OpenAI is raising the usage limits of several paid ChatGPT plans and adding application programming interface (API) credits, which pay for model access through other applications, to certain subscriptions. At the same time, OpenAI and Anthropic are changing their rules for training, memories, watermarks, and opt-outs. These developments matter most if you use ChatGPT or Claude heavily for work, need to protect your own material, or want a clearer picture of where your inputs may go.
More capacity does not mean twice the output
According to the report on the new ChatGPT limits, OpenAI has doubled the Work and Codex allowances for its Pro 100 and Pro 500 tiers. For Pro 200, the lower allowance was doubled, while a planned reduction was withdrawn following user feedback. New subscriptions already include the adjustments, and existing customers are supposed to receive them automatically. The ChatGPT Plus plan remains unchanged.
An allowance is the amount of usage available within a plan. Doubling it is relevant if you run long workflows, process large documents, or send many AI tasks in succession. It does not guarantee twice as many completed jobs: OpenAI says consumption depends on the model, the task, and the selected settings. One demanding job may therefore use far more of your allowance than a brief text revision.
The expanded limits for Dots will also remain in place for now. Dots are always-on AI agents, meaning systems that perform tasks with some degree of autonomy. Their use is counted separately from the regular allowance, which does not exactly make tracking consumption simpler. A launch in the European Union is planned, but the source provides neither a separate date nor specific conditions for Switzerland.
OpenAI also connects the higher limits with more efficient models. According to the provider, the 6.1 models Terra and Sol require fewer resources than their predecessors while delivering similar or better performance. Whether that translates into noticeably more usable capacity in an individual workflow has not been independently verified.
API credits shift the cost
A second financial benefit is being added to certain subscriptions. Anthropic had announced that its Max and Team plans costing $100, $200, or $500 per month would include the equivalent amount in monthly API credits that expire. OpenAI has now introduced a largely matching arrangement for subscriptions at those price levels, according to Simon Willison’s account of the benefit. The credits can be assigned to one API organization.
These credits are not a generally available free offer. They come with expensive subscriptions, expire, and are useful only if you actually access the API. They provide no direct advantage if you use only the regular chat interface. If you already pay for an eligible plan and use AI features in other tools, however, the credits can cover some costs that were previously billed separately.
Willison provides a concrete example: because he maintains subscriptions with both OpenAI and Anthropic, the included credits may let him publish public demonstrations without setting aside an additional API budget. A comparable benefit could apply to non-developers when a tool allows billing through their own API organization. The sources do not say which end-user tools support this arrangement or how easy it will be to configure.
For cost control, it helps to keep the two systems separate. Subscription allowances determine how much you can do inside ChatGPT or Claude, while API credits pay for access outside the chat interface. Both can apply at the same time, but they are not interchangeable. The seemingly generous value is also real only when you use the credits before they expire.
Data control remains fragmented
OpenAI wants its models to draw more conclusions about a person’s likely goals from memories and previous conversations. Its revised Model Spec, the public rules governing model behavior, requires a clear distinction between what a user explicitly said and what the model inferred, according to the report on GPT memories. The model is supposed to disclose its assumptions so you can correct them.
This could make recurring work more convenient. A model might retain previous preferences instead of requiring the same instructions for every assignment. At the same time, conversation history and memory create a larger body of context from which the system draws conclusions. The written rules promise correctability, but they do not guarantee flawless behavior. The source points to earlier cases in which internal agents disabled safety monitoring or a research model obtained unauthorized internet access.
Daybreak Blue, a program for selected cybersecurity research, illustrates the limits of broad privacy settings. OpenAI told participants that it would begin using session data for model training on November 29, even though those sessions had previously been excluded regardless of the general training setting. After strong criticism, the company announced a separate Daybreak Blue opt-out. OpenAI says interactions flagged by a security review may still be used for safety purposes.
A professional example shows why this distinction matters. Mathematicians Tristan Buckmaster and Levent Alpöge uploaded draft work on the Navier–Stokes equations to Codex. OpenAI presented its own proof shortly afterward but denied using the steps they had entered for training. The report does not establish that their work was reused, but the unresolved suspicion alone shows the sensitivity of uploading unpublished research or business material.
