Gemini Flash 3.6: Not Smarter, But Noticeably Cheaper

On July 21, Google released three models: Gemini 3.6 Flash (replacing 3.5 Flash, priced at $1.50/$7.50 per million tokens), Gemini 3.5 Flash-Lite (ultra-fast and cheap at $0.30/$2.50), and Gemini 3.5 Flash Cyber (for vulnerability hunting, government-only). The headline figure is a 17% reduction in output tokens compared to 3.5 Flash per Artificial Analysis, reaching 65% savings on the DeepSWE coding benchmark (97k tokens vs 276k). The overall intelligence index (50 points) is unchanged. On benchmarks, 3.6 Flash improved on DeepSWE (37%→49%), MLE-Bench (49.7%→63.9%), OSWorld-Verified (78.4%→83.0%), and long-context recall (~27%→54%), but lagged behind top competitors like Grok 4.5 and Claude Sonnet 5 on pure coding (SWE-Bench Pro 58.7% vs 64.7%) and knowledge tasks (GDPval-AA Elo 1421 vs 1607 for Sonnet 5). A notable regression: LlamaIndex CEO Jerry Liu found ParseBench chart reading dropped from 45% to 31%, and document scores from 69.9 to 66.8, likely a trade-off from post-training focus on code/agents. The Flash-Lite model is a value standout: 350 tok/s output, beating older Gemini 3 Flash on some tests (SWE-Bench Pro 54.2% vs 49.6%, OSWorld-Verified 74.0% vs 65.1%) and cheaper than Claude Haiku 4.5. Flash Cyber, built on 3.5 Flash, found 55 unique issues on Chrome V8 (vs 47 for base, 36 for Claude Opus 4.6) and scored 83.2% on CyberGym benchmark (fourth among specialized models). Google also noted internal testing of Gemini 3.5 Pro with partners and the start of Gemini 4 pre-training.