Gemini 3.5 Flash is just... fine

하이라이트

• Gemini 3.5 Flash achieves a throughput of 278 tokens per second, exceeding the speed of GPT 5.5 and Opus 4.7.

• Artificial Analysis benchmarks indicate the actual cost of running Gemini 3.5 Flash in agentic workflows is 5.5 times higher than its predecessor, Gemini 3 Flash.

• Gemini 3.5 Flash averages 49 turns per task in agentic evaluations, reflecting a tendency to consume more tokens than other models.

• The model performs well in basic UI design tasks but lacks the coding reasoning necessary for complex full-stack applications when compared to Opus 4.7.

• Google is replacing the open-source Gemini CLI with a closed-source version written in Go, effective June 18th.

타임라인

1. 00:00 - 02:05

   Gemini 3.5 Flash Capabilities and Benchmarks

   Google claims frontier performance for the new model, highlighting multimodal inputs including video and audio. While internal data shows strong results, third-party testing suggests lower coding proficiency. Speed remains the standout feature, significantly outpacing established frontier models.

   • The model supports a one-million token context window and 64,000 output tokens.
   • Internal benchmarks show the model competing with GPT 5.5 in coding and outperforming Opus 4.7 on Terminal Bench.
   • Third-party coding indices rank the model below Kimi K2.6 and Gemini 3.1 Pro.
   • The model achieves a peak speed of 278 tokens per second.
2. 02:06 - 03:40

   Cost and Agentic Efficiency

   Marketing claims suggest cost-effectiveness, but real-world testing reveals that the model's high turn count per agentic task inflates overall expenses. It currently ranks as more expensive to run than some higher-reasoning models that deliver superior coding results.

   • The model is priced at $1.50 per million input tokens and $9 per million output tokens.
   • Actual operational costs during benchmark testing proved to be 75% higher than Gemini 3.1 Pro.
   • High token consumption stems from an average of 49 turns per task in agentic workflows.
3. 03:41 - 06:36

   Anti-Gravity 2 App and CLI Changes

   The new application interface mirrors common industry standards for AI-assisted coding tools. While it produces functional UI code, it lacks the depth required for complex project structures compared to longer-running models. The shift to a closed-source CLI signals a change in the developer experience strategy for Google.

   • Anti-Gravity 2 functions as a standalone agent app similar to existing tools like Codecs and Cursor.
   • The model excels at generating simple UI elements but struggles with complex, full-stack application logic.
   • Google is discontinuing the current Gemini CLI in favor of a closed-source Go-based implementation.