Google releases Gemini 3.5 Flash at I/O 2026, a fast frontier model targeting agentic workflows, coding, and multimodal tasks via Antigravity, Gemini API, and AI Studio.
Key Takeaways
Outperforms Gemini 3.1 Pro on Terminal-Bench 2.1 (76.2%), GDPval-AA (1656 Elo), MCP Atlas (83.6%), and CharXiv Reasoning (84.2%).
Claims 4x output tokens per second versus other frontier models, landing top-right on Artificial Analysis index.
Priced at $1.50/$9.00 per million input/output tokens; Google positions it as under half the cost of comparable frontier models.
Antigravity harness enables collaborative subagent orchestration for long-horizon tasks like legacy codebase migration and multi-agent game development.
Enterprise partners Shopify, Macquarie Bank, Salesforce, Ramp, Xero, and Databricks are piloting agentic workflows in production.
Gemini Spark, a 24/7 personal AI agent built on 3.5 Flash, is rolling out to AI Ultra subscribers in the US next week.
Hacker News Comment Review
The pricing jump is a major friction point: 3.5 Flash at $1.50/$9.00 is a 3-5x increase over 2.5 Flash, and when accounting for higher token usage, real-world eval costs are estimated at roughly 9x Gemini 2.5 Flash and 2.4x Gemini 2.5 Pro.
Commenters running SVG and visual generation benchmarks found 3.5 Flash competitive with 3.1 Pro on creative output quality while using far fewer tokens, though individual results were mixed on spatial reasoning tasks like bicycle anatomy.
Quota exhaustion in Antigravity after just two prompts on the AI Pro plan raises concerns about production viability for agentic, multi-step use cases.
Notable Comments
@GodelNumbering: Documents the per-million pricing jump: 2.5 Flash was $0.30/$2.50; 3.5 Flash is $1.50/$9.00, similar to what 2.5 Pro cost.
@reconnecting: Flags a January 2025 knowledge cutoff against a May 2026 release, a 16-month training-to-deployment lag.
@hmate9: Reports quota exhausted in two Antigravity prompts on the AI Pro plan, calling it “seriously unusable” if not a bug.
