Chuyển tới nội dung chính

1 bài viết được gắn thẻ "performance-analysis"

Xem tất cả thẻ

Claude AI in November 2025: A Month of “Extreme Reasoning”, Skills & Computer-Use

· 3 phút để đọc
Claude Dev
Claude Dev
Claude Dev

November 2025 saw Anthropic move the Claude family from “helpful chatbot” to “agentic co-worker”.
Below are the three headline moves you may have missed while you were wrapping up Q4.

1. “Extreme Reasoning” drops – Opus 4 & Sonnet 4 think longer, code better​

  • What changed

    • Claude Opus 4 becomes Anthropic’s flagship coding model, built for multi-hour agent loops.
    • Claude Sonnet 4 gets a 2 × speed-up plus higher instruction-fidelity.
    • Both models ship in two modes:
      1. Fast (sub-second)
      2. Extended-thinking (up to 5 min), letting the model search, test and debug its own outputs.

  • Why it matters
    Internal benchmarks show a 42 % jump on SWE-bench (real GitHub issues) vs. 3.5-Sonnet, with pass@1 above 70 % when the agent is allowed to iterate.
    Early adopters report 7-hour autonomous coding sessions that complete full feature branches without human hand-off [^14^].

  • How to try

    • Available today on claude.ai (Max/Team/Enterprise) and the Anthropic API.
    • Toggle “Extended thinking” in the UI or set thinking_budget_tokens in the API.

2. “Skills” GA – turn Claude into your company’s mini-employee​

  • What it is
    Skills are portable folders that bundle instructions, Python/R scripts, brand guidelines, SQL queries—anything Claude needs to repeat a workflow.
    Think “Excel macro” meets “GPT”, but version-controlled and shareable across seats.

  • Ships with 20 pre-builds

    • “Quarterly-earnings parser” (pulls tables from PDFs, writes CEO summary)
    • “Canva brand-guard” (auto-crops to template, exports 4 sizes)
    • “Jira→Slack sprint digest”

  • Who gets it
    Pro, Max, Team and Enterprise plans. API & Agent SDK support landed Nov-18 [^3^].

3. Computer-Use graduates from beta – Claude now drives your desktop​

Originally teased in October, the 3.5-model that can see pixels, move the cursor and type is now production-grade.

  • New in November

    • Multi-app workflows (e.g., pull data from Snowflake, paste chart into Google Slides, export PDF).
    • Vision accuracy ↑ 18 % on OSWorld leaderboard.
    • SOC-2 Type II compliance ⇒ approved for regulated industries [^15^].

  • Pricing
    $0.60 / successful task (success = user clicks “Approve”). Free tier gets 25 tasks/month until Jan-2026 promo ends.

Quick hits you might have scrolled past​

  • Web-search leaves beta – now on every paid tier, citations auto-inserted [^5^][^12^].
  • 1-hour prompt-cache – keep a 1 M-token context hot for <$0.20, perfect for book-length docs [^14^].
  • GitHub Actions for Claude Code – run nightly test-fix loops without a server [^14^].

Looking ahead​

Anthropic’s roadmap slide (leaked Nov-29) hints at:

  • Memory v2 – cross-conversation recall for individual free users (Dec).
  • Claude 4 Haiku – 200 Hz, sub-$0.10 / 1 K tokens, aimed at embedded devices (Q1-26).
  • European region – GDPR-compliant endpoints in Ireland (Feb-26).

Bottom line​

November 2025 marks the moment Claude stopped asking you for perfect prompts and started bringing its own toolkit to work.
If you haven’t given Extended-thinking or Skills a spin, schedule a 30-minute sandbox before year-end—your 2026 self will thank you.

Happy building!