Scale · founder · 8 min read

Two AI Coding Plans Shrank This Month. Neither Price Moved.

Claude Code lost 17% of its weekly limits on Sep 14. Codex loses GPT-5.5 on Oct 14. Same invoice, less product — and one clear pattern behind both.

Published September 18, 2026 ·

pricing budget claude code codex models

Your AI coding bill this month is the same as last month. What you get for it is not.

Two things happened in the last three weeks, from two different vendors, in two different products. Neither one changed a sticker price. Both made the thing you are paying for smaller. And if you only watch prices — which is what most people watch — you would not have noticed either.

What changed, in order

September 14: Claude Code weekly limits dropped 17%.

Anthropic announced on X in late August that it was “permanently raising standard weekly limits in Claude Code by 25%” for Pro, Max, Team, and seat-based Enterprise plans, effective September 14.

That sentence is true. It is also not what happened to you.

Since May, Claude Code had been running a temporary 50% boost on those same weekly limits. The boost expired September 13. The permanent 25% increase started September 14. So if your baseline allowance was 100 units, you had 150 the day before and 125 the day after. That is 25% more than the pre-promotion floor and 17% less than what you actually had.

Developers said so, loudly. Anthropic deleted the original thread and reposted with the number in plain language: “Compared to today, this works out to a 17% reduction in weekly limits on Claude Code.” BleepingComputer covered the sequence on August 29.

Credit where it is due: Anthropic corrected itself within hours, in public, with the unflattering number spelled out. That is a better response than most vendors on this site have managed. It does not change the arithmetic.

October 14: Codex loses GPT-5.5.

OpenAI announced on September 15 that GPT-5.5 retires from ChatGPT, ChatGPT Work, and Codex on October 14, across every plan — consumer, Business, Enterprise, Edu. If you use GPT-5.5 in Codex, you move to GPT-5.6 Sol or GPT-6 Astra.

GPT-5.5 launched April 23, 2026. It is being retired from the product surface after roughly six months. The GPT-5.4 class was pulled from Codex at the end of August. This is now a rhythm, not an event.

The part worth actually remembering

Read the OpenAI announcement again and notice the carve-out: the retirement does not apply to the API. Anything authenticated with an API key keeps working.

That is the whole lesson, and it is more useful than either individual story.

The subscription is the surface that churns. The API is the surface that doesn’t. OpenAI will pull a model out from under your Codex workflow with 29 days’ notice while leaving the same model available to anyone calling it with a key. Anthropic will re-baseline your weekly allowance while its per-token API prices sit unchanged.

This is not vendors being sneaky, exactly. It is a structural difference in what you bought. An API price is a published rate against a metered unit — changing it is a price change, it is legible, and customers scream. A subscription is a bundle, and the contents of a bundle can be adjusted without touching the number on the invoice. Included credits, weekly caps, default effort levels, which models are in the box: all of it moves, none of it is a “price increase.”

We have been making this argument on this site since June. The pricing calendar exists because the advertised price was identical before and after every single scheduled change we could find. The June billing reset and the Copilot credit cliff were the same shape. September added two more, from the two vendors most likely to be in your stack.

What this costs you, concretely

For Claude Code, the honest answer is: we can’t tell you, and neither can Anthropic.

Weekly limits are not measured in prompts. Anthropic says usage varies by conversation length, model choice, tool usage, and effort level. A week of short, cheap edits and a week of long agentic runs on Opus draw down the same allowance at wildly different rates. So “17% less” translates to somewhere between “you’ll never notice” and “you hit the wall on Thursday instead of Sunday,” depending entirely on how you work.

The practical read: if you were already bumping into weekly limits, you now hit them about a day and a half earlier in the week. If you weren’t close, this is invisible. There is no middle ground where you should care mildly.

For Codex, the cost is migration labour, and it is small but not zero. Before October 14, check:

• Workspace defaults and saved model settings
• Managed configurations, if your org has them
• Custom agents
• Scheduled tasks — these fail silently and nobody notices for a week
• Any script or automation with gpt-5.5 hardcoded in it

That last one is where the actual breakage lives. A hardcoded model string in a CI job or a cron script does not throw a helpful error; it throws whatever the platform throws when a model name stops resolving, at 3am, in a log nobody reads.

Five things to do this month

1. Grep your repos and automations for gpt-5.5. Do it today, not on October 13. It takes four minutes and it is the only item here with a hard deadline.

2. Pick your Codex replacement deliberately. GPT-5.6 Sol and GPT-6 Astra are not interchangeable — different cost profiles, different behaviour on long agentic runs. If you have prompts tuned against 5.5, budget an afternoon to re-check them rather than assuming a clean swap. See our take on GPT-6 Astra and GPT-5.6 Sol.

3. Find out where your Claude Code week actually ends. Not the limit number — the day. Run a normal week and note when you get throttled. That single observation is worth more than any published allowance figure.

4. Stop tracking sticker prices and start tracking your own unit. Cost per shipped feature, cost per working week, whatever maps to your business. Every change on this page is invisible to a price tracker and obvious to a per-output tracker.

5. If model stability is load-bearing for you, move that workload to the API. This is the real takeaway. A scheduled job, a production pipeline, anything where a model disappearing breaks a customer promise — that belongs behind an API key, not behind a ChatGPT login. You will pay per token instead of a flat rate, and you will stop being re-platformed on someone else’s schedule.

What we are not saying

We are not saying either vendor did something underhanded. Anthropic ran a promotion, ended the promotion, and left customers better off than before it started — that is a normal thing to do, and the only real misstep was the framing, which it fixed the same day. OpenAI gave 29 days’ notice, named the replacements, and explicitly protected API users. Compared to Google retiring Gemini CLI outright or Microsoft’s silence on Copilot security, both of these are well-behaved.

We are also not saying you should panic-migrate. If Claude Code’s limits were never your constraint, this changed nothing for you. If you don’t use GPT-5.5 in Codex, October 14 is not your problem.

What we are saying is that the invoice stopped being a useful signal, and it stopped a while ago. Two vendors, three weeks, zero price changes, and both products got smaller. If your tool budget is a spreadsheet of monthly prices, that spreadsheet has been telling you the same comfortable number all year while the thing it describes quietly changed underneath it.

Track the bundle, not the bill.