Token-based Credits

Runcell calculates Ask/Chat and Agent usage from the actual number of tokens used by the selected model. Input tokens and output tokens are priced separately because models often charge different rates for reading context and generating responses.

Token-based Credits currently apply to Ask/Chat and Agent mode. Other AI features, such as code completion, code apply, image generation, predictive interaction, visualization analysis, and title generation, still use their standard feature-based credit rules.

Runcell converts model usage into Credits with a fixed rate:

1 USD of model usage = 100 Credits
1 Credit = 0.01 USD of model usage

How Credits Are Calculated

For each request, Runcell calculates input and output usage separately:

input_credits = input_tokens / 100,000 * input_credits_per_100k_tokens

output_credits = output_tokens / 100,000 * output_credits_per_100k_tokens

total_credits = input_credits + output_credits

For example, if a model costs 20 Credits / 100K input tokens and 80 Credits / 100K output tokens, then a request with 10,000 input tokens and 1,000 output tokens has a theoretical cost of:

10,000 / 100,000 * 20 = 2 Credits
1,000 / 100,000 * 80 = 0.8 Credits

Total = 2.8 Credits

Final deductions are rounded up to the minimum billing unit of 0.01 Credit, so very small requests may still show as 0.01 Credit.

Actual Usage Is Usually Lower

The rates below are theoretical token rates before cache optimization. In most continuous work sessions, the actual Credits used are usually much lower than the theoretical value. As a rough planning estimate, real usage is often around 1/5 to 1/4 of the theoretical token cost, though this is not a strict fixed discount.

This is because Runcell uses Prompt Cache Prefix optimization. When a conversation or agent task keeps reusing the same context, the repeated prefix can be cached and reused instead of being priced like entirely new context each time.

Model Token Credit Rates

The table below converts public OpenRouter token prices into Runcell Credits. It shows the approximate theoretical Credits used for every 1K and 100K input or output tokens before cache optimization.

Prices were checked from OpenRouter public model pricing on May 14, 2026. The 100K tokens columns are intended as a large-context reference point for comparing models.

Anthropic

OpenAI

Google

Other Providers

Why Cache Optimization Lowers Credits

AI work often happens as a continuous task rather than as isolated one-off messages. In those cases, much of the context is repeated from one model call to the next. Runcell uses Prompt Cache Prefix optimization so repeated context can be reused more efficiently.

When a cached prefix is hit, the token cost for that cached portion can theoretically be as low as about 1/10 of the normal input token cost. In real conversations and agent tasks, however, not every token is a cache hit: each new turn adds fresh context, new outputs, and sometimes new tool results that still need to be processed normally.

There can also be a higher cost when cache entries are first written. Depending on the model and cache duration, cache write cost is commonly around 1.25x normal input cost and can be higher in some cases, such as longer-lived cache entries. Runcell handles this with an internal strategy that decides when cache optimization is worthwhile.

This is especially helpful in two common scenarios:

• Multi-turn conversations: In an ongoing conversation, previous context is reused across turns. Cache hits are usually high, so the actual Credits used can be much lower than the theoretical table rate.
• Tool-heavy agent tasks: When the AI calls multiple tools during a task, the intermediate context often shares the same prefix. That repeated context can benefit from cache optimization, reducing the cost of long agent workflows.

The exact savings vary by model, context shape, and task flow. For typical continuous Runcell usage, 1/5 to 1/4 of the theoretical table cost is a practical estimate, not a guaranteed calculation rule.

Models Without Current OpenRouter Pricing

Some older or preview model names may not have a current public price in OpenRouter's model list. When a public price is not available, Runcell cannot show a stable token-to-Credits estimate in this table.

Quick Reference

Use the 1K tokens columns for small requests and the 100K tokens columns for larger context-heavy tasks. The actual Credits used by a request depend on the selected model, final token counts, and how much context can benefit from cache optimization.