ai pricing is switching to usage credits. here is how to not get burned

the shift that hit this month

On July 8, Anthropic's top-tier Fable 5 model moved to usage-credit billing for all subscribers: reportedly $10 per million input tokens and $50 per million output tokens. One analysis worked out that a medium-complexity code refactor, around two million tokens, costs about $20 in output credits alone.

This is not one company being greedy. It is the whole industry converging on the same truth: agentic AI consumes wildly different amounts of compute per user, and flat subscriptions cannot survive users who run agents all day. Metered billing is coming for every serious AI tool. Here is how to stay ahead of it.

step 1: learn to estimate tokens

Tokens are the meter. The rough math that covers most cases:

• 1,000 tokens is about 750 words of English text.
• Code is denser: a medium source file often runs 1,000 to 3,000 tokens.
• Agents multiply everything, because every step re-reads context. A "simple" agent task can quietly consume 50 times the tokens of the question that started it.

The habit to build: before a heavy job, ask what the model will read (input) and write (output), in rough thousands of tokens. You only need to be right within 2x for budgeting to work.

step 2: know the current price spread

The per-million-token prices in play right now, from the cheap end up:

• Z.ai's ZCode (open-weight, GLM-5.2): about $1.40 input / $4.40 output.
• Claude Sonnet 5: $2 / $10 intro until August 31, then $3 / $15.
• Fable 5 class: $10 / $50.

That is a 10x-plus spread between tiers. The single highest-leverage cost decision you can make is matching the model tier to the task instead of sending everything to the most powerful model out of habit.

step 3: route by difficulty

The pattern every efficient team lands on:

1. Default cheap. Summaries, drafts, formatting, extraction: bottom tier. It is more capable than you think.
2. Escalate on failure. If the cheap model's output fails your check, rerun on the tier above. Escalating 20 percent of tasks costs far less than premium-by-default.
3. Reserve the top tier for work whose failure is expensive: architecture decisions, complex refactors, final passes.

step 4: put a number on it before it puts one on you

• Set billing alerts at a threshold that would annoy you, not bankrupt you. Every serious platform supports them.
• For agents, cap steps and context size. Runaway loops are where surprise bills are born.
• Re-estimate monthly. Prices moved three times while I drafted this post. The intro discounts, like Sonnet 5's until August 31, are real windows worth timing heavy work around.

the honest framing

Usage pricing sounds worse than flat fees, but it is mostly honest pricing: the all-you-can-eat era was subsidized, and subsidies end. The people who get burned are the ones who never look at the meter. Now you know where the meter is. Look at it, route by difficulty, and the same budget that bought you a subscription last year can run a small fleet of agents this year.

You're welcome.