How Cursor's Credit System Works:
The Complete 2026 Guide
Cursor replaced its old "fast request" system with a credit pool in June 2025. Your subscription now equals a dollar-amount budget that drains based on which AI models you use and how much you use them. Here is everything you need to know.
The Old System vs the New
Before June 2025 (Retired)
- Pro included 500 "fast requests" per month
- Each request counted as one, regardless of model or size
- After 500 fast requests, you got "slow" mode (longer wait times)
- Simple to understand but economically unsustainable for Cursor
- Different models cost Cursor vastly different amounts but users paid the same
After June 2025 (Current)
- Subscription equals a credit pool in dollars ($20 on Pro)
- Each request costs tokens at the model's API rate
- Auto mode is unlimited on all paid plans (free to use)
- Manual model selection drains from the credit pool
- Pay-as-you-go overages available at the same API rates
The shift made sense economically: a single Claude Sonnet request could cost Cursor 10 times more than a Gemini request, but both counted as one "fast request" under the old system. The credit pool aligns user costs with actual model costs. The downside is complexity. The upside is sustainability and model choice.
How Credits Work Now
1. Subscription = Pool
Your monthly plan price becomes your credit pool. Pro gives you $20, Pro+ gives approximately $70, Ultra gives approximately $400.
2. Requests Drain Pool
When you manually select a model, each request deducts credits based on token count multiplied by the model's API rate per token.
3. Auto Mode is Free
Auto mode lets Cursor pick the model. It is unlimited on all paid plans and never touches your credit pool. Most devs use it for 70-80% of work.
4. Overages at Same Rate
When your pool runs out, enable pay-as-you-go at the same API rates with no markup, or wait for the next billing cycle reset.
How a Single Request is Priced
Every AI request in Cursor involves sending tokens (your code context, chat history, instructions) and receiving tokens (the AI response). The cost of a request is calculated as:
Request cost = (input tokens x input rate) + (output tokens x output rate) + (cached tokens x cache rate)
For example, a typical Claude Sonnet request with 2,000 input tokens and 1,000 output tokens costs approximately: (2,000 / 1,000,000 x $3) + (1,000 / 1,000,000 x $15) = $0.006 + $0.015 = $0.021. In practice, with context windows and caching, the median request costs around $0.089 for Claude Sonnet. Cursor handles all of this math internally and just deducts from your pool.
Model Cost Comparison
Different models drain your credit pool at dramatically different rates. Choosing the right model for each task is the key to making your credits last.
| Model | Input / 1M tokens | Output / 1M tokens | Cache / 1M tokens | Median Request | Requests per $20 |
|---|---|---|---|---|---|
| Claude Sonnet 4.6 | $3 | $15 | $0.3 | ~$0.089 | ~225 |
| GPT-4.1 | $2 | $8 | $0.5 | ~$0.031 | ~650 |
| GPT-4o | $2.5 | $10 | $0.63 | ~$0.044 | ~455 |
| Gemini 2.5 Pro | $1.25 | $5 | $0.32 | ~$0.036 | ~550 |
What $20 buys by model
Claude Sonnet costs roughly 2.4x more per request than Gemini 2.5 Pro. GPT-4.1 offers the best value among paid models. Auto mode is unlimited and free.
Auto Mode: The Key to Value
Auto mode is the single most important feature to understand in Cursor's pricing system. When you enable Auto mode, Cursor automatically selects the most cost-effective model for each task. It is unlimited on all paid plans and does not drain your credit pool at all.
Cursor's Auto mode internally uses token pricing of approximately $0.25 per million cache read tokens, $1.25 per million input tokens, and $6 per million output tokens. But these costs are absorbed by Cursor, not passed to you. From your perspective, Auto mode requests are simply free.
The quality of Auto mode is surprisingly good. Cursor routes requests to the most appropriate model based on the complexity of the task. Simple completions might go to a fast, lightweight model, while complex multi-file reasoning gets routed to a more capable model. Most developers report that Auto mode handles 70 to 80 percent of their work without any noticeable quality difference versus manually selecting a premium model.
The optimal strategy for most developers: use Auto mode as your default and only manually select a specific model (like Claude Sonnet) when you need its specific strengths for complex reasoning, architectural decisions, or nuanced code review. This approach typically means your $20 credit pool only needs to cover 20 to 30 percent of your actual requests.
