Claude Sonnet 4

• 1. overview (552 chars)

  Claude Sonnet 4 (actual version is Claude Sonnet 4.6) is a mid-range model of the Claude 4 series released by Anthropic in February 2026. It is positioned as a general-purpose large model that takes into account performance and cost control. As the default model for free and Pro users of the Claude platform, Sonnet 4.6 has been fully upgraded in core capabilities such as code writing, computer operations, and long-context reasoning. Its performance is close to or even surpasses the higher-end Opus 4.6 in some dimensions, but the price is only one-fifth of the latter. It is considered to be one of the most cost-effective AI models currently on the market.

• 2. Product background (763 chars)

  Claude Sonnet 4 is developed by American AI security company Anthropic. Anthropic was founded in 2021 by former OpenAI core members Dario Amodei and Daniela Amodei, focusing on AI safety and alignment research. Since its establishment, the company has completed multiple rounds of financing, with investors including Google, Salesforce, Zoom and other technology giants.   Since the first release of the Claude series models in 2023, it has formed a product line of three grades: Haiku (lightweight and fast), Sonnet (balanced performance), and Opus (high-end capabilities). Sonnet 4.6 is the third generation of the Sonnet series, which includes the Claude Sonnet 3.5 and Sonnet 4.5.   On February 17, 2026, Anthropic officially released Claude Sonnet 4.6, replacing Sonnet 4.5 as the default model. Also released at the same time were the high-end version Opus 4.6 and the lightweight version Haiku 4.5.

• 3. Core functions (1086 chars)

  Claude Sonnet 4.6 surpasses previous generation products in multiple dimensions, especially in the following areas:   **Programming skills**: code generation, bug fixing, unit test writing, cross-file reconstruction, and understanding of complex project structures. Achieving 79.6% on the SWE-bench Verified benchmark, close to Opus 4.6’s 80.8%.   **Long context reasoning**: Supports context windows up to 200K tokens (API users can pay for the 1M token version). In the long text retrieval task, the accuracy of finding the correct information from 256,000 tokens soared from 10.9% in Sonnet 4.5 to 90.3%.   **Computer Use**: Ability to understand and operate graphical user interfaces, complete automated tasks such as web form filling, data entry, cross-application operations, etc., and achieve near-human level operation accuracy.   **Multi-modal understanding**: Supports image, screenshot, PDF and chart parsing, able to understand visual content and make inferences.   **Adaptive thinking mechanism**: Dynamically allocate computing resources according to task complexity, respond quickly to simple tasks, and think deeply about complex tasks.   **Agent task planning**: Built-in tool calling interface supports multi-step task decomposition and execution.

• 4. Technical architecture (478 chars)

  Sonnet 4.6 uses a number of advanced technologies:   - **Mixed Expert Architecture (MoE)**: sparse activation structure, reducing inference calculation load - **Dynamic Computing Scheduling**: Adaptive thinking mechanism, allocate computing resources according to task complexity - **Long context optimization**: improved attention algorithm and positional encoding - **Visual-Text Fusion**: Cross-modal reasoning in a unified semantic space - **RLHF Alignment Training**: Reinforcement learning based on human feedback to improve output stability

• 5. Compare with competing products (438 chars)

  | Dimensions | Claude Sonnet 4.6 | GPT-4o | Gemini 2.5 Pro | DeepSeek V3 | |------|-------------------|----------|----------------|-------------| | Coding ability | ★★★★☆ | ★★★★★ | ★★★★☆ | ★★★★☆ | | Instructions to follow | ★★★★★ | ★★★★☆ | ★★★★☆ | ★★★☆☆ | | Long text | ★★★★★ | ★★★★☆ | ★★★★★ | ★★★☆☆ | | Value for money | ★★★★★ | ★★★☆☆ | ★★★★☆ | ★★★★★ |   Sonnet 4.6 has clear advantages in terms of coding capabilities and instruction compliance, but is slightly inferior to GPT-4o and Gemini 2.5 Pro in terms of multi-modal support.

• 6. Pricing plans (304 chars)

  Claude Sonnet 4.6 adopts a token-based API pricing model: - **Enter price**: $3/million tokens - **Output Price**: $15/million tokens - **1 million token version**: $6/million tokens (input), $22.5/million tokens (output)   Compared with Opus 4.6 ($15/million token input, $75/million token output), the price of Sonnet 4.6 is only one-fifth of the former.

