes configurable thinking levels: minimal, low, and higher. Developers can prioritize low-cost, low-latency execution for high-volume tasks. They can also engage higher thinking levels for multi-step subagent workloads. Computer use is a built-in tool here too.

Gemini 3.5 Flash Cyber in CodeMender: cheap agents that find and patch bugs

Gemini 3.5 Flash Cyber is the most specialized release. It is built on 3.5 Flash and fine-tuned to find, validate, and patch software vulnerabilities. The design premise is the search-space problem. Finding deep flaws means exploring an immense execution search space. A single call to one massive model becomes a bottleneck.

The answer is a cheap model called many times. Inside CodeMender, Google’s code-security agent, multiple 3.5 Flash Cyber agents run in parallel. CodeMender invokes the model up to five times, then merges the sub-agent findings into one report. On the CyberGym benchmark, this setup reaches competitive performance against much larger models.

The internal evaluations are striking. On Google’s Big Sleep evaluation, Flash Cyber significantly surpassed mainline 3.5 Flash and 3.6 Flash. On the V8 JavaScript engine, it found 55 unique confirmed issues at a fixed number of invocations. That compares to 47 for mainline 3.5 Flash and 36 for Claude Opus 4.6. It caught 10 issues the other two models missed. In one real-world test, Google’s Cloud Vulnerability Research team used it to find remote-code-execution flaws in public APIs within two hours.

https://ift.tt/RhYonqP

Community Reaction

Reaction split along predictable lines. Builders welcomed the price and efficiency. The delayed flagship drew the loudest criticism. On Hacker News, some argued Google is over-selling capacity it cannot reliably provision, citing frustrating hands-on coding sessions. The gated Flash Cyber release opened a dual-use debate about who should hold automated exploit-finding tools.

The dashboard below aggregates that discussion by platform. It is a qualitative editorial synthesis, not a scraped dataset, and the method note is embedded.

Availability

Gemini 3.6 Flash and 3.5 Flash-Lite are available starting today. Developers can access them through the Gemini API via Google AI Studio and Android Studio. Gemini 3.6 Flash is also in Google Antigravity and rolling out in GitHub Copilot. Enterprises get both models in the Gemini Enterprise Agent Platform, with 3.6 Flash in the Gemini Enterprise app. Everyone can use them via the Gemini app, and 3.5 Flash-Lite is rolling out in Google Search. Start with the Developer Guide.

Key Takeaways

  • Gemini 3.6 Flash cuts output tokens by 17% (up to 65% on DeepSWE) and drops the output price from $9.00 to $7.50 per 1M.
  • Gemini 3.5 Flash-Lite runs at 350 tokens/sec for $0.30/$2.50 per 1M and beats the older 3 Flash on SWE-Bench Pro and OSWorld-Verified.
  • Gemini 3.5 Flash Cyber powers CodeMender with cheap multi-agent scans; it found 55 unique V8 issues versus 47 and 36 for 3.5 Flash and Opus 4.6.
  • Flash Cyber is gated to governments and trusted partners under a limited-access pilot due to dual-use risk.

The post Google Releases Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber: A Cheaper, More Token-Efficient Flash Tier Built for Agentic Workloads appeared first on MarkTechPost.



from MarkTechPost https://ift.tt/1MV0asm
via IFTTT

Comments

Post a Comment

Popular posts from this blog

Microsoft AI Proposes BitNet Distillation (BitDistill): A Lightweight Pipeline that Delivers up to 10x Memory Savings and about 2.65x CPU Speedup

By
Image
Microsoft Research proposes BitNet Distillation , a pipeline that converts existing full precision LLMs into 1.58 bit BitNet students for specific tasks, while keeping accuracy close to the FP16 teacher and improving CPU efficiency. The method combines SubLN based architectural refinement , continued pre training , and dual signal distillation from logits and multi head attention relations. Reported results show up to 10× memory savings and about 2.65× faster CPU inference , with task metrics comparable to FP16 across multiple sizes. What BitNet Distillation changes? The community already showed that BitNet b1.58 can match full precision quality when trained from scratch, but converting a pretrained FP16 model directly to 1.58 bit often loses accuracy, and the gap grows as model size increases. BitNet Distillation targets this conversion problem for practical downstream deployment. It is designed to preserve accuracy while delivering CPU friendly ternary weights with INT8 activa...
Read more

Technical Deep Dive: Automating LLM Agent Mastery for Any MCP Server with MCP- RL and ART

By
Image
Table of contents Introduction What Is MCP- RL? ART: The Agent Reinforcement Trainer Code Walkthrough: Specializing LLMs with MCP- RL Explanation: Under the Hood: How MCP- RL Generalizes Real-World Impact and Benchmarks Architectural Overview Practical Integration Summary Introduction Empowering large language models (LLMs) to fluidly interact with dynamic, real-world environments is a new frontier for AI engineering. The Model Context Protocol (MCP) specification offers a standardized gateway through which LLMs can interface with arbitrary external systems—APIs, file systems, databases, applications, or tools—without needing custom glue code or brittle prompt hacks each time. Still, leveraging such toolsets programmatically, with robust reasoning across multi-step tasks, remains a formidable challenge. This is where the recent combination of MCP- RL (a reinforcement learning loop targeting MCP servers) and the open-source ART (Agent Reinforcement Trainer) library ...
Read more

Google AI Releases LangExtract: An Open Source Python Library that Extracts Structured Data from Unstructured Text Documents

By
In today’s data-driven world, valuable insights are often buried in unstructured text—be it clinical notes, lengthy legal contracts, or customer feedback threads. Extracting meaningful, traceable information from these documents is both a technical and practical challenge.  Google AI’s new open-source Python library, LangExtract , is designed to address this gap directly, using LLMs like Gemini to deliver powerful, automated extraction with traceability and transparency at its core. Key Innovations of LangExtract 1. Declarative and Traceable Extraction LangExtract lets users define custom extraction tasks using natural language instructions and high-quality “few-shot” examples. This empowers developers and analysts to  specify exactly which entities, relationships, or facts to extract, and in what structure . Crucially, every extracted piece of information is  tied directly back to its source text —enabling validation, auditing, and end-to-end traceability. 2....
Read more