When to Use Auto Mode
- Routine code completions and suggestions
- Simple chat questions about syntax or APIs
- Standard refactoring and code cleanup
- Test generation and boilerplate code
- Documentation and comment writing
When to Select a Specific Model
- Complex multi-file architectural changes
- Debugging subtle logic errors that need deep reasoning
- Code review requiring nuanced understanding
- Performance optimisation with specific constraints
- When Auto mode gives unsatisfying results on a specific task
What Happens When Credits Run Out
Option 1: Auto Mode Continues
Auto mode is unlimited and keeps working even after your credit pool hits zero. For most developers, this means you can keep coding with AI assistance. You just lose the ability to manually select specific premium models until your pool resets.
Option 2: Pay-As-You-Go Overages
Enable overages in Cursor Settings. Additional usage is billed at the same API rates as your credit pool with no markup or penalty. Charges appear on your next invoice. You can set spending limits to prevent surprise bills.
Option 3: Wait for Reset
Your credit pool resets at the start of each billing cycle. If you are near the end of the month and want to avoid overages, you can use Auto mode exclusively and wait for the reset. Unused credits do not roll over.
BYOK: Bring Your Own Key
On any paid Cursor plan, you can bring your own API keys from Anthropic, OpenAI, Google, and other providers. When you use your own keys, requests bypass Cursor's credit pool entirely. You pay the API provider directly at their standard rates.
BYOK is recommended for two scenarios. First, heavy users who regularly exhaust their credit pool and want to avoid upgrading to a more expensive plan. Second, developers who need access to specific model versions or configurations not available through Cursor's default model selection.
To set up BYOK: open Cursor Settings, navigate to the Models section, and enter your API keys. You can configure Cursor to use BYOK for specific models while using the credit pool for others, giving you fine-grained control over your spending.
Supported BYOK Providers
Anthropic
Claude Sonnet, Claude Opus
$3-15/M input, $15-75/M output
OpenAI
GPT-4o, GPT-4.1, o3-mini
$2-2.5/M input, $8-10/M output
Gemini 2.5 Pro
$1.25/M input, $5/M output
Azure OpenAI
Same as OpenAI
Varies by deployment
The Pricing Controversy: What Actually Happened
In June 2025, Cursor abruptly switched from its fixed "500 fast requests" system to the credit-pool model described on this page. The transition was poorly communicated and caught many users off guard. Here is the timeline:
Cursor announces the credit-pool system. Pro users who were used to 500 predictable requests per month suddenly found their usage billed by token consumption at API rates. Many users reported their effective usage dropped significantly compared to the old system.
Community backlash on Reddit, Hacker News, and Twitter. Users complained about the lack of warning, confusing billing, and perceived price increase for heavy users. Some reported monthly costs doubling or tripling compared to the old flat-rate system.
Cursor CEO acknowledged 'mishandling' the transition in a public blog post. The company issued refunds to users who experienced unexpected charges during the transition period. They also improved the Settings dashboard to show real-time credit usage.
Cursor adjusted Teams pricing and added more granular usage controls. The admin dashboard for Teams was expanded with spending limits and per-user usage analytics to address enterprise concerns about unpredictable costs.
The credit system stabilised. Auto mode (unlimited on paid plans) was positioned as the primary way to use Cursor, with manual model selection as a secondary option. Most users adapted, though some migrated to competitors like GitHub Copilot or Claude Code.
The controversy matters for context: if you see older reviews or articles complaining about Cursor's pricing, they may be referring to the transition period rather than the current system. The credit-pool model is now stable and well-understood, even if the rollout was rough. Whether it is a better deal than the old system depends entirely on your usage patterns and model preferences.
Credit Optimisation Tips
Default to Auto Mode
Set Auto mode as your default and only switch to manual model selection when you specifically need a particular model's capabilities. This alone keeps most developers well within their credit pool.
Use Gemini for Quick Tasks
When you do need manual model selection, prefer Gemini 2.5 Pro for straightforward tasks. It costs roughly 2.4x less per request than Claude Sonnet while handling most coding tasks well.
Reserve Claude for Complex Work
Save Claude Sonnet for tasks that genuinely benefit from its stronger reasoning: complex architectural decisions, subtle bug analysis, and nuanced code review. Do not use it for simple completions.
Monitor Usage Weekly
Check your credit usage in Cursor Settings at least weekly. If you are burning through credits faster than expected, identify which models and tasks are consuming the most and adjust your workflow.
Set Spending Limits
Configure overage spending limits in Settings to prevent surprise bills. Start with a conservative limit and increase it as you understand your usage patterns over a few billing cycles.
Consider BYOK for Overflow
If you occasionally exceed your pool, bringing your own API key for one specific model can be cheaper than upgrading your entire plan. Keep BYOK as a safety valve rather than a primary strategy.