Pricing

Vellum uses a prepaid credit balance: usage is deducted from your credits as you use the assistant. You can add credits anytime from the Billing page via Stripe Checkout. Applicable taxes may be added during checkout.

In the app, your Billing screen shows a Credit Balance plus a breakdown of settled and pending amounts:

• Credit Balance: the current amount available after pending compute is considered.
• Settled Balance: charges that have already settled.
• Pending Usage: estimated in-flight compute that may not be fully settled yet.

The purchase and use of Vellum Credits is governed by Section 6 of our Terms of Service.

Vellum makes available certain features and functionalities within the Services, as designated by Vellum from time to time, that are accessible exclusively through the use of prepaid credits (“Vellum Credits”). These designated features and functionalities are referred to as “Credit-Eligible Features.” Credit-Eligible Features currently include inference, web search, image creation, and paid third-party APIs accessed through Vellum's managed OAuth (for example, Twitter). No alternative direct-payment method is available for Credit-Eligible Features.

You may fund your Vellum Credit balance (“Vellum Balance”) by purchasing Vellum Credits in ten-dollar ($10.00 USD) increments, up to one hundred dollars ($100.00 USD) per top-up, through the payment methods made available in the Services, or at such other amounts as determined by Vellum from time to time.

Four categories of work consume credits today: LLM inference (the biggest line item by far), web search, image generation, and paid third-party APIs you reach through Vellum's managed OAuth (for example, Twitter).

Inference is itself broken into a set of Actions you'll see attributed in your usage dashboard. Here's what each one does:

Conversation with your assistant

• Main agent. Your assistant's response when you chat with them. The biggest chunk for most people when actively using the app.
• Inference. One-off model calls from skills or utilities that don't fit a more specific category.

Memory subsystem (mostly background)

• Memory consolidation. Promoting short-term observations into long-term memory pages.
• Memory extraction. Pulling concrete facts, preferences, and entities out of a conversation so they can be remembered.
• Memory retrieval. Looking up relevant memories when you ask a question or start a task.
• Recall. Targeted, deeper memory lookups across notes, knowledge base, and past conversations.

Conversation polish (background)

• Conversation summarization. Summarizing finished or long conversations so your assistant can refer back to them efficiently.
• Conversation title. Auto-generating a short title for each new conversation.
• Conversation starters. Suggested prompts the app surfaces when idle.
• Empty-state greeting. The hello your assistant shows when you open the app with no active conversation.
• Context compactor. Shrinking a long conversation's context so it still fits in the model's window without dropping anything important.

Autonomy (background)

• Heartbeat agent. Periodic check-ins where your assistant reflects, plans, and decides whether anything needs your attention.
• Filing agent. Filing notes, decisions, and learnings into your personal knowledge base.
• Notification decision. Deciding whether to push you a notification or stay quiet.

Other

• Unknown Task. LLM calls that haven't been tagged with a specific subsystem yet. We're cleaning up the remaining attribution gaps.

A lot of background work is configurable. You can ask your assistant to disable or reduce the frequency of “heartbeats” and “memory compaction,” or to use a less expensive model for these actions. We're actively working on making background spend more visible and easier to control.