OpenAI ships hard budget caps for agent spending in Codex
Token budgets that abort the turn when they run out, and delegation controls that decide when an agent may spawn more agents. The bill-shock era is getting its first circuit breakers.
- Codex 0.144.0 adds configurable rollout token budgets that track usage across agent threads, warn on remaining budget, and abort turns when the budget is exhausted.
- Multi-agent delegation is now controllable per thread and per turn: disabled, on explicit request only, or proactive.
- Usage-limit reset credits get formal structure, with type, expiration, and redemption status.
The most requested feature in agent procurement is not a capability, it is a brake. OpenAI shipped one: Codex 0.144.0 adds configurable token budgets that track usage across agent threads, remind the agent what remains, and abort the turn outright when the budget is exhausted, per the Codex changelog.
Budgets, credits, and a leash for delegation
The release also formalizes usage-limit reset credits, each with a type, expiration, and redemption status, and adds delegation controls that decide when an agent may hand work to subagents: disabled, only on explicit request, or proactive, settable per thread and per turn, per OpenAI's subagent documentation. That last control matters because delegation is where spend multiplies. A single turn that fans out into parallel subagents is a cost event, not a convenience, and analysts covering the release described the multi-agent reasoning mode as a higher-cost lever that should be budgeted separately from per-token pricing.
FinOps catches up to the agent era
Runaway token bills are the number-one objection in agent purchasing right now, and the anecdotes are not subtle: developers have reported agent-tool bills jumping from tens to thousands of dollars a month. We wrote about the economics underneath that in the 25x markup. Hard caps change the negotiation: a CFO who could not previously bound worst-case spend on a multi-agent workflow now has a config file to point at. Coming the same week OpenAI built subagent orchestration into its API, the pairing is coherent. First you sell the accelerator, then you sell the brakes, and both invoices go to the same buyer.