Your AI coding plan changed again. What does a shipped change actually cost?

Both major coding agents are sold as flat monthly plans with usage limits on top. Claude Code plans combine a rolling five-hour window with a weekly cap; Codex plans have published per-window allowances on lower tiers and weekly allowances on Pro. Through 2026 those limits changed several times. Anthropic was reported to raise its limits in the spring and again in September. At the end of October, third-party trackers and subscriber emails report that the included Codex usage of the $200 Pro tier is being cut roughly in half.

Every one of these changes started a new round of the same debate: which plan is the better deal? The debate usually compares sticker prices and published allowances. It rarely asks what each plan costs for the work a team actually ships.

A subscription makes AI look like a fixed cost. It is not. The real price of AI in a subscription is the plan fee divided by the work it produced. When the allowance shrinks, the same fee buys fewer sessions, and the cost of every change goes up while the invoice stays exactly the same.

• Hidden unit price

  A smaller allowance at the same fee is a price increase per session. Nothing on the invoice shows it.

• Overflow to pay-as-you-go

  When the cap is hit, work moves to credits or the API at per-token prices. That is where bills spike, and it often lands on a different invoice.

• Waiting at the cap

  An engineer who stops until the window resets is not free. Their time is usually the most expensive line in the whole change.

• Tool sprawl

  Teams hedge by paying for both tools. Two subscriptions can be sensible, but only if you know which one the shipped work came from.

Benchmarks that compare cost per task between agents are useful for API users, who pay per token. For subscribers, the marginal session is free until the cap and expensive after it, so per-task prices say little about what a team pays.

There is a second trap. In agent sessions most tokens are context re-sent on every turn, so the same task can cost very different amounts depending on how it is run. A plan with a generous allowance can still be the expensive choice if sessions drift, retry and end without a commit, and a tighter plan can be enough for a team whose sessions are short and focused.

The question to ask after every plan change is not “did our bill go up” but “did the cost of a shipped change go up”. Answering it takes a few steps:

1. 1

   Attribute usage to work

   Connect AI sessions to the commits, pull requests and deploys they led to, per project.
2. 2

   Price subscription usage honestly

   Translate subscription usage into an amount from the provider’s price list, and label it as an estimate. It is not what you were billed, and it should not pretend to be.
3. 3

   Add the overflow

   Credits and API spend that started when a cap was hit belong to the same changes, even when they arrive on another invoice.
4. 4

   Count what never shipped

   Sessions without a commit and pull requests closed unmerged consume allowance too. They are often the reason a team hits the cap.
5. 5

   Re-run after every change

   When a provider changes limits, compare the cost per change before and after. That is the number that tells you whether to switch plans.

Nythrex follows this approach. Every change gets its cost from AI session to production; AI on a subscription is labelled as an estimate from the price list, never as a reported fact; and a total with a missing price is marked incomplete instead of being treated as zero. Claude Code and Codex cost tracking · How the chain works.

Providers are still learning how to price agents that can run for hours, and limits will keep moving in both directions. Arguing about which plan is the best deal this month is a game you cannot win. Knowing what your shipped work costs on each plan is a question you can answer, and the answer stays useful whatever the next announcement says.

• OpenAI Help Center — Using Codex with your ChatGPT plan
• Anthropic Help Center — About Claude’s Pro plan usage
• BenchLM — Codex usage limits, plans and public changes
• Morph — Claude Code vs Codex pricing and limits (September 2026)
• Vantage — The hidden cost driver in agentic coding sessions (April 2026)