Google ships Gemini 3.5 Flash and managed agents in the API
A coding/agent model plus Google-hosted sandboxed 'Antigravity' agents; older image previews shut down 25 June
Summary
Google Deepmind opened the Gemini 3.5 series with 3.5 Flash, pitched at agents and coding and said to outperform Gemini 3.1 Pro on key benchmarks. Google also moved Managed Agents and a general-purpose "Antigravity" agent to public preview: autonomous, stateful agents that run in Google-hosted, isolated Linux sandboxes, able to plan, write and run code, manage files and browse the web. File Search gained multimodal (image) search via gemini-embedding-2, and the API added event-driven webhooks to replace polling. Older image-preview models (gemini-3.1-flash-image-preview, gemini-3-pro-image-preview) shut down 25 June. The push follows Google's I/O 2026 repositioning of the Gemini app against ChatGPT and Claude.
By the numbers
- 3.5 Flash, first Gemini 3.5-series model, agent/coding focused.
- 25 Jun 2026, shutdown of two image-preview models.
- Public preview, Managed Agents, Antigravity agent, multimodal File Search.
- gemini-embedding-2, model behind multimodal search.
Why it matters
Google is competing on agent infrastructure, hosted, sandboxed execution, not just model quality, lowering the bar for developers to ship autonomous agents on Google's stack. Bundling agents into the API and the consumer Gemini app leverages distribution that pure-play labs lack.
What to watch
- Independent benchmarks of 3.5 Flash vs. Claude Opus 4.8 and GPT-5.5.
- Adoption and safety incidents from Antigravity's sandboxed code execution.
- A 3.5 Pro/Ultra tier to follow Flash.