Cursor Official Docs: Claude Sonnet 5 Model Integration, Usage Pools, and Agent Tool Configuration

One-sentence takeaway

Cursor positions Claude Sonnet 5 as the primary mid-tier coding model to replace Sonnet 4.6, supporting a 1M context window, thinking mode, and a complete Agent tool suite, with no long-context multiplier fees charged for contexts exceeding 200k tokens.

Use cases

• Suitable tasks: code completion within Cursor IDE, Agent autonomous coding, multi-file refactoring, terminal command execution, automated browser debugging, and rules retrieval.

• Unsuitable tasks: extremely complex cross-architecture theoretical design (the official docs recommend upgrading to Opus 5 for such tasks) .

• Applicable model version: Claude Sonnet 5 (Cursor-managed model pool) .

• Applicable client, agent, or API: Cursor IDE (Agent mode, Chat mode, Composer) .

• Recommended reasoning levels and parameters: enable the thinking mode variant (Thinking Variant) ; use default effort for everyday tasks, and increase effort for complex debugging.

Ready-to-use content

Cursor model and pricing configuration notes

{
  "model": "claude-sonnet-5",
  "pricing": {
    "input_per_mtok": 2.0,
    "cache_write_per_mtok": 2.5,
    "cache_read_per_mtok": 0.2,
    "output_per_mtok": 10.0
  },
  "context_window": 1000000,
  "long_context_markup": false,
  "usage_pool": "third_party_other",
  "agent_tools": [
    "Search files and folders",
    "Web",
    "Read files (vision-capable)",
    "Edit files",
    "Run shell commands",
    "Browser (screenshots & visual verify)",
    "Image generation",
    "Ask clarifying questions",
    "Fetch rules"
  ]
}

Recommended Cursor Rules configuration example (.cursorrules)

# Claude Sonnet 5 Project Rules
- Execution Mode: Prefer test-driven implementations; write minimal unit tests before modifying core logic.
- Conciseness: Avoid unsolicited boilerplate or helper utilities unless directly required by the user requirement.
- Tool Usage: Inspect files before proposing diffs; verify terminal output after running build or test commands.

Test/workflow steps

1. In Cursor Settings > Models, toggle on and enable Claude Sonnet 5, and set it as the default Agent model.

2. Verify the third-party model usage pool quota in your individual or team plan (individual plans include at least $20 of third-party usage per month) .

3. Submit tasks in the Agent pane; Sonnet 5 will automatically orchestrate file search, reading, code editing, and terminal command execution.

4. For large codebases exceeding 200k context tokens, there is no need to worry about tiered markups—the system bills uniformly at the base per-token rates.

Original evidence and data

• Official pricing: $2/M input, $2.5/M cache write, $0.2/M cache read, and $10/M output.

• Billing characteristics: All Sonnet 5 prompts are billed at base per-token rates, with no separate long-context multiplier even when context exceeds 200k tokens.

• Tokenizer changes: The updated tokenizer breaks down identical input into more tokens, resulting in a slight increase in token count compared to earlier Sonnet models.

Scope and limitations

• Belongs to the third-party usage pool, consuming the monthly $20 base credit or additional pay-as-you-go balance without drawing from Cursor's proprietary fast model pool.

• In ultra-long sessions, due to tokenizer density adjustments and thinking process outputs, the overall token consumption rate is faster than Sonnet 4.6.

Source excerpts or observations (short compliant quotations only)

• Official documentation quote: “Claude Sonnet 5 is Anthropic's latest medium-tier model and replaces Sonnet 4.6... ideal default choice for day-to-day coding.”

• Official billing quote: “All Sonnet 5 prompts are billed at the base per-token rates... even when context exceeds 200k.”