Gemini 3.5 Flash is here and Google’s boldest AI claim yet is that it’s not just smarter, it’s four times faster than competing frontier models, at less than half the cost. Unveiled at Google I/O 2026, the model outpaces its own Pro-tier predecessor on coding and agentic benchmarks, redefining what “efficient AI” really means.
Google opened its annual developer conference on Monday with a declaration that its AI has stopped being a chatbot and started being a builder. The centrepiece announcement was Gemini 3.5 Flash the first model in the new 3.5 family which the company describes as its strongest release yet for coding and autonomous agent work.
The headline claim is striking: Flash delivers 289 tokens per second, running at four times the output speed of comparable frontier models, while costing less than half as much per task. An optimised deployment pushes that speed advantage to twelve times, Google says, with equivalent output quality.
“Frontier-level performance at 4× the speed of comparable frontier models often at less than half the cost.”
Google, official launch statement, May 19 2026
A Flash model that beat its own Pro
Perhaps the most pointed detail in Google’s announcement is what Gemini 3.5 Flash was measured against: its own previous flagship. According to DeepMind chief technologist Koray Kavukcuoglu, Flash “outperforms our latest frontier model, 3.1 Pro, on nearly all benchmarks” including coding, agentic tasks, and multimodal reasoning. Gemini 3.1 Pro had only launched in February 2026.
The inversion a speed-optimised, lower-cost model surpassing the premium tier is deliberate. Google argues that for the workloads that actually matter in production today (long agent loops, terminal automation, multi-file coding, parallel subagent coordination), raw intelligence is only useful if it arrives quickly enough to keep pace.
| Benchmark | What it measures | Score |
| Terminal-Bench 2.1 | Coding performance | 76.2% |
| GDPval-AA | Real-world agentic tasks | 1656 Elo |
| MCP Atlas | Scaled tool-use reliability | 83.6% |
| CharXiv Reasoning | Multimodal understanding | 84.2% |
Built for the age of agents
Speed is not just a vanity metric at this scale. When agentic systems run in production, multiple agents operate simultaneously on long-running tasks, and latency compounds with every parallel process. Google co-developed Gemini 3.5 Flash alongside Antigravity its new agent-first IDE specifically to optimise for that environment, where a single user query might spawn dozens of subagents working in parallel.
At I/O, Google demonstrated agents building a full operating system from scratch not a scripted demo, but a live run of software agents dividing work, spawning sub-processes, and stitching components together. Flash was the model powering the feat. It’s the clearest signal yet that the AI race has moved from benchmark leaderboards to real-world deployment stamina.
This matters especially for developers building with the Gemini API latency at this scale is the difference between a product that feels instant and one that makes users wait.
Availability and what comes next
Status Default Model Powers 3.5 Pro
Generally Available Gemini App Gemini Spark Coming June 2026
Via Gemini API + AI Mode in Search Personal AI Agent Internal testing
Gemini 3.5 Flash is generally available today via the Gemini API, and is now the default model across the Gemini app and AI Mode in Google Search. It also powers Gemini Spark, the company’s new personal AI agent. Meanwhile, Gemini 3.5 Pro the premium sibling remains in internal testing ahead of a planned June 2026 launch.
For a deeper look at how Flash compares to other leading models released this year, see our 2026 frontier model speed benchmark roundup. The gap between what Flash delivers and what rivals charge for similar output is, at this moment, difficult to ignore.
Google I/O 2026, Gemini 3.5 Flash, Agentic AI, Coding models, AI speed, DeepMind
Antigravity IDE
