Gemini 3.6 Flash Is 17% More Token-Efficient Than Its Predecessor — and Cheaper Too

Google recently confirmed that Gemini 3.6 Flash is 17% more token-efficient than its predecessor — and cheaper too. This flagship update consumes fewer output tokens than Gemini 3.5 Flash on the Artificial Analysis Index. However, the bigger story involves simultaneous improvements in speed, cost, and agentic benchmarks. Google achieved these gains without sacrificing quality in knowledge work or coding.

This combination is rare because model releases typically trade one dimension for another. Therefore, 3.6 Flash avoids that common tradeoff effectively. Fewer tokens per task means lower costs for entire agentic workflows, not just individual tokens.

How Gemini 3.6 Flash Is 17% More Token-Efficient Than Its Predecessor — and Cheaper Too

First, coding benchmarks show significant progress. The new model scores 49% on DeepSWE by Datacurve, up from 37%. This 12-point jump suggests fewer execution loops and more precise edits. Additionally, OSWorld-Verified scores rose to 83.0%, while MLE Bench reached 63.9%.

These computer use improvements matter because Google now builds this capability natively into the Gemini API. For knowledge-intensive work, GDPval-AA v2 shows a score of 1421 versus 1349 previously. Customers at Figma, Harvey, and Hebbia report strong performance in document parsing and chart analysis.

Finally, pricing sits at $1.50 per million input tokens and $7.50 per million output tokens. This rate is lower than 3.5 Flash. Combined with efficiency gains, the cost reduction per completed task is substantial.

Flash-Lite Targets High-Volume Throughput

Gemini 3.5 Flash-Lite addresses a different need entirely. It delivers 350 output tokens per second, making it the fastest model in the 3.5 series. Pricing is just $0.30 per million input tokens and $2.50 per million output tokens. Consequently, it suits high-volume pipelines where latency matters more than deep reasoning.

Notably, Flash-Lite beats older models on key benchmarks. It scores 54.2% on SWE-Bench Pro versus 49.6% for 3 Flash. A throughput-focused model outperforming previous mid-tier options signals major architectural progress. The model also supports configurable thinking levels and built-in computer use tools.

Flash Cyber Focuses on Security

Gemini 3.5 Flash Cyber is fine-tuned specifically for finding cybersecurity vulnerabilities. Within Google’s CodeMender infrastructure, it matches frontier-level security models on the CyberGym benchmark. However, Google restricts access to governments and trusted partners via a limited pilot.

Defenders need a head start before such capabilities become widely accessible. After all, a model optimized to find vulnerabilities at scale poses genuine dual-use risks.

Future Updates and Availability

Google confirmed Gemini 3.5 Pro is currently in partner testing. More notably, pre-training for Gemini 4 has begun. This ambitious run suggests the gap between Flash tier and frontier models is compressing fast.

Developers can access 3.6 Flash and 3.5 Flash-Lite today via the Gemini API and Google AI Studio. The efficiency gains are large enough that swapping models should reduce operational costs immediately. Ultimately, these updates quietly reshape what is economical to build.