Everyone Shipped Agent Cost Control in the Same 72 Hours

August 6: Sapiom raised $35M for a router that sends each model call to the cheapest capable model. August 7: Databricks published its internal cost playbook. August 8: Toolport hit Product Hunt with a free MCP gateway cutting tool-definition overhead 96%. Three parties, zero overlap, one premise.

The Databricks post is the substantive one because it names numbers and adopters. Patrick Wendell, Erich Elsen and team argue agentic coding spend is on a curve that "will eventually overtake revenue," and they lay out four levers converged on by early large-scale adopters including Stripe, Coinbase, Uber and Ramp. Moving to open-source or lower-cost models is the single greatest lever. Dynamic request and task routing cuts average task cost more than 30%. Developer-visible tripwires and budgets with progressive friction, not hard caps. And token overhead reduction, where Databricks cut generated tokens by almost 50% through harness tuning alone. Both infrastructure pieces are free: the Omnigent meta-harness and Unity AI Gateway. Their stated goal is a "dual mandate," broad low-friction access with a roughly fixed cost envelope per user.

Fifty percent fewer generated tokens from tuning the harness. Not the model, not the prompt. The scaffold.

Sapiom's number is the loud one: it reportedly took agent company Polsia's monthly Anthropic bill from about $1.2M to about $100K. That's model arbitrage as a venture-scale business, and Anthropic itself participated in the round alongside Dragonfly, Accel, Gradient, Coinbase Ventures, Okta Ventures, Menlo Ventures and VanEck. 270M+ transactions across 100,000+ daily agents in six months.

Toolport is the one I'd install today. Free, open source, local-first, it puts every MCP server behind one port and exposes three meta-tools the agent searches on demand instead of loading every schema up front. Claimed results: roughly 900 tokens of tool overhead versus the usual load, 96% less tool-definition overhead per request, up to 91% fewer total tokens at the same task success, and 99.5% reduction on a real 415-tool catalog. Each MCP tool definition runs 200 to 500 tokens. Five servers routinely burn 30,000 to 60,000 tokens before the user types anything.

Meanwhile Anthropic made spend a protocol concept. The August 7 platform release notes add session budgets as a hard cap at public list rates, with a session that hits the cap pausing under a new budget_reached stop reason and resuming if the budget is raised. Deployments propagate the budget to every session they start. Same week: Claude Code 2.1.225 names the gateway spend cap and its reset time in limit messages instead of just failing.

A year ago token spend was a billing-page afterthought. It's now a stop reason in the API.