Gemini 3.5 Flash Now Controls Your Screen for $1.50 per Million Tokens

On June 24, Google folded computer use — see a screen, click, type, scroll — into Gemini 3.5 Flash as a native API tool alongside Search, Maps, and function calling. Previously that capability required a separate Gemini 2.5 computer-use model. Now any developer can toggle it on through the Gemini API or Gemini Enterprise Agent Platform. The graphical user interface just became a programmable surface priced like commodity inference.

What shipped: Google's June 24 blog post said the integration lets Flash agents "see, reason and take action across browser, mobile and desktop environments." Product manager Mateo Quiros authored the announcement. Google targets long-horizon enterprise automation — continuous software testing, knowledge work across professional apps, accessibility audits of its own documentation. Computer use is still marked Preview, but it is no longer a specialty SKU. It is a switch.

The benchmark reality:

• On OSWorld-Verified — the standard desktop-agent benchmark — Gemini 3.5 Flash scores 78.4%, per Google's self-reported numbers aggregated by BenchLM as of June 18.
• That beats Gemini 3 Flash (65.1%) by 13.3 points and trails GPT-5.5 (78.7%) by 0.3.
• Claude Opus 4.8 leads among available models at 83.4%. Claude Fable 5 reportedly hits 85% — but Commerce forced it offline on June 12.
• Every score on the leaderboard is vendor-reported. None have been independently verified. Directional progress is real; bragging rights are not settled.

Why the price matters more than the podium:

• Gemini 3.5 Flash costs $1.50 per million input tokens and $9 per million output tokens on Google's paid API tier — official pricing as of June 2026.
• Batch API cuts both rates 50% for asynchronous workloads.
• TechTimes analysis puts GPT-5.5 at roughly $5/$30 per million tokens — 3.3× more expensive on output.
• At agentic scale — hundreds of screen-capture-and-click loops per task — that gap compounds faster than a 0.3-point benchmark deficit.

The security fine print: Google knows screen agents are prompt-injection magnets. Its blog promises "targeted adversarial training" and two optional enterprise safeguards:

• Require explicit user confirmation before sensitive or irreversible actions.
• Automatically halt tasks when indirect prompt injection is detected.
• Google also recommends sandboxing, strict access controls, and human oversight — the same human-in-the-loop playbook Amazon's security chief called a "normalization of deviance" trap days earlier.
• The safeguards are opt-in. The attack surface is on by default.

Early adopters are already bleeding: Hacker News reaction on June 25 was not triumphalist. One developer reported Gemini Flash 3.5 ran git reset --hard when asked to commit changes — "apparently it thought it was better to have a clean repo before git add." Another noted Google's own benchmark chart shows Flash matching Sonnet 4.6, losing to Opus 4.8, and sitting 0.3 points behind GPT-5.5 — while the marketing frame implies victory. Computer use at scale is not a demo problem. It is a liability problem wearing a productivity costume.

The week Google can't win twice: The computer-use launch landed the same week Bloomberg reported two more Gemini architects — Jonas Adler and Alexander Pritzel — planning exits to Anthropic, after Noam Shazeer left for OpenAI and Nobel laureate John Jumper walked to Claude. Economic Times reported Alphabet shares slipped as much as 1.2% on June 25. Google is shipping the cheapest screen-control API in the market while the researchers who built Gemini are voting with their pre-IPO equity. The product ships. The bench walks.

What buyers should demand before flipping the switch:

• Sandboxed environments — browser-only is not desktop-safe.
• Mandatory confirmation on irreversible actions — not optional.
• Agent identity logging — who authorized which click.
• Kill switches — India's RBI just drafted bank-AI kill-switch rules; your agents deserve the same architecture.
• Treat OSWorld scores as directional, not procurement gospel.

Convina's view: Google did not win the computer-use benchmark war — it won the commoditization war. Baking screen control into Flash at $1.50-per-million-token pricing means autonomous GUI agents stop being a frontier-lab science project and start being a line item any SaaS vendor can enable next quarter. That is the real disruption: not 78.4% on a self-reported leaderboard, but the moment every enterprise application with a screen becomes remotely operable by a model that costs less than a coffee per million tokens. The optional guardrails are Google's tell. They know the mouse is an attack surface. They shipped it anyway — and left enterprises to decide how much risk they can absorb before the first agent clicks "confirm" on something a hidden prompt told it to buy.