How to use Claude Sonnet 4.6 inside VS Code: three setup paths compared

This article was originally published on aicoderscope.com

Claude Sonnet 4.6 is the strongest daily-driver coding model available right now: 1 million token context window, $3 per million input tokens, and fast enough for interactive use. Getting it into VS Code is a 10-minute job — but there are three meaningfully different ways to do it, and they suit different workflows. This guide walks through each path with exact steps, then tells you which one to start with.

Pricing verified against Anthropic's model overview page on May 9, 2026.

---

What you're choosing between

Path	Extension	Cost model	Best for
A	Claude Code (official)	Claude.ai subscription or API pay-per-use	Most developers; Anthropic-native experience
B	Cline	API pay-per-use (BYOK)	Autonomous agent workflows; full cost control
C	Continue.dev	API pay-per-use (BYOK)	Open-source preference; mixed-model setups

All three give you Claude Sonnet 4.6 inside the VS Code editor. The difference is in how much setup, how much control, and what happens to your token bill.

---

The API cost picture first

Before picking a path, understand what Claude Sonnet 4.6 actually costs:

• Input tokens: $3.00 per million
• Output tokens: $15.00 per million
• Batch API (for non-interactive tasks): 50% off both rates
• Prompt caching: cached input drops to $0.30 per million (90% reduction)
• Context window: 1 million tokens (~750,000 words)

In practice, a typical hour of chat-assisted coding — describing a function, reviewing a diff, asking why a test fails — consumes roughly 50,000–100,000 input tokens and 5,000–20,000 output tokens. That's $0.15–$0.60 per hour. Heavy autonomous agent sessions (Cline doing a multi-file refactor) run $0.50–$1.00 per task.

Compared to Cursor Pro at $20/month flat, direct API access is cheaper for occasional use and more expensive for heavy daily use. The crossover is roughly 15–20 hours of active coding per month.

New Anthropic accounts get approximately $5 in free credits to test the API. There is no ongoing free tier — once the starter balance is gone, you need to purchase credits or enable auto-reload at platform.claude.com.

---

Path A: Official Claude Code extension

The Claude Code VS Code extension is Anthropic's own product. It's the path with the least friction and the tightest integration.

Requirements

• VS Code 1.98.0 or higher (check via Help → About)
• An Anthropic account (free to create at claude.ai) OR an API key from platform.claude.com

Install

1. Press Ctrl+Shift+X (Windows/Linux) or Cmd+Shift+X (Mac) to open Extensions
2. Search for Claude Code, publisher: Anthropic
3. Click Install
4. A Spark icon (✦) appears in the Editor Toolbar once a file is open

First sign-in

Click the Spark icon and sign in with your Claude.ai account in the browser that opens. If you prefer to use an API key directly, set ANTHROPIC_API_KEY in your shell environment and launch VS Code from that terminal with code . so it inherits the variable.

What you get

The Claude Code panel opens as a sidebar or editor tab. The key features:

• @-mentions: type @filename to pull a specific file into context, or @src/components/ with a trailing slash for a folder. Press Alt+K (Windows/Linux) or Option+K (Mac) to insert a reference to your current selection with line numbers.
• Permission modes: click the mode indicator at the bottom of the prompt box. Default mode asks before each file edit. Plan mode has Claude describe what it will do and waits for your approval before touching anything — useful when you're giving it a multi-step task. Auto-accept is for when you're certain and want flow.
• Checkpoints: hover over any message to rewind. You can fork the conversation from a point in history, revert file changes to that point, or both. This is the escape hatch when an agent session goes sideways.
• Multiple conversations: Ctrl+Shift+Esc (Windows/Linux) opens a new conversation in a separate tab. Useful for running one refactor while asking questions about another part of the codebase.
• MCP server support: run claude mcp add in the integrated terminal to connect external tools, then manage them with /mcp in the chat panel.

The extension includes the Claude Code CLI. Open the integrated terminal (Ctrl+`) and run claude for terminal-based workflows — these share the same conversation history as the panel.

Billing

If you use a Claude.ai Pro subscription ($20/month), Claude Code usage is included but rate-limited. For heavier use, switch to API-key billing at platform.claude.com — you pay per token at the rates above, with no monthly floor.

---

Path B: Cline

Cline (formerly known as Claude Dev, now published by Cline Bot Inc.) is the most-installed third-party AI coding agent for VS Code: 3.9 million installs as of May 2026, version 3.82.0. It's free to install; you bring your own API key.

Install

1. Open Extensions (Ctrl+Shift+X)
2. Search Cline, publisher: Cline Bot Inc.
3. Click Install

Configure with Claude Sonnet 4.6

1. Click the Cline icon in the Activity Bar (left sidebar)
2. Click the gear icon → API Provider → select Anthropic
3. Paste your Anthropic API key from platform.claude.com/settings/keys
4. Under Model, select claude-sonnet-4-6
5. Click Save

What Cline does differently

Cline is explicitly an autonomous agent. You describe a task, and it plans the steps, creates and edits files, runs terminal commands, and navigates the browser — all while showing you exactly what it intends to do before each action. This is different from the Claude Code extension, which is more of a collaborative assistant you direct step by step.

Token usage is displayed in real time in the Cline sidebar. You can see your running cost per task. This transparency is useful: complex refactoring tasks where Cline has to read dozens of files first can accumulate 200,000–400,000 input tokens before writing a single line.

Cline also supports OpenRouter, Google Gemini, AWS Bedrock, local models via Ollama/LM Studio, and any OpenAI-compatible API — so if you want to route to Claude through a different provider (e.g., AWS Bedrock for enterprise data residency), Cline handles it.

When to use Cline over the official extension

Cline wins when the task is: "take this spec and build a working feature, checking in with me at each decision point." The official Claude Code extension wins for: "I'm writing code, I want a smart pair programmer in the sidebar." Cline is more autonomous; Claude Code is more collaborative.

---

Path C: Continue.dev

Continue.dev is an open-source VS Code extension licensed under Apache 2.0. Its differentiating feature: it lets you mix models. You can use a small, fast local model (like Qwen2.5-Coder 7B via Ollama) for tab completion — where latency matters most — while routing chat and agent tasks to Claude Sonnet 4.6. This is the cheapest overall configuration for a developer who writes a lot.

Install

1. Open Extensions (Ctrl+Shift+X)
2. Search Continue, publisher: Continue
3. Click Install

Configure Claude Sonnet 4.6 in config.yaml

Continue uses a config.yaml file for configuration. Open it by clicking the Continue icon in the sidebar → gear icon. Add an Anthropic provider block:

models:
  - name: Claude Sonnet 4.6
    provider: anthropic
    model: claude-sonnet-4-6
    apiKey: ${{ secrets.ANTHROPIC_API_KEY }}
    roles:
      - chat
      - edit
      - apply

For tab completion, add a separate block pointing to a local model to avoid the la