• 7. Access method (239 chars)

  - **Online Platform**: https://claude.ai (Free and Pro users use Sonnet 4.6 by default) - **API interface**: accessed through Anthropic API, AWS Bedrock, Google Vertex AI - **Developer Console**: https://console.anthropic.com - **API Model ID**: `claude-sonnet-4-20260301`

• 8. target users (307 chars)

  - **Software Developer**: Code generation, refactoring, Review - **Enterprise Knowledge Management**: long document analysis, summarization, Q&A - **Office Automation**: report writing, data sorting, daily task processing - **Smart Customer Service**: Integrated into enterprise Q&A system - **Data Analysis**: chart interpretation, report analysis

• 9. positive review (464 chars)

  The community gradually formed a consensus: regard Opus as the "thinker" and Sonnet as the "executor". The recommended workflow is to use Opus for planning and architecture and Sonnet for specific tasks.   In Claude Code, you can use `/plan` to let Opus create a plan and then assign it to Sonnet subagents for execution. This division of labor is considered best practice.   User feedback Sonnet 4.6 performs well in command compliance, with format deviation reduced from approximately 15% to less than 3%, which is very important for automated processes.

• 10. negative feedback (618 chars)

  **Inadequate creative writing skills**: Sonnet 4.5 has seen a significant decline in writing quality since January 2026, and 4.6 is improved but not as good as when 4.5 was first released. There is still too much modern slang in the writing of historical fiction, and it is impossible to fully grasp the historical context.   **Context window limit**: 1 million tokens. Context is only available to API users and requires additional payment. Web and mobile users are still limited to 200,000 tokens, which has disappointed many users.   **Difficulty in model selection**: It is difficult for ordinary users to determine which model to use in which scenario. Some users hope that Claude can automatically select the most appropriate model.

• 11. media opinion (251 chars)

  Industry media generally believe that Claude Sonnet 4.6 is a price-performance revolution. TechCrunch and other media pointed out that Sonnet 4.6 provides performance close to Opus at only one-fifth the price of Opus, and is one of the most balanced and practical AI models currently on the market.

• 12. Expert analysis (300 chars)

  Experts believe that Sonnet 4.6 greatly surpasses the previous generation in code generation, instruction following and long text understanding, and is currently the most worthy of consideration for developers' daily use. It is especially suitable for scenarios that require processing of extremely long documents and complex code base maintenance.

• 13. Competitive product landscape (284 chars)

  In the large model market in 2026, Sonnet 4.6’s main competitors include OpenAI’s GPT-4o, Google’s Gemini 2.5 Pro, domestic DeepSeek V3 and Zhipu GLM-4, etc. Sonnet 4.6 has clear advantages in terms of coding capabilities and instruction compliance, but is slightly inferior to GPT-4o and Gemini 2.5 Pro in terms of multi-modal support.

• 14. critical voice (551 chars)

  **"Old models are weakened" controversy**: Every time a new model is released, users complain that old models are weakened. Some users believe that Anthropic deliberately reduces the performance of old models in order to promote new models, but some users refute this. This is a common comment every time it is released. Objective data shows that the model is indeed continuously improving.   **Disappointment from the creative writing community**: Many users who rely on AI-assisted writing complained that the quality of creative writing in Sonnet 4.6 has declined and cannot meet tasks that require creativity such as novel creation and copywriting.

• 15. Potential risks (442 chars)

  **High-risk decision-making scenarios**: Model output still requires manual review and is not suitable for direct use in high-risk decision-making scenarios such as professional legal and medical care.   **Data Privacy**: Enterprise users need to pay attention to data privacy and protection issues when using APIs to avoid leakage of sensitive information.   **Risk of over-reliance**: Over-reliance on AI-assisted programming may lead to the degradation of developer skills and needs to be maintained in moderation.

• 16. Suitable for whom (342 chars)

  - **Software Developer**: Daily coding, code review, and refactoring tasks - **Technical Team**: Development and operation and maintenance teams that require cost-effective AI assistance - **Enterprise knowledge worker**: processing long documents, writing reports, data analysis - **Automation process developer**: Agent developers who need stable instruction following and tool invocation

• 17. Not suitable for whom (335 chars)

  - **Creative Writer**: Users who need high-quality creative writing (it is recommended to choose Opus or other specially optimized models) - **Multi-modal application developers**: applications that require video and audio processing (Gemini 2.5 Pro or GPT-4o is recommended) - **Individual developers with extremely limited budget**: You can consider the more cost-effective DeepSeek V3

• 18. alternative (168 chars)

  - **Requires stronger reasoning skills**: Claude Opus 4.6 - **Requires stronger multi-modal capabilities**: GPT-4o or Gemini 2.5 Pro - **Requires lower cost**: DeepSeek V3 or domestic large model

• 19. Summarize (491 chars)

  Claude Sonnet 4.6 is one of the best overall value for money AI models in 2026. It has reached the ceiling of mid-range models in three dimensions: code generation, instruction following, and long text understanding, while the price is only one-fifth of the high-end model. While creative writing capabilities are still lacking, Sonnet 4.6 is useful and affordable enough for most practical scenarios.   For enterprises and individual developers who need digital employees who can work, run fast, have good brains and are not expensive, Claude Sonnet 4.6 is currently the best choice.

