Models in Cursor
Cursor gives you access to frontier models from every major provider — OpenAI, Anthropic, Google, xAI — plus Cursor's own purpose-built models. You can switch models per task within the same session. The default Auto mode picks the best available model automatically and is unlimited on paid plans.
Auto Mode
Auto mode is the default and the recommended starting point. Cursor selects the model based on the task complexity, current server load, and model performance. It tends to route to capable frontier models for complex agent tasks and faster models for quick completions. Auto mode is unlimited on all paid plans — it does not draw from your monthly usage credits. For most users, staying in Auto mode is the right choice.
Available Models
Cursor supports models from every major provider. The exact list changes as new models are released — check the model selector in Cursor for the current set. As of mid-2026, the categories are:
Anthropic — Claude family
Claude models (Sonnet, Opus, Haiku) from the Claude 4 generation. Claude Sonnet is popular for coding tasks — strong at reasoning through complex refactors, writing tests, and understanding large codebases. Opus is the most capable but slowest and most expensive on credits.
OpenAI — GPT and o-series
GPT-5 family for general-purpose coding, and the reasoning-focused o-series (o3, o4) for problems that benefit from extended thinking — algorithm design, complex debugging, architecture decisions. o-series models are slower and use more credits per call.
Google — Gemini family
Gemini Pro models with very long context windows. Useful when you need to load many large files into context simultaneously. Also strong at structured data extraction and tasks that involve processing large amounts of text.
xAI — Grok
Grok models, particularly useful for tasks that benefit from real-time web knowledge. Available as an alternative provider in Cursor's model selector.
Cursor — purpose-built models
Cursor's own models optimized for specific tasks. The Tab completion model is always Cursor's own — built for speed and prediction accuracy in an editing context, not general conversation. Additional Cursor models may appear in the selector for specific use cases.
How Usage Credits Work
Every paid plan includes a monthly credit pool equal to the plan price: Pro = $20, Pro+ = $60, Ultra = $200. When you manually select a specific model (rather than Auto), Cursor charges tokens at that model's per-token rate against your credit pool. If you exhaust credits, Cursor charges usage-based overage to your card.
| Mode | Uses Credits? | Notes |
|---|---|---|
| Auto mode (Chat/Agents) | No — unlimited | Cursor picks model; no credit charge |
| Manual model selection | Yes — draws from pool | Charged at that model's per-token rate |
| Tab completion | No — unlimited | Always uses Cursor's own Tab model |
| Cloud Agents | Yes — draws from pool | Charges based on model and tokens used |
Which Model for Which Task
Daily coding, multi-file edits
Auto mode — let Cursor decide
Fastest path, no credit drain
Complex refactor or architecture
Claude Sonnet or Opus
Strong at multi-step reasoning across large context
Algorithm design or hard debugging
o3 or o4 (reasoning models)
Extended thinking produces more reliable answers on hard problems
Long-context tasks (many large files)
Gemini Pro
Highest context window in the Cursor model roster
Tab completions
Always Cursor Tab (automatic)
Purpose-built for speed; not in the manual selector
MCP Integration
On Pro and above, Cursor supports the Model Context Protocol (MCP) — a standard for connecting external tools and data sources into the model's context. Configure MCP servers to give agents access to databases, internal APIs, documentation systems, or any external resource. MCP servers are configured per-project in .cursor/mcp.json or globally in Cursor settings.
Checklist: Do You Understand This?
- Auto mode is unlimited on paid plans — Cursor picks the model, no credit charge
- Manual model selection draws from your monthly usage credit pool ($20 on Pro, $60 on Pro+, $200 on Ultra)
- Tab completion always uses Cursor's own purpose-built model — not in the manual selector
- For complex reasoning tasks, o-series models (o3/o4) outperform general chat models
- MCP support (Pro+) lets agents connect to external tools via the Model Context Protocol