Ditto AI (SN118) Uses Cascade (SN91) to Forecast AI Credit Usage and Recommend Plans

AI credits can disappear faster than expected. One day your usage looks manageable, then a few heavy sessions leave you wondering whether your balance will last until the next reset. 

Checking your current balance tells you what remains, but not what you are likely to need.

Ditto AI (SN118) is addressing that problem with a new in-chat credit forecasting feature that analyses past usage, estimates future consumption, and recommends a suitable plan.

From tracking credits to predicting costs

Users can ask questions directly in Ditto’s chat interface, such as the following;

1) where their credits have been going.

2) whether their current pace will exhaust their balance before the next reset.

3) which plan best fits their needs.

Ditto combines billing history with information about the user’s requirements to estimate future consumption and suggest suitable plans. 

The chat provides a short summary of recent activity, projected usage and plan options, while detailed forecasts, model information, and pricing comparisons are available in expanded results.

Users can access these insights without building spreadsheets or opening a separate forecasting tool.

How Bittensor Subnet 91 powers the forecast

The feature uses Ephemeris, a forecasting API developed by Cascade, which operates Bittensor Subnet 91 (SN91).

Ephemeris provides access to multiple time-series forecasting models through a single API, similar to how OpenRouter gives developers access to multiple language models.

Ditto specifically requests Chronos-2, an open time-series foundation model developed by Amazon. Instead of processing words, it analyses numerical sequences to estimate how values may change over time.

Here is how the system works.

1) Usage data is prepared: Ditto converts billing records into time-based usage data, starting with hourly intervals. These intervals widen as more history becomes available, and past usage is recalculated at current prices so forecasts reflect current rates.

2) Estimates future consumption: The model produces a range of possible outcomes, including lower (p10), median (p50) and upper estimates (p90). This helps Ditto account for uncertainty instead of presenting one projection as guaranteed.

3) Ditto checks the results: It tests the model against the user’s past usage and compares its performance with a simple estimate based on recent spending. Chronos-2 is used only when it clearly performs better and meets stricter requirements for predicting when credits could run out. Accounts with insufficient history fall back to a simpler estimate until there is enough data.

Privacy and plan recommendations

Forecasting requires usage data, but Ditto says it sends only anonymised numerical sequences to Ephemeris while prompts, memories and account details stay private.

The feature also helps users choose the cheapest plan that meets their stated requirements and estimated 30-day usage. Ditto provides a recommendation, while the final subscription choice is completed through Stripe.

A few limitations are worth knowing.

1) Usage analysis is available on every plan, but full forecasts require sufficient usage history and available capacity within a shared daily budget.

2) Forecast results are cached for the remainder of the UTC day so repeated requests can reuse existing results.

3) Bring-your-own-key (BYOK) usage is tracked separately.

Users can read the full explanation in Ditto’s blog post.

Why this is useful for Bittensor

Ditto is an AI assistant built around persistent memory across apps and chats, including Telegram, Slack, Discord and MCP clients.

By integrating Cascade’s forecasting service, it can offer a specialised capability without building every part of the forecasting stack itself.

This is a practical example of Bittensor subnets serving different roles in a wider product ecosystem. Cascade’s SN91 supplies forecasting infrastructure, while Ditto uses it to solve a specific problem for its users.

As more subnet projects connect their services, the network’s value can extend beyond individual subnets to the products built with them.

➛ Users can explore Ditto’s features and plans at heyditto.ai and view its pricing.