← All checklists

Checklist 10 of 10 · October 2026

Agent budget checklist

A launch price that doubles on 1 December 2026, and the five numbers to set before it does

A frontier model went on sale on 13 October 2026 at two dollars per million input tokens and ten per million output, for seven weeks. From 1 December the same tokens cost four and twenty. Output is five times input and includes the reasoning nobody reads. A pilot priced on the launch rate meets a December invoice at twice the price. Five checks.

The checks

Everything starts ticked. Untick what you will not do; the choice is saved on this device only, and Print keeps only what you kept.

Why. The discount has an end date; the budget line does not. A cost model built on the launch rate is wrong by a factor of two from the first of December.

The check.

  • Put the list price in the cost model today, with the discount's end date on the ticket that owns the line.
  • Set the spend alert at the list price now, so the alarm is tested while the bill is still small.
price_per_million: {input: 4.00, output: 20.00, cached: 0.20}   # list price
alert_usd_per_day: 120   # list-price projection, from today

Source: trade-press report of the launch pricing, 13 Oct 2026

Why. Agents are output machines, and output costs five times input. A request cap written when the ceiling was sixty-four thousand tokens does nothing against a million-token task.

The check.

  • Size max output tokens to the job, per route, and log every request that hits the ceiling.
  • A million-token task is a batch job with its own approval, at half the standard rate.
  • Count ceiling hits every day; a rising count is a prompt problem before it is a budget problem.
routes: {support-answer: {max_output_tokens: 4000}, code-migration: {max_output_tokens: 64000, tier: batch}}

Source: trade-press report of the launch pricing, 13 Oct 2026 (the 64,000-token ceiling raised to a million); the full price table, 14 Oct 2026 (batch at half the standard rate)

Why. Cached input is ninety-five percent off: at the list price, twenty cents per million against four dollars. The order of the prompt decides the bill more than the model does.

The check.

  • Stable content first: system prompt, tool schemas, reference documents; the user's turn last.
  • Put the cache hit rate per route next to latency and error rate; a falling rate pages someone before the invoice does.
  • Treat a prompt change that breaks the cache as a cost regression in review.
cache_hit_rate{route="support-answer"} < 0.8 for 30m   # page: prefix changed

Source: the full price table, 14 Oct 2026 (cached input $0.10 now, $0.20 from 1 December)

Why. The output rate includes the model's reasoning; the published price table says so in as many words. The answer on screen is a fraction of what was billed.

The check.

  • Log output tokens per task, split into visible answer and reasoning, and watch the ratio per route.
  • Set the reasoning effort per route; the cheapest setting that passes the eval is the default.

Source: the full price table, 14 Oct 2026 (output prices include thinking tokens)

Why. A price step with a known date is the easiest incident to prepare for. The launch rate runs to 30 November 2026 and the list rate applies from 1 December; both dates were published on launch day.

The check.

  • Run one day of traffic through the cost model at the list price, compare it with the alert thresholds, and name who approves the line after the step.

Source: trade-press report of the launch pricing, 13 Oct 2026; the vendor's launch post, 13 Oct 2026

Prices are the vendor's published standard tier: $2 input / $10 output per million tokens to 30 November 2026, $4 / $20 from 1 December, cached input at 95 percent off, batch at half the standard rate. The thresholds in the code blocks are illustrative; set yours from your own traffic.

Take this list with youThe PDF is the handout: this list on one or two pages. The Markdown is for your own docs, your wiki, or an assistant you want to hand the list to. Print gives you this list only, without the navigation.

Something in your production AI broke in a way that is not on this list? Tell me what broke.

New checklists, as they are published

One list at a time, each built the same way: a dated failure, the control that stops it, and the primary source so you can check it. Nothing else goes to this address, no digest and no newsletter.