AI Progress Tracker

    Claude Sonnet 4 · claude-sonnet-4-20250514

    Anthropic · 22 May 2025

    Ask the Advisor about this model
    64from 11
    benchmarks

    Claude Sonnet 4 is Anthropic’s hybrid-reasoning model for coding, analysis and tool-driven workflows, with standard API pricing of $3 per million input tokens and $15 per million output tokens. Claude Sonnet 4.5 subsequently shipped; this profile remains specific to the original Sonnet 4.

    context: 200Kimagetextfile$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 mathematics70.5%22 May 2025View
    GPQA DiamondPhD-level science questions75.4%22 May 2025View
    MMMUMultimodal (image + text) understanding74.4%22 May 2025View
    SWE-bench VerifiedResolving real GitHub issues72.7%22 May 2025View
    Tau-benchTool-using customer-workflow agents80.5%22 May 2025View
    Terminal-BenchLong-horizon command-line agent tasks35.5%22 May 2025View
    OTIS Mock AIME 2024-25Competition mathematics71.1%22 May 2025View
    FrontierMath4.1%4 Jul 2025View
    MATH Level 584.4%22 May 2025View
    Aider PolyglotCode editing across six programming languages56.4%24 May 2025View
    LiveCodeBenchFresh competitive-programming problems68.5%1 May 2025View

    Weighted mean of 11 verified benchmark results: AIME 70.5%, GPQA Diamond 75.4%, MMMU 74.4%, SWE-bench Verified 72.7%, Tau-bench 80.5%, Terminal-Bench 35.5%, OTIS Mock AIME 2024-25 71.1%, FrontierMath 4.1%, MATH Level 5 84.4%, Aider Polyglot 56.4%, LiveCodeBench 68.5%.

    Capabilities

    • Supports both standard responses and extended thinking for more demanding reasoning tasks.
    • Generates, edits and debugs code, including repository-level software engineering tasks.
    • Can alternate extended thinking with tool calls in supported workflows.
    • Processes text and images for document analysis and visual question answering.
    • Supports long-context analysis, including a documented 1M-token beta option for eligible users.

    What improved

    • Higher published SWE-bench Verified performance than Claude Sonnet 3.7.
    • Higher published AIME 2025 performance than Claude Sonnet 3.7.
    • Anthropic reports improved instruction following and more precise responses.
    • Anthropic reports reduced shortcut-taking and reward-hacking behavior relative to Claude Sonnet 3.7.
    • Introduced interleaved thinking and tool use for extended-reasoning workflows.

    Known flaws & limitations

    • Published GPQA Diamond and MMMU results are below Claude Sonnet 3.7’s results, so benchmark gains are not uniform.
    • The published Terminal-Bench result shows substantial remaining limitations on complex terminal workflows.
    • Tool-assisted performance varies by domain; strong coding results do not establish reliable autonomous operation.
    • The 1M-token context option has eligibility requirements and premium pricing above 200K input tokens.
    • Generated code, factual claims and tool-derived conclusions still require independent validation.

    Risks, safety & security issues

    • Anthropic has documented Claude Code misuse for data theft and extortion; this is application-level evidence, not confirmation of a particular Sonnet checkpoint.
    • Malicious actors can disguise harmful requests as legitimate security work to evade safeguards.
    • Broad filesystem, network and credential access can magnify the consequences of unsafe agent actions.
    • Untrusted documents, websites and tool outputs can carry prompt-injection instructions into agentic workflows.

    Good fit for

    • Supervised software development, debugging and repository maintenance.
    • Document analysis and synthesis across large collections of text.
    • Tool-assisted research and operational workflows with explicit permissions.
    • Mathematical and technical problem solving with independent checks.

    Poor fit for

    • Unsupervised agents with unrestricted production, credential or financial access.
    • High-stakes decisions without qualified human review.
    • Workflows requiring consistently reliable autonomous terminal operation.
    • Applications that treat untrusted retrieved content as authoritative instructions.
    Security & risk dashboardLatest vulnerabilities, advisories and real-world incidents for Claude Sonnet 4

    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.