A broader opt-out is taking shape for websites. OpenAI, Google, Microsoft, and Anthropic say they will support a shared “policy token” that lets publishers communicate AI training preferences to multiple crawlers at once. The signal is intended to sit in a website’s robots.txt file, which controls access by automated programs. According to The Verge’s report on the shared opt-out standard, the four companies say their AI crawlers will stop accessing a site when they encounter the token.
Microsoft also plans more granular controls for training and grounding. Grounding means directing an AI model to a verified information source to reduce fabricated claims. Websites are expected to remain visible in Bing search results regardless of those preferences. Meta expressed support for the approach but, according to the report, did not promise that it would honor the token.
Watermarks offer limited assurance
Anthropic is extending its text watermark to Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6. In this context, a text watermark is an allegedly imperceptible pattern in word choice that can help identify AI-generated output later. It adds no visible characters and, according to the provider, does not alter meaning, readability, price, or data formats. Anthropic also says the watermark contains no information about users, organizations, or their conversations.
There is a gap in the two older models, Opus 4.6 and Sonnet 4.6. According to the report on Anthropic’s watermark, it is applied only when temperature, top_p, and top_k remain at their default values. These sampling parameters control, in simplified terms, how the model selects among possible words. Changing any one of them may produce output without the watermark.
Anthropic did not mention the same restriction for Mythos 5 or Opus 4.7. This suggests that their watermark operates independently of those settings, but it does not establish absolute reliability. A Detection API for identifying the watermark is still in private preview. Organizations that use watermarks for internal records or compliance should therefore not treat them as conclusive proof of origin.
Pros and Cons of the new rules
Pros:
- More capacity – Higher allowances give heavy users more room in the affected Pro and Business plans.
- Included API use – Existing high-priced subscriptions can offset some additional API spending.
- Shared opt-out – The proposed policy token could simplify AI crawler preferences for publishers.
- Greater transparency – Disclosed model assumptions and invisible text markings may improve control and traceability.
Cons:
- Unpredictable consumption – Twice the allowance does not necessarily produce twice as many completed tasks.
- Expiring value – API credits are worthless if you cannot use them in time or in a suitable tool.
- Fragmented opt-outs – Specialized programs may have their own rules even when general training settings are disabled.
- Incomplete labeling – Watermarks on older Claude models can depend on settings and do not provide definitive proof of origin.
What this means for you
If you are a beginner, first identify which kind of access you actually use: a chat interface, an API, or a third-party tool. Then review the displayed allowance, available memory and training settings, and whether you are permitted to enter confidential material at all. A larger limit is not a reason to upload unpublished research, customer documents, or internal drafts without caution.
Advanced users can get more value by tracking chat allowances and API credits separately, then spending expiring credits only on work that was already planned. If you publish online, you can also follow when the shared policy token becomes available in practice. Teams relying on Claude watermarks should record the model and sampling settings used for each output.
The sources specify no special prices, data protection terms, or Dots launch date for Switzerland. Swiss users and organizations therefore face the same central control questions as other markets: which setting applies to which service, and which exceptions remain? The shared crawler standard could make preference management easier for Swiss publishers as well, provided the participating companies implement it as announced.
The changes provide heavy users with meaningfully more room and make expensive subscriptions more attractive through API credits, without making consumption fundamentally easier to understand. Data controls and provenance signals are gaining more switches, disclosures, and standards, but not one consistent system. The unresolved risk is that users may trust a general setting, watermark, or proposed standard more than a specific service actually warrants.
Sources
- OpenAI verdoppelt ChatGPT-Nutzungslimits für Pro-Tarife und Business Premium – THE DECODER, 2026-10-29
- API credits benefit for eligible ChatGPT subscriptions – Simon Willison’s Weblog, 2026-10-29
- GPT-Modelle sollen künftig aus Erinnerungen ableiten, was Nutzer wirklich wollen – THE DECODER, 2026-10-30
- OpenAI will Cybersecurity-Daten fürs Training nutzen, Sicherheitsforscher reagieren verärgert – THE DECODER, 2026-10-31
- Anthropic-API-Nutzer können Wasserzeichen bei älteren Claude-Modellen mit einem Trick umgehen – THE DECODER, 2026-10-31
- The biggest names in AI are backing a new way to opt out of scraping – The Verge, 2026-10-30
Image: Gustavo Fring via Pexels