Gemini 3.5 Flash Performance and Cost Analysis
Better Stackgo watch the original →
the gist
Gemini 3.5 Flash offers industry-leading inference speeds but fails to meet Google's cost-efficiency claims, often proving more expensive and less capable at coding than competing frontier models.
Performance and Cost Discrepancies
Gemini 3.5 Flash delivers high-speed inference, reaching 278 tokens per second, which outperforms models like Opus 4.7 and GPT 5.5. Despite Google's marketing claims of frontier performance at half the cost, third-party benchmarks from Artificial Analysis indicate the model is less efficient than advertised. On coding tasks, the model scores 45 on the Artificial Analysis coding index, trailing behind Kimi K 2.6 and Gemini 3.1 Pro. The model is notably token-hungry, averaging 49 turns per task on agentic evaluations, which drives actual operational costs to $1,552 per intelligence index run, making it 5.5 times more expensive than Gemini 3 Flash and 75% more expensive than Gemini 3.1 Pro.
Antigravity 2.0 and CLI Changes
Google released Antigravity 2.0 as a standalone agent application, moving away from its previous IDE-integrated form factor. The interface follows standard AI-assisted coding patterns, featuring a side-by-side chat and diff view. While the model demonstrates competence in simple UI generation, such as single-file HTML cafe websites, it struggles with complex full-stack applications compared to Opus 4.7, often producing generic AI-styled interfaces. Additionally, Google is deprecating the open-source Gemini CLI on June 18th, replacing it with a closed-source Antigravity CLI rewritten in Go.