No in-depth report yet

• 

  Melissa.BakerSr

  —

  绝了

---

• 

  Kelly_Anderson168

  —

  新手请教一下，Sonnet 4.6 和 DeepSeek V3 哪个更推荐？主要用来学编程和做课程项目，预算有限。

---

• 

  谭萍

  —

  刚把团队 70% 的调用切到 Sonnet 4.6，跑了三周无质量事故，月账单直接腰斩。代码生成和指令遵循确实够用了，复杂重构才需要切 Opus。

---

• 

  IBCinter46

  —

  Sonnet 4.6 的指令遵循比 4.5 强太多，以前写自动化流程经常格式跑偏，现在格式偏差从 15% 降到 3% 以内，终于不用每次都手动修正了。

---

• 

  blackrabbit930

  —

  太强了！

---

• 

  EdwardBaker_2021

  —

  有个问题想请教大家，Sonnet 4.6 的 100 万 token 上下文是需要单独付费的吗？网页端还是只有 20 万？

---

• 

  BAgra

  —

  用 Sonnet 4.6 做代码 Review 两周了，体验非常好。能准确识别跨文件的代码依赖，给出的重构建议也很实用。唯一缺点是复杂系统设计还是得用 Opus。

---

• 

  Brian.Hernandez_2020

  —

  卡成 PPT

---

• 

  Mark_Nelson_Plus

  —

  我们团队用 Sonnet 4.6 做 Bug 检测，与 Opus 的差距显著缩小，现在可以并行跑更多审查，捕捉更广泛的 bug 类型，而且成本不增加。这对小团队来说太重要了。

---

• 

  Christina_Morales_202308

  —

  免费的

---

• 

  RBell_2022

  —

  想问问大家，Sonnet 4.6 和 GPT-4o 比哪个更强？主要用来写代码和做数据分析。

---

• 

  TPatel369700

  —

  Sonnet 4.6 在长文档推理方面确实有突破，从 25.6 万 token 中找出正确信息的准确率从 4.5 的 10.9% 飙升至 90.3%，这个提升太夸张了。处理超长合同和论文的时候特别有用。

---

• 

  KeithFloresSr

  —

  不推荐用来写小说，创意写作能力还是不行。我试了好几次，写出来的东西总有现代俚语， historical novel 根本把握不好历史背景。还是用 Opus 或者专门的小说 AI 吧。

---

• 

  JWilson168

  —

  yyds

---

• 

  VPowell_770

  —

  我们公司用 Sonnet 4.6 做智能客服，效果出乎意料的好。指令遵循准确，多轮对话不跑偏，而且成本只有 Opus 的五分之一。一天 2000 次调用，月成本才一千出头，比原来的方案便宜太多了。

---

• 

  RuthHall_20209

  —

  有个小技巧分享：在 Claude Code 里用 /plan 让 Opus 创建计划，然后分配给 Sonnet 子代理执行。这样既保证了架构质量，又控制了成本，是目前最优的工作流程。

---

• 

  VEdav

  —

  回不去了

---

• 

  MacitYalçın

  —

  Sonnet 4.6 的 Computer Use 能力确实提升了，能理解和操作 GUI，填写表单、跨应用操作都接近人类水平。我们在做 RPA 流程的时候试了一下，成功率比 4.5 高很多。

---

• 

  HCampbell_2024

  —

  价格太香了，输入 $3/百万 token，输出 $15/百万 token，比 Opus 便宜五倍。我们算过账，客服场景一天 2000 次调用，Sonnet 月成本 ¥1026 vs Opus ¥5130，月差 ¥4000+，而且回答质量几乎无差别。

---

• 

  KennethSimmons_2020

  —

  有个疑问，Sonnet 4.6 支持图片输入吗？能做 OCR 和图表解读吗？看到文档说支持多模态，但不确定具体能力边界。

---

• 

  枫叶_8

  —

  这也太贵了

---

• 

  DrHarryWood_2024

  —

  我们团队用 Sonnet 4.6 做前端开发，视觉输出比之前精致很多，布局、动画、设计感都上了一个台阶。Rakuten AI 测试里生成了最佳的 iOS 代码，确实有点东西。

---

• 

  FCruz

  —

  Warning: 不要用 Sonnet 4.6 做高风险决策，模型输出还是需要人工审核的，不适合直接用于专业法律、医疗等场景。我们之前差点犯了这个错误，还好最后有人工复核。

---

• 

  ZoeLo

  —

  Sonnet 4.6 在 SWE-bench Verified 基准测试达到 79.6%，接近 Opus 4.6 的 80.8%，这个成绩对于中端模型来说已经很夸张了。日常写代码完全够用，只有复杂系统级重构才需要上 Opus。

---

• 

  yellowfish230

  —

  Anthropic 今天发布了 Sonnet 4.6，正式替代 Sonnet 4.5 成为免费和 Pro 用户的默认模型。价格是 $3/$15 per million tokens，比 Opus 便宜五倍，但性能接近 Opus 水平。这是一次性价比革命！

No reviews yet