The sticker price is not the bill

Fraser's week in AI tools: cheaper tiers everywhere, stolen keys as the expensive bug, and models that don't write

Three things this week, one mechanism, one table. Everything here comes from official posts, docs and public reporting. Vendor benchmark numbers are labeled as vendor-reported.

1. Every lab shipped a cheaper tier, and every one comes with fine print

The list prices moved down almost everywhere this week.

Anthropic cut Fable 5.2 to $8/$40 per million input/output tokens, down from $10/$50 (release notes, Nov 9). Google opened Gemini 4 Carbon at $2/$10, but that is an introductory rate that runs until Nov 30, after which it doubles to $4/$20 (Google's post, Nov 10). Then on Nov 12 Google added Gemini 4 Hydrogen at $1.50/$7.50, flat across a 1M-token context. Meta's Muse Ember came in at $0.10/$0.50 up to 200K tokens of context.

The fine print is the part to read. Hydrogen is cheaper than Argon, but it replaces 3.8 Flash, which VentureBeat lists at $0.75/$3.75 through Dec 31, so for Flash users it is a per-token increase. Alibaba's Qwen4-Max preview is only available through its Token Plan subscription and retires on Nov 25. Cheap tiers are real. So are the clocks attached to them.

1. The most expensive bug of the week was a leaked key

Railway's incident report (Nov 10) put fraudulent AI spend from stolen customer keys at about $740K, by its own estimate. The report also listed how each provider handled refunds, and the answers ranged from refunds already started to case by case to no refunds at all. Fireworks widened its own breach disclosure the same week. And an Apiiro study, reported Nov 7, found secrets in 1,122 of 104,185 public files tied to Claude Code's "don't ask again" setting.

The takeaway is boring and important. An API key is a prepaid card that anyone holding it can spend. Use one key per service, set a spend cap on each, and check your git excludes before you push.

1. Some of your AI calls don't need to write anything

AWS's Strands team published a case study (Nov 9) about moving internal ticket labeling from Sonnet 4.6 to a small decision model. By their own numbers, cost per ticket fell from $0.0222 to about $0.0005. Two days later Google launched the Gemini Decisions API in preview: $0.10 per million input tokens, $0 for output, and up to 1,000 yes/no, choice or score questions per request.

A decision model returns probabilities for a fixed set of options instead of generating text. If an LLM call in your product ends with "pick one of these", it is worth asking whether you are paying for prose you then throw away.

Mechanism of the week: why the price per token is not the bill

Your bill is roughly price × tokens. This week showed that the tokens side moves at least as much as the price side, and that it moves for reasons you set yourself.

Thinking counts as output. Reasoning models think before they answer, and that thinking is billed as output tokens. Alibaba's Qwen4-Max docs state that thinking and the answer share the same max_tokens budget, so if you size the budget for the answer alone, the reply gets cut off.

Effort settings multiply cost. Gemini 4 Hydrogen has four thinking levels. Artificial Analysis puts its cost per task at $0.29 on low, $0.80 on medium, $1.87 on high (the default) and $3.16 on xhigh. That is roughly an 11x spread from one setting on the same model. Defaults also differ between surfaces: Anthropic sets Fable 5.2's effort to high in the API and to medium in its apps and Claude Code.

Models differ in how much they think. Google's own post says Carbon uses more tokens than Argon. A lower price per token can still lose to a model that finishes in fewer tokens.

Prices move on a schedule. An introductory rate has an end date, and a preview has a retirement date. If you are comparing models for anything longer than a weekend project, compare the price you will be paying in December, not the one on the launch post.

So when a new model launches, the useful question is not "how much per million?" It is "how much per task, at the setting I will actually use, at the price that will apply next month?"

The table

The image above lists this week's list prices with dates and sources. Treat it as a snapshot. Two of those rows have an end date attached.

That's the week. If you have a model you moved to because of a price cut, I'd like to hear whether the bill actually followed.

Sources: Anthropic release notes and pricing page; Google blog posts on Gemini 4 Carbon, Gemini 4 Hydrogen and the Gemini Decisions API; Meta's Muse Ember post; Alibaba Cloud Model Studio's qwen4-max-preview docs; Railway's engineering incident report; Fireworks AI security bulletin; The Hacker News on the Apiiro study; Strands Agents blog; Artificial Analysis via The Decoder; VentureBeat.