On June 24, 2026, Google officially deeply integrated and enabled the "Computer Use" (computer operation) tool in the Gemini 3.5 Flash model, markedly enhancing its agentic capabilities. This update moves the AI from a "question‑and‑answer" machine toward an "action‑oriented" agent: developers can use screenshot analysis and precise action outputs to directly control browsers, desktop applications and mobile interfaces, autonomously completing complex cross‑software workflows. Meanwhile, the Gemini 3.5 Pro version, originally planned for June, has been postponed again to July, underscoring Google's cautious approach to quality control and release cadence.
This feature enhancement is seen as a key step in the Gemini series' shift toward "frontier intelligence + real‑world action." Since Gemini 3.5 Flash was unveiled at Google I/O on May 19, it has been widely adopted as an efficient model, and the June 24 Computer Use update makes it more competitive in real agent scenarios.
Key innovations: native Computer Use integration, step‑change in agentic capability
The Computer Use enhancement in Gemini 3.5 Flash is the headline feature of this update. Previously available in preview, the capability is now built into the Flash model, with substantial gains in performance and usability, marking a breakthrough in GUI interaction and autonomous operation.
Main technical innovations and capability gains:
- Screenshot‑driven precise actions: The model analyzes real‑time screenshots to interpret interface elements and outputs precise action sequences (e.g., clicks, typing, scrolling, drag‑and‑drop). It supports cross‑platform operations across browsers, desktop apps and mobile interfaces, suitable for automated workflows, UI testing, data processing and similar scenarios. Early user feedback indicates strong performance in tasks such as flight searches, document audits and electrical design.
- OSWorld‑Verified benchmark leading: It achieved a 78.4% success rate on the OSWorld‑Verified Agentic computer use benchmark, approaching or matching top models (for example GPT‑5.5 at 78.7%). This result reflects strong GUI grounding and multi‑step task execution in real open environments, outperforming earlier agent tools.
- Multi‑agent parallelism and long‑horizon execution: Combined with existing agentic optimizations, the model supports parallel task handling, long‑term planning and self‑iteration. Developers can build complex agents in platforms such as Antigravity to enable end‑to‑end automation.
- Multimodal and long‑context foundation: A 1M tokens context window (about 1 million input tokens) and up to 64k output tokens, with native support for text, images, video and audio. Knowledge is current through January 2025, making it suitable for large‑document and multimedia tasks. A "thinking levels" mechanism lets users flexibly control reasoning depth, quality and speed.
- Security and enterprise‑grade protections: Built‑in safety filtering to block suspicious inputs and risky operations, with fine‑grained permission controls and audit logging suitable for production deployments.
These innovations make Gemini 3.5 Flash especially well suited for software development automation, office workflow optimization, test engineering and industry agent platforms, significantly lowering the barrier to building action‑capable AI.
API pricing: cost‑effective for high‑frequency agent scenarios
Gemini 3.5 Flash's pricing continues the Flash series' efficiency positioning; with Computer Use enhancements it remains competitively priced (official 2026 pricing):
- Input price: $1.50 / million tokens (paid tier)
- Output price (including thinking tokens): $9.00 / million tokens
- Context caching: $0.15 / million tokens/hour (storage price, substantially cheaper than direct re‑input)
- Grounding with Google Search: first 5,000 queries per month free, then metered
- Free tier: limited quotas available in the Gemini App, AI Studio and other tools to support developer testing.
Compared to prior Flash models, 3.5 Flash is somewhat more expensive, but performance improvements (especially on agentic and coding tasks) yield better overall "intelligence per dollar." It is more cost‑effective than Gemini 3.1 Pro (roughly $2/$12 entry), making it attractive for high‑concurrency, long‑context or frequent agent calls. Enterprise customers can access regional pricing and volume discounts via Vertex AI / Gemini Enterprise Agent Platform.
Developer feedback indicates complex agent tasks increase token consumption, but combined with caching and high‑speed outputs (hundreds of tokens/s), overall economics remain strong.
Pro edition delay: background and strategic significance
Gemini 3.5 Pro, originally scheduled for a June release, has been postponed to July. The Pro build is already in internal testing and limited enterprise preview; the extra time is being used to gather feedback and optimize complex reasoning and long‑context performance. Sundar Pichai had previously said at I/O that delivery would come "next month," but the timeline has been adjusted to July.
The delay reflects Google's strict quality standards for flagship models. The Pro edition is expected to bring stronger reasoning, larger context (rumored 2M tokens) and a "Deep Think" mode, further rounding out the 3.5 family. The Flash edition's feature enhancements mean users can access core agentic capabilities without waiting.
▶️ Enjoy this article? Creating content takes real effort!
Subscribe to our YouTube channel:
AI Rosetta
Your likes, subscriptions and shares are our greatest support! 🙏
For more AI news visit: https://www.airosetta.com