Gemini 3.5 Flash arrives with agentic intelligence

Google announced Gemini 3.5 today — a new family of models headed by the 3.5 Flash variant, with 3.5 Pro following in June. The framing is "frontier intelligence with action": models built to plan, build, and iterate on long-horizon tasks rather than just answer in a single turn.

What's new

3.5 Flash is positioned as a frontier-tier model focused on agentic workflows — multi-step coding, debugging, and audit tasks — at meaningfully higher throughput than other models in the same tier. Google reports 4× faster output tokens/second compared to peer frontier models, while keeping accuracy on long-horizon benchmarks.

Reported benchmarks

Google's numbers for 3.5 Flash:

• Terminal-Bench 2.1 — 76.2%
• GDPval-AA — 1656 Elo
• MCP Atlas — 83.6%
• CharXiv Reasoning (multimodal) — 84.2%

The terminal-bench and MCP scores in particular signal the model's positioning around long-running developer agents — the kind of tasks where a model has to use tools, recover from errors, and stay coherent across many turns.

Availability

Shipping today:

• The Gemini app and AI Mode in Google Search
• Google Antigravity — the agentic development platform
• The Gemini API via Google AI Studio and Android Studio
• Gemini Enterprise and the Gemini Enterprise Agent Platform

Code paths through the unified google-genai SDK don't change; existing callers just pass model="gemini-3.5-flash".

What's next

The blog post positions the 3.5 series as "built with frontier safeguards" — Google's framing for capability and safety scaling in lockstep. 3.5 Pro lands in June 2026; pricing and rate-limit details weren't shared with the Flash launch.

---

Source: Gemini 3.5: frontier intelligence with action, Google's Keyword blog · May 19, 2026 · authored by Koray Kavukcuoglu, Jeff Dean, Oriol Vinyals, and Noam Shazeer.