Billing: Every Agent Counts

A different billing model

Remember one thing before going live: every token consumed by every agent is billed. It fits on one line; its consequences show up on your invoice. Many teams discover multi-agent through a convincing demo on a few queries, open it to their users, and see a tenfold difference from their forecasts at the end of the month. This lesson ensures you don’t join them.

What gets billed

For every request, you pay for three categories of tokens, for every agent involved—leader and sub-agents alike: input tokens, output tokens, and reasoning tokens. The third is the one you’ll forget because it’s invisible. Sub-agents’ internal reasoning is encrypted and inaccessible to you, but it still counts. Add to that tool calls made by any agent and coordination exchanges between agents: internal team discussion is work, so it’s tokens.

What this looks like in practice

With a standard request to a Grok model, you pay for one model’s tokens. With four agents, count on at least four times more, and generally more since agents interact. With sixteen agents, the gap ranges from ten to fifty times depending on topic complexity and number of searches triggered. A single request to sixteen agents equipped with active search tools easily generates tens of thousands of tokens, and far more on an investigation that casts a wide net.

Tools weigh heavily in this total, for three cumulative reasons: the request sent to the tool by the agent, the response the tool returns, and the processing of that response. Multiply by sixteen agents each performing several web searches and you see why volume climbs so fast.

Four levers to control costs

The simplest lever is also the most effective: make four agents your default, systematically. Switch to sixteen only when depth is truly at stake, which represents a minority of cases. The vast majority of questions get an excellent answer with four agents, and the user will see no difference other than a shorter delay.

Next comes prompt precision. A vague prompt scatters agents, and every unnecessarily explored path costs money. “Tell me everything about AI in Europe” sends sixteen agents in sixteen directions without stop criteria; “What are the 3 main AI regulatory frameworks in the EU voted in 2025-2026?” bounds the work and produces a more useful answer for less. Writing rigor here is an economy measure as much as a quality measure.

Third point, and it deserves blunt statement: multi-agent does not belong everywhere. It is justified when the answer requires cross-referencing multiple sources, when you need to conduct complex comparative analysis, or when you request a news synthesis on a broad topic. For a factual question with a single answer, a standard model costs a fraction of the price and answers just as well, sometimes better.

Finally, the organizational side: monitor consumption instead of discovering it. Set budget alerts on your xAI account and token limits in your application, so you are warned while spending drifts, not after.

A reminder with budget consequences

The multi-agent model is not accessible via the Chat Completions API. If your infrastructure relies on it, adopting multi-agent implies migrating to the Responses API with the OpenAI SDK, or to the xAI SDK. This migration cost is part of the equation just like token price: integrate it into your decision rather than discovering it during development.

Key takeaways

• All tokens from all agents are billed (input, output, reasoning)
• Tool calls from each agent add billed tokens
• 16 agents consume significantly more than 4 agents
• Always start with 4 agents and scale up if needed
• Reserve multi-agent for requests that truly justify it
• Chat Completions API is not supported for multi-agent