Gemini 3.5 Flash Computer Use Ties GPT-5.5 at 1/3 Cost

Verdict: Gemini 3.5 Flash is now a genuine contender for agentic desktop automation — it matches GPT-5.5 on OSWorld-Verified at roughly one-third the cost, with enterprise-grade security controls built in. The standalone Gemini 2.5 Computer Use model is effectively obsolete.

On June 24, 2026, Google shipped a capability upgrade that reshapes the computer-use landscape: Computer Use is now a native, built-in tool inside Gemini 3.5 Flash — no separate model required. The feature was previously accessible only through a standalone Gemini 2.5 computer-use model launched in October 2025; that model is now superseded for this use case.

The headline number: 78.4 on OSWorld-Verified, the standard benchmark for computer-use agents across Ubuntu, Windows, and macOS. GPT-5.5 sits at 78.7 — a 0.3-point gap that independent coverage describes as effectively a three-way tie with Claude Opus 4.7 at 78.0. The more significant differentiator is not the benchmark gap but the cost: Gemini 3.5 Flash runs at $1.50/M input tokens and $9/M output, versus GPT-5.5 at $5/$30. For high-volume agentic workloads, that differential compounds fast.

But there's a catch worth stating upfront: the Computer Use tool is not yet exposed for Gemini 3.5 Flash in the public API. The published 78.4 score reflects internal evaluation; the API documentation still points to gemini-3-flash-preview for computer-use workloads. Google describes this as a public preview — GA pricing and full API availability are still coming. This is not a launch you can ship to production today; it's a signal of where the Flash line is heading.

What happened: Gemini 3.5 Flash gets native Computer Use

Before this release, combining computer use with other Gemini capabilities required multi-model orchestration: one model for screen interaction, another for Search grounding or function calling. Native integration removes that constraint. Developers invoke computer use as a single tool parameter alongside Search, Maps, and custom function calls — all within a single model context.

The technical upgrade includes an intent field added to each action response. Earlier computer-use models returned bare pixel coordinates; Gemini 3.5 Flash now outputs a brief natural-language explanation with each action — for example, "Click the search box to type the destination." For developers debugging long-horizon agent runs, this makes the reasoning chain visible at every step without separate logging infrastructure.

The generation-over-generation improvement is substantial: Gemini 3 Flash, the predecessor, scored 65.1 on OSWorld. The 13.3-point gain between model generations is the largest single jump on the current leaderboard.

Why it matters

This isn't just a benchmark flex — it's an architectural shift that changes the economics of agentic automation. A single Gemini 3.5 Flash(opens in new tab) agent can now see a screen, look up information on Search, interact with legacy software via UI, and ground responses in Maps — all without routing between models. For teams building software testing agents, office automation, or enterprise workflows, this collapses engineering complexity that previously required bespoke multi-model pipelines.

The competitive picture, with scores:

Two caveats matter: first, all OSWorld-Verified scores are self-reported by vendors — none have been independently verified as of June 2026. Second, Fable 5's 85.0 is the true leaderboard ceiling, but it's been suspended since the US export-control directive on June 12. That makes Opus 4.8 at 83.4 the effective ceiling among currently available models.

Google also shipped the security controls that make Computer Use viable for enterprise deployments:

• Adversarial training specifically for computer-use prompt injection scenarios
• Explicit user confirmation for sensitive or irreversible actions — the model asks before clicking "Delete" or "Send"
• Auto-stop on indirect prompt injection detection — if a webpage tries to inject instructions, the session terminates
• Defense-in-depth deployment guidance recommending sandboxing, human-in-the-loop verification, and strict access controls

This is the right packaging for production. Computer Use without guardrails is a security incident waiting to happen — a browser-based agent with power-user access is exactly the attack surface that makes prompt injection dangerous. Google shipped both the capability and the controls in one release, which is more than most vendors manage on a first launch.

What changes for you

If you're running Computer Use workloads on the standalone Gemini 2.5 Pro(opens in new tab) model: plan your migration. The 2.5 computer-use model is effectively deprecated for this use case. However, wait for GA pricing and full API availability before cutting over production systems — the public API does not yet expose Computer Use for Gemini 3.5 Flash.

If you're evaluating computer-use models for agentic automation: add Gemini 3.5 Flash to your benchmark list alongside GPT-5.5 and Claude Opus 4.8. The 0.3-point OSWorld gap from GPT-5.5 is negligible in practice; the 70%+ cost advantage at scale is not.

If you're building enterprise agents: the security controls Google shipped are worth studying even if you plan to use a different model. The combination of adversarial training + user-confirmation gates + auto-stop on injection detection is a template for any production computer-use deployment.

FAQ

Can I use Computer Use in Gemini 3.5 Flash today?

Not yet through the public API. Google announced the capability on June 24 as a public preview; the documented API model for computer use remains gemini-3-flash-preview. The 78.4 OSWorld-Verified score and the enterprise safeguards are real, but GA pricing and full API availability have not been announced. Watch the Gemini API documentation(opens in new tab) for the production-ready endpoint.

How does Gemini 3.5 Flash compare to GPT-5.5 for computer use?

On OSWorld-Verified, the gap is 0.3 points — 78.4 vs 78.7 — which is effectively a tie. The real difference is cost: $1.50/$9 per 1M tokens for Gemini 3.5 Flash versus $5/$30 for GPT-5.5. All OSWorld-Verified scores are self-reported by vendors and have not been independently verified as of June 2026.

What happened to the standalone Gemini 2.5 Computer Use model?

It is still accessible via its existing API endpoint, but Google has positioned Gemini 3.5 Flash as the recommended model for computer-use workloads going forward. The standalone model cannot simultaneously use Search, Maps, or function calls within the same agent context — a limitation the native Flash integration eliminates. Migration planning is the right posture; emergency migration is not (the 2.5 endpoint still works).