# Claude Code review: experiences, strengths and weaknesses | llmnet Apps [← Apps overview](/en/) # Claude Code: review By Ivo Donker · October 28, 2026 — created with AI assistance (writer: DeepSeek V4.1 Flash; critic: gpt-oss-120b via Groq) Claude Code: [Review](/en/claude-code-review)[Install & set up](https://apps.llmnet.nl/en/claude-code-installatie)[News](/en/claude-code-nieuws)[Technical tutorial](https://apps.llmnet.nl/en/claude-code-tutorial) Current as of September 2026, compiled from our own notes and news archive; text machine-generated and editorially checked. Check versions and commands in the official documentation. Claude Code is Anthropic's coding agent: you give an instruction in plain language and the agent reads your code, edits files, runs commands and makes commits. In this review you can read what it is, where its strengths and weaknesses lie, what it costs and who it is suitable for (as of September 2026). ## What it is and who it is for Claude Code runs in your terminal, as an extension in VS Code and JetBrains, as a desktop app and in the browser. The core is the same everywhere: a Claude model with access to your project directory, a shell and external tools via MCP. It is made for developers who would rather delegate a task than get line-by-line autocomplete: having a bug traced and fixed, a refactor across dozens of files, writing tests, preparing a PR. It works both interactively and unattended. With claude -p it runs in scripts and CI pipelines, and via the Agent SDK you can use it to build your own agents in Python or TypeScript. ## How it works At every session, Claude Code loads your project instructions: CLAUDE.md, or from version 2.1.277 an existing AGENTS.md if there is no CLAUDE.md. In addition, it keeps its own "auto memory" with lessons from earlier sessions. Every action falls under permission rules: anything not permitted in advance is first asked of you. With hooks you enforce fixed steps, such as lint after every edit. An OS-level sandbox limits which files and domains shell commands may touch. You extend it with skills (reusable ways of working), subagents (subtasks with their own context) and plugins that bundle these components. ## Strengths - Model quality per euro. The default model Claude Opus 5.5 (since 22 September 2026) costs, according to Anthropic, $4 per million input tokens and $20 per million output tokens. For typical workloads, that is about 40% cheaper than Opus 5. Fast mode provides up to 2.5 times the speed for $8 input and $40 output per million tokens. - Secure to configure. Permission rules, hooks and an OS sandbox are built in and can be managed in layers: user, project and organisation. A cloned repository cannot grant itself broader permissions than you allow. - Easy to automate. Headless mode with JSON output, a schema for structured responses, cost estimation per run and a --bare mode for reproducible CI. - Open to other tools. It reads AGENTS.md, /init takes over Cursor and Copilot rules, and MCP provides access to a large ecosystem of servers. ## Weaknesses and risks - High release cadence. Several versions often appear per week. The latest channel delivers new features immediately, but sometimes also a regression. The stable channel lags by about a week and skips versions with major bugs; for production work, that is the sensible choice. - Costs can add up. Long sessions, large CLAUDE.md files and many subagents fill the context quickly. On a subscription you notice this in the limits, via the API in the bill. - A supply chain you have to watch. Plugins, MCP servers and hooks are third-party code. In 2026, the "Plugin4Shell" vulnerability affected several coding agents: attackers could bypass commit pinning in plugin marketplaces. Claude Code patched the vulnerability in version 2.1.179. Keep updating and install only plugins you trust. - No free entry point. The free claude.ai plan does not provide access; you need a Pro, Max, Team or Enterprise subscription, or API credit. - Windows is second-class. Native Windows works, but the sandbox does not; for that you need WSL 2. ## Price and licence You install the software itself for free; you pay for model usage. This can be through a Claude subscription (Pro, Max, Team, Enterprise) with usage limits, or per token through a Console account or a cloud provider such as Amazon Bedrock, Google Cloud or Microsoft Foundry. Claude Code is not open source. Current prices are on Anthropic's pricing page. ## Alternatives - Codex (OpenAI): a comparable terminal agent with GPT models, open source. - Cursor: an editor (VS Code fork) with agent features; more pleasant if you mainly work graphically. - OpenCode: an open-source terminal agent that works with many different model providers. - Hermes Agent (Nous Research): open source, strong in self-hosted and long-running agent setups. ## Conclusion Claude Code is one of the strongest choices for developers who want to hand serious tasks to an agent while keeping that agent tightly constrained at the same time. It suits best those who work in the terminal or in CI and are prepared to spend half an hour on permissions, hooks and a good CLAUDE.md. Anyone who only wants an occasional suggestion in the editor, or who insists on open source or running their own model, is better off looking at Cursor, OpenCode or Hermes.