AI Progress Tracker

    Deepseek Reasoner · DeepSeek-V3.2 (deepseek-reasoner, thinking mode)

    DeepSeek · 1 Dec 2025

    Ask the Advisor about this model
    70from 7
    benchmarks

    DeepSeek-V3.2 powers the documented deepseek-reasoner thinking mode, combining text-based reasoning, coding and tool use. The API name is a rolling alias rather than a fixed checkpoint, so reproducibility depends on recording the underlying model version and evaluation date.

    context: 128Ktext$0.28 / $0.42 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 mathematics93.1%Dec 2025View
    GPQA DiamondPhD-level science questions82.4%Dec 2025View
    Humanity's Last Exam (without tools)25.1%Dec 2025View
    SWE-bench VerifiedResolving real GitHub issues73.1%Dec 2025View
    Terminal-BenchLong-horizon command-line agent tasks46.4%Dec 2025View
    OTIS Mock AIME 2024-25Competition mathematics87.8%16 Dec 2025View
    Aider PolyglotCode editing across six programming languages74.2%3 Oct 2025View

    Weighted mean of 7 verified benchmark results: AIME 93.1%, GPQA Diamond 82.4%, Humanity's Last Exam (without tools) 25.1%, SWE-bench Verified 73.1%, Terminal-Bench 46.4%, OTIS Mock AIME 2024-25 87.8%, Aider Polyglot 74.2%.

    Capabilities

    • Thinking mode supports multistep mathematical and scientific reasoning.
    • Integrates reasoning with tool calls for multistep agent workflows.
    • Supports code generation and repository-level software repair.
    • Accepts text contexts of up to 128K tokens.
    • Returns reasoning content separately from the final answer through the API.

    What improved

    • DeepSeek Sparse Attention reduces long-context computation compared with the dense-attention V3.1-Terminus architecture.
    • Integrates thinking with tool use, extending reasoning across tool-assisted workflows.
    • Expanded reinforcement learning and large-scale synthetic agent training strengthen reasoning and multistep task execution.

    Known flaws & limitations

    • The documented API interface is text-only, without native image or audio understanding.
    • Reasoning tokens consume output capacity and are billed as output tokens.
    • Published evaluations show substantial remaining errors on difficult knowledge and agentic tasks.
    • The deepseek-reasoner alias is not an immutable checkpoint, limiting reproducibility without version tracking.

    Risks, safety & security issues

    • Service-side storage and logging can expose conversations and credentials; a DeepSeek database exposure was publicly documented in January 2025.
    • Incorrect reasoning and generated code can cause downstream harm when used without independent validation.
    • Tool-connected deployments can turn erroneous decisions into external actions unless permissions, sandboxing and approval gates constrain execution.

    Good fit for

    • Mathematical and scientific problem-solving with independent answer checks.
    • Code generation, debugging and repository repair backed by automated tests.
    • Long-document analysis within the documented context limit.
    • Supervised tool-assisted workflows with narrowly scoped permissions.

    Poor fit for

    • Native image, audio or video understanding.
    • Unsupervised high-stakes decisions requiring reliably correct answers.
    • Autonomous execution with unrestricted access to production systems or credentials.
    • Strictly reproducible deployments that rely only on a rolling API alias.
    Security & risk dashboardLatest vulnerabilities, advisories and real-world incidents for Deepseek Reasoner

    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.