May 19, 2026AgentsAPI

Gemini 3.5 Flash Beats Last-Gen Pro on Coding

Google dropped Gemini 3.5 Flash at I/O 2026 and said the quiet part out loud: this Flash model beats the previous Gemini 3.1 Pro on coding and agentic benchmarks. Four times the throughput of frontier competitors on output tokens per second. And cheaper.

The number doesn't matter as much as the positioning. Google is telling developers: stop buying Pro for agent workloads. Buy Flash, run more turns, fire more parallel calls, and you'll come out ahead on both latency and bill. The whole industry has been moving toward thinking models. Google is going the other direction with a model that is fast and good enough.

The kicker is buried in the announcement: Gemini 3.5 Flash was itself co-developed using Antigravity, Google's own agent platform. A frontier lab publicly admitting that its agent harness is part of the model training loop. Self-improving infrastructure is no longer a research paper, it's a product page.

Rolling out today across the Gemini app, Search, and the API.

Announcement: https://blog.google
← Previous
Karpathy Walks Into Anthropic
Next β†’
Gemini Spark Goes 24/7 on Your Behalf
← Back to all articles

Comments

Loading...
>_