ChatGPT Pro and Claude Code limits shrank while API prices fell

Last verified: September 30, 2026, against OpenAI's and Anthropic's own pages and the reporting linked below.

Fair warning, we build a local-first BYOLLM tool, so we are not neutral. We will keep the facts sourced and mark what nobody has confirmed. Claude Code and Codex are capable tools, and plenty of people should keep their plan. The point is narrower: a subscription allowance is something the vendor can shrink at the same price, and this month two vendors did.

What exactly changed, and when?

Anthropic announced a Claude Code cut on August 29 that took effect September 14, and OpenAI's lower Pro 200 allowance applies to new subscribers now and to grandfathered ones after October 29. OpenAI also added a $500 tier. Here is the dated version.

OpenAI now lists three Pro plans: Pro 100 at $100 a month, Pro 200 at $200, and Pro 500 at $500. Pro 500 is the only tier with Ultrafast, and OpenAI puts its allowance at 25 times ChatGPT Plus. OpenAI does not publish a multiplier for Pro 100 or Pro 200 anywhere we could find.

How big was the Claude Code cut, really?

17%, measured against the boosted limit, which means limits are still above where they started. Index the pre-boost weekly limit at 100. The temporary boost took it to 150 (a 50% increase) from May through September 13. On September 14 a permanent 25% increase over the pre-boost level took effect, so the limit is 125.

125 divided by 150 is 0.833, a 16.7% drop from what people had been getting. That matches Anthropic's own line that this works out to a 17% reduction in weekly limits on Claude Code. It also means you have 25% more than before the boost. Both statements are true, and which one you feel depends on when you started using it. Anthropic cited difficulty securing compute to sustain the higher limits, and said it would add more visibility and control over usage.

What changed on ChatGPT Pro 200?

The price stayed at $200 a month and the allowance went down. OpenAI says new Pro 200 subscriptions include “a lower usage allowance than previously offered,” reflecting more efficient models. Existing subscribers who were active at the eligibility cutoff, or in the seven days before it, keep the old allowance through October 29, 2026, as long as the subscription stays active, then move to the lower one. Only affected users get an email.

How much lower? OpenAI has not published the number. Some coverage reports the allowance was roughly halved. We have not verified that, so treat it as a report, not a fact. On Hacker News, the Pro 500 thread drew about 250 comments, with commenters debating whether the lower allowance at the same price was fair, and whether $6,000 a year justifies buying local GPUs instead.

Two details matter if you are deciding what to do. Keeping the old Pro 200 allowance does not give you Pro 500 features or Ultrafast. And buying credits on Pro 100 or Pro 200 does not unlock Ultrafast either. Credits only let you keep going past your included allowance.

What should Pro 200 subscribers do before October 29?

Measure your real usage first, price it at API list rates, and only then decide. A short checklist:

1. Check your email. OpenAI says only affected subscribers get one, with the details.
2. Look at your usage history. Note how much of your allowance you actually burn in a normal week and in your heaviest week.
3. Price that usage at list API rates. The next two sections have the per-token numbers.
4. Do not upgrade to Pro 500 reflexively. Pay $500 only if you are hitting the wall on the lower allowance and the extra usage or Ultrafast is worth it to you.
5. Try the alternatives before the date, not after. A second tool set up in advance costs you an afternoon. Switching under deadline pressure costs more.

Did API prices go up or down in the same weeks?

Down or flat on the three frontier models that shipped this month. Per million tokens, from the vendors' own pages:

Sonnet 5.5's “up to 30% less per task” and Opus 5.5's “40% less than Opus 5 on typical workloads” are Anthropic's claims about token efficiency, not price cuts. Test them on your code before you count on them. The list prices themselves are not claims. They are on the pages.

Is pay-per-token actually cheaper?

Sometimes. It depends entirely on how much you use, and heavy agentic users can blow past $200. We are not going to invent a typical usage number. Here is the reasoning that holds up with only the published prices.

• The ceiling is real. $200 of GPT-6.1 Sol output at $10 per million tokens is 20 million output tokens. Input at $2 per million is 100 million tokens for the same $200, and more than that when cached input ($0.10 on Sol) kicks in. That is a lot of chat and edits for one person.
• Agents change the math. Autonomous coding loops re-read context on every step. Input tokens pile up fast, and long agent runs are where a flat plan wins. Our agentic cost wall post walks through why.
• Caching helps a lot. Cached input is $0.10 on Sol and $0.20 on Sonnet 5.5 against $2 uncached. A well-behaved agent that reuses context pays far less than the headline input price.
• Flat plans reward heavy use and punish light use. If you rarely hit your allowance, you are paying $200 for capacity you do not use, and per-token may well be cheaper. Price your actual token usage to find out.

The honest answer: if you were near your limit every week, run the numbers, because per-token may cost you more than $200. If you were nowhere near it, per-token may cost less, but only your own token counts can tell you. The cost calculator post and the earlier Copilot credits piece show the same pattern from other angles. Either way, per-token has one property a subscription does not: the rate is published and you see exactly what you use. This month those rates went down or held flat.

Where does Bodega One Code fit?

It moves the pricing decision to you: bring your own key, or run a model locally. Bodega One Code is a local-first coding IDE and CLI with BYOLLM support across 10+ LLM provider presets, including OpenAI, Anthropic, DeepSeek, and OpenRouter, plus local runtimes like Ollama and LM Studio.

• Bring the key, pay list price. No allowance to run out of, no multiplier to decode. You pay the provider's per-token rate.
• Switch providers per task. Use a cheap model for routine edits and a stronger one for the hard problem. When a vendor reprices, you change a dropdown, not your workflow.
• Run local for no per-token cost. Qwen3.8-27B is Apache 2.0 with a 262K context window. The Q4_K_M quant is about 17 GB and fits a 24 GB GPU at moderate context lengths. See our local LLM guide.

Caveats, since you should have them before you download anything:

• Local open models are not Opus or Astra class on the hardest tasks. Expect to route the difficult work to a cloud model.
• You need the hardware. A ~24 GB GPU for a 27B model at that quant and moderate context, plus electricity.
• Cloud models still bill at your provider's rates, and your prompts go to that provider.
• Bodega One Code is free for everyone during the open beta, commercial use included. It is a beta and it has bugs.

None of this says cancel your subscription. It says keep a second door open. If your workflow only runs inside one vendor's plan, that vendor sets your price and your limits. If it runs on keys and models you can swap, a cut like this one is an inconvenience instead of a deadline.

Sources

• @ClaudeDevs on X, August 29, 2026 (17% reduction in Claude Code weekly limits): x.com/ClaudeDevs/status/2093742322525810912
• BleepingComputer, Anthropic is cutting Claude Code's current weekly limits by 17 percent: bleepingcomputer.com/news/artificial-intelligence/anthropic-is-cutting-claude-codes-current-weekly-limits-by-17-percent
• OpenAI Help Center, About ChatGPT Pro tiers: help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers
• Hacker News, ChatGPT Pro 500 discussion: news.ycombinator.com/item?id=49896975
• OpenAI, DevDay 2026 recap: openai.com/index/devday-2026-recap
• OpenAI, Introducing GPT-6.1 Sol: openai.com/index/introducing-gpt-6-1-sol
• Anthropic, Claude Sonnet 5.5: anthropic.com/claude-sonnet-5-5
• Anthropic, Claude Opus 5.5: anthropic.com/claude-opus-5-5