Salesforce already sells extra API capacity for integration traffic at roughly $83 per million calls. The agent meter it is rolling out lands somewhere between $5,000 and $100,000 per million — a 60x to 1,200x multiple on the same endpoint, same record, same write, with the only variable being whether a human's integration or an agent made the call. That arithmetic, drawn from published pricing, frames a wave of agent-access price increases now hitting almost every pre-AI B2B vendor.

The two most visible movers are taking different routes. HubSpot's increases target its own agents: Breeze credits, per-resolution pricing, and custom agents metering since July. Its MCP server remains free for agents customers bring. Salesforce meters third-party agents instead — every successful call an agent makes through MCP or the API becomes a Flex Credit charge, agents must be registered, and existing customers migrate to the new billing at renewal.

The pattern extends beyond the majors. The source, a SaaStr writer running agent infrastructure, reports a niche CRM used for five-plus years now demanding extra payment for agent use of its API, and another vendor deprecating the API their AI VP of Marketing, "10K," runs on — with no replacement and no meaningful notice. The consistent divide: vendors whose pricing was built for agents from the start aren't adding meters. Vendors whose pricing assumed humans clicking a UI, with those humans logging in less, are moving the meter to the API call.

The workaround is cheap

The measured counter-response is architectural, not adversarial. When a vendor announces agent API charges, the buyer's agents route around it: sync records to an internal database, read from the copy, write back to the vendor only when something changes. Agents read far more than they write, so most meterable calls are lookups against data the customer already holds. Moving a database used to be a hard project; the source calls it "a weekend project now."

Buyers also change procurement. "How does this price agent access" is now a real evaluation question, and a bad answer is disqualifying.

The spiral mechanics

The asserted risk for vendors: metered access reduces usage, new agents get built against other systems from day one, less work happens on the platform, and the system-of-record moat — never the software itself, but the fact that everything touched it — erodes. That leads to a renewal conversation where the vendor holds a smaller footprint and the customer holds more options than the year before. This is the author's argued thesis, not a measured outcome; the price increases are new enough that no churn data exists yet.

What the infrastructure market charges

The cost-gap comparison is concrete. Firebase has charged $0.06 per 100,000 document reads for years without anyone calling it a tax — per-call pricing is not itself the issue. On storage, Salesforce extra data storage runs about $125 per month per 500MB, roughly $3,000 per GB-year, while Neon charges about $0.35 per GB-month, or $4.20 per GB-year. Several hundred times more, for the same bytes on comparable hardware. The gap is wide enough that agents bypass it by default. As the source puts it, when one lookup costs thousands of times more through the vendor than through a copy, "the agent isn't making a hard call. It's making an obvious one."

Atlassian shows a workable middle path with Rovo credits. Overage billing starts December 3, 2026, at $0.01 per credit, with a basic action costing 10 credits — $0.10, the same rate as an Agentforce action. Two design choices soften it: every paid plan includes a pooled allowance (25 credits per user per month on Standard, 70 on Premium, 150 on Enterprise), and reads don't draw credits today. The allowance is thin and overage billing is on by default unless an admin caps it, but Atlassian published the rate, the date, and gave admins a switch.

The source's proposed fixes for vendors: publish and cap the rate so buyers can budget, let a registered agent replace a seat rather than stack on one, or price outcomes instead of calls — the model some of these vendors already use for their own agents. The problem isn't the meter; it's the additive stack of seat, storage premium, API tiers, and a per-call charge — three bills for one piece of work.

Incumbent stacks stay, at least for now: ten years of context lives there, and that context is what makes the agents good. But 10K alone runs 30,000+ API calls a day, and no vendor charging a significant tax on that volume will make the next evaluation shortlist.