AI Progress Tracker

    Claude Sonnet 4.5 · claude-sonnet-4-5-20250929

    Anthropic · 29 Sep 2025

    Ask the Advisor about this model
    67from 10
    benchmarks

    Claude Sonnet 4.5 is an Anthropic model designed for coding, reasoning, computer use and multi-step agent workflows. Its published launch results demonstrate improvements over Claude Sonnet 4, but factual reliability, prompt injection and the consequences of excessive tool permissions remain important limitations.

    context: 200Ktextimagefile$3.00 / $15 per 1M tokens (in / out)

    Benchmark results

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

    BenchmarkScoreSource
    AIMECompetition mathematics87%29 Sep 2025View
    GPQA DiamondPhD-level science questions83.4%29 Sep 2025View
    SWE-bench VerifiedResolving real GitHub issues77.2%29 Sep 2025View
    Terminal-BenchLong-horizon command-line agent tasks50%29 Sep 2025View
    OTIS Mock AIME 2024-25Competition mathematics77.8%28 Oct 2025View
    FrontierMath15.2%16 Nov 2025View
    SimpleQA VerifiedShort factual questions (hallucination resistance)30.7%27 Aug 2026View
    MATH Level 597.7%21 Oct 2025View
    LMArena TextHuman preference Elo from blind chat votes (not a percentage)1438 Elo22 Oct 2026View
    LMArena WebDevHuman preference Elo for building web apps (not a percentage)1385 Elo23 Oct 2026View

    Weighted mean of 8 verified benchmark results: AIME 87%, GPQA Diamond 83.4%, SWE-bench Verified 77.2%, Terminal-Bench 50%, OTIS Mock AIME 2024-25 77.8%, FrontierMath 15.2%, SimpleQA Verified 30.7%, MATH Level 5 97.7%.

    Capabilities

    • Repository-level software development, debugging and code modification.
    • Multi-step agent workflows using external tools.
    • Computer interaction through screenshot interpretation and interface actions.
    • Mathematical and scientific reasoning with extended thinking.
    • Analysis and synthesis of long documents and codebases.

    What improved

    • Higher SWE-bench Verified performance than Claude Sonnet 4 in Anthropic's evaluations.
    • Improved computer-use performance over Claude Sonnet 4.
    • Better sustained execution of long-running coding and agent tasks.
    • Reduced sycophancy and improved prompt-injection resistance in Anthropic's evaluations.

    Known flaws & limitations

    • Can generate incorrect facts, faulty reasoning and defective code, including plausible but unsupported conclusions.
    • Computer-use and tool workflows can fail or require human intervention.
    • Improved prompt-injection resistance does not eliminate susceptibility to malicious external instructions.
    • The 1M-token context window is an extended-context option rather than the standard window.
    • The listed prices are standard USD rates per million input and output tokens; long-context requests and additional services can cost more.

    Risks, safety & security issues

    • Instructions hidden in websites, documents or tool results can redirect an agent or induce disclosure of sensitive information.
    • Coding and tool-use capabilities can be repurposed for reconnaissance, exploitation and data theft.
    • Broad filesystem, network or account permissions can turn model mistakes into unauthorized changes or data exposure.
    • Confident but incorrect reports can mislead operators about whether an action succeeded or what information was obtained.

    Good fit for

    • Supervised software engineering with automated tests and code review.
    • Tool-assisted research and document analysis with source checking.
    • Multi-step business workflows with scoped permissions and approval gates.
    • Computer-use automation in controlled, recoverable environments.

    Poor fit for

    • Unsupervised high-stakes decisions requiring consistently correct factual judgments.
    • Autonomous execution with unrestricted production credentials or destructive permissions.
    • Security-sensitive workflows that treat external content as trusted instructions.
    Security & risk dashboardLatest vulnerabilities, advisories and real-world incidents for Claude Sonnet 4.5

    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.