AI Progress Tracker

    Gemini 3 Pro · Gemini 3 Pro Preview (last recorded; current status unverified)

    Google · 18 Nov 2025

    Ask the Advisor about this model
    62from 9
    benchmarks

    Gemini 3 Pro launched in preview as Google's multimodal reasoning model for complex questions, software development and tool-assisted workflows. Current availability, lifecycle status and pricing: Not yet confirmed by public reporting.

    context: —textimageaudiovideo

    Benchmark results

    Every figure links to where it was published. Compare like with like — harnesses and prompt setups differ between vendors.

    BenchmarkScoreSource
    ARC-AGI-2Novel abstract reasoning puzzles31.1%November 2025; vendor launch evaluationView
    GPQA DiamondPhD-level science questions91.9%November 2025; vendor launch evaluationView
    Humanity's Last ExamFrontier expert questions across fields37.5%November 2025; vendor launch evaluation,View
    SWE-bench VerifiedResolving real GitHub issues76.2%November 2025; vendor launch evaluationView
    Terminal-BenchLong-horizon command-line agent tasks54.2%November 2025; vendor launch evaluationView
    OTIS Mock AIME 2024-25Competition mathematics91.4%19 Nov 2025View
    FrontierMath37.6%21 Nov 2025View
    LMArena TextHuman preference Elo from blind chat votes (not a percentage)1479 Elo22 Oct 2026View
    LMArena WebDevHuman preference Elo for building web apps (not a percentage)1440 Elo23 Oct 2026View

    Weighted mean of 7 verified benchmark results: ARC-AGI-2 31.1%, GPQA Diamond 91.9%, Humanity's Last Exam 37.5%, SWE-bench Verified 76.2%, Terminal-Bench 54.2%, OTIS Mock AIME 2024-25 91.4%, FrontierMath 37.6%.

    Capabilities

    • Multimodal reasoning across text, images, video and audio.
    • Analysis of large documents and other long-context inputs.
    • Advanced scientific question answering, supported by the published GPQA Diamond evaluation.
    • Code generation and repository-level software engineering with agent tooling.
    • Tool-assisted workflows involving planning and terminal interaction.

    What improved

    • Google's launch evaluation reports higher GPQA Diamond performance than Gemini 2.5 Pro.
    • Google's launch evaluation reports higher SWE-bench Verified performance than Gemini 2.5 Pro.
    • Google's launch evaluation reports higher Humanity's Last Exam performance without tools than Gemini 2.5 Pro.
    • Google describes stronger multimodal reasoning than Gemini 2.5 Pro.

    Known flaws & limitations

    • The launch Humanity's Last Exam result without tools leaves most questions unsolved.
    • The published software-engineering and terminal evaluations show substantial remaining task failures.
    • Benchmark results depend on evaluation settings and agent scaffolding; they do not establish reliability in arbitrary deployments.
    • The maximum context window does not establish dependable retrieval or reasoning across every part of an input.

    Risks, safety & security issues

    • Incorrect answers or generated code can cause downstream harm when accepted without independent validation.
    • Tool-enabled deployments can amplify mistakes unless permissions, execution environments and consequential actions are constrained.
    • Sensitive long-context inputs require data minimization, access controls and explicit retention policies.
    • Malicious packages marketed to AI-tool users can compromise host systems independently of the model's safeguards.

    Good fit for

    • Assisted scientific and technical research with source checking.
    • Software development with automated tests and human code review.
    • Multimodal document and media analysis.
    • Supervised agent workflows in restricted execution environments.

    Poor fit for

    • Unsupervised high-stakes decisions requiring reliably correct answers.
    • Autonomous execution with unrestricted credentials or production-system access.
    • Workflows requiring guaranteed exhaustive understanding of very long inputs.
    Security & risk dashboardLatest vulnerabilities, advisories and real-world incidents for Gemini 3 Pro

    References

    Profile last reviewed 10/27/2026 and re-checked automatically every few hours against the vendor's published documentation, papers and public leaderboards. Scores come from the cited benchmark results above — confirm limits and pricing against the provider's own documentation before building on them.