Agent budgets: your guardrail can lie

On September 10, 2026, my budget guardrail for coding agents still allowed GPT 5.6 Luna up to $6 over five hours. The published limit for that model was $3. The server could cut off the task before the local warning arrived. The script ran; its protection depended on a wrong value.

The same check found grok-4.5 removed from the catalog while it still occupied the deep-review role. Direct calls returned 400 Model is unavailable. An available budget for an unavailable model cannot run a review.

These personal observations date from September 4 and 10, 2026, with omp v18.1.17 for the September 10 reading. They were not repeated on October 8. The public Go documentation changes over time: the historical amounts below are not a current price list.

Which amount does your counter track?

Quota dollars value consumed tokens at catalog rates. They do not tell you the subscription bill. A $3 threshold over five hours describes a usage allowance; it does not mean an extra three dollars will be charged.

The Go documentation publishes monthly limits per model, with windows of 20% over five hours, 50% over a week and 100% over the month. In the September 10 reading, a $15 monthly limit therefore meant $3 over five hours. The old local aggregate of $12/5 h was no longer enough to decide whether Luna could start.

What it costs

Prices checked on September 10, 2026

Luna: old local threshold Incorrect historical threshold, expressed as API-valued usage.

$6 / 5 h

Luna: observed published limit 20% of a $15 monthly quota. Neither amount is the subscription price.

$3 / 5 h

The price catalog had drifted too. For DeepSeek V4 Pro, my estimate used $0.435 for input and $0.87 for output per million tokens, which were MiMo prices. The September 10 reading showed $0.66 and $1.98 off-peak. A correctly filtered window with the wrong price still produces a wrong inference cost calculation.

Server windows and local detail

By September 10, omp usage exposed server windows, with usedFraction and resetsAt. The public command in omp v18.1.17 also supports JSON output:

omp usage
omp usage --json

The reading showed 0% over five hours, 1% over seven days and 44% over the month. My local scan found only $5.89 for the cycle from August 26 to September 10. Sessions from another machine and calls that were not persisted escaped that scan. Only server telemetry gave the account’s actual global usage.

My personal model-budget script remained useful for another purpose: attributing consumption to models in the pi and omp sessions present on the machine, then estimating a review. Its detail is local and its estimate depends on the recorded prices. It cannot reconstruct missing consumption or replace server windows.

A useful green light carries a date

The proposed method is to reconcile these readings before a long task. Start with the server window and its reset, then check the model’s limit in the documentation. Finally, compare that limit with the locally attributable consumption and the task estimate. A discrepancy leaves the available budget uncertain; it calls for a manual decision.

In my personal tooling, these existing commands provide that last reading:

model-budget report
model-budget check gpt-5.6-luna
model-budget estimate --depth deep

These commands describe the script observed in September, not a native Go feature. They do not automate server reconciliation. To keep your own guardrail auditable, store the catalog date and harness version alongside its thresholds. After a model change, also check routing and fallback chains: replacing Grok involved several entries.

What remained uncertain

Reviewing the guardrail itself had uncovered counting bugs, a stale cache and an unreachable warning band. Five rounds on August 8 represented roughly $1 in API-valued usage:

What it costs

Prices checked on August 8, 2026

Guardrail review, five rounds Documented historical consumption. This amount is neither an extra bill nor a guaranteed review price.

~$1 in API-valued usage

Guidelines vs guardrails argues for enforced constraints. Here, their effectiveness also depends on the data feeding them. The harness is a budget line already covers upkeep; the issue here is whether the green light is valid.

The layering guide remains useful for assigning roles, with its July figures treated as historical. For cost decisions, the optimization guide complements the method. No post-recalibration outcome is measured here. You can start by dating your thresholds and comparing a planned task with the server window: if the readings diverge, your local counter should remain an estimate.