Claude Ships Read-Only Personal Finance, Character.AI Reportedly Files for Chapter 11, and ChatGPT Sites Swaps Site Counts for a Shared Usage Cap

Three stories from the past few days: a major assistant connecting to bank accounts with deliberately limited powers, a companion-chatbot company reportedly entering bankruptcy, and a quiet change in how one product meters what users can build. Details come from secondary coverage and trackers, so treat specifics as the reporters' own.

1. Claude's Personal Finance: Read Access, No Write Access

On November 5, Anthropic announced Personal Finance in Claude, a US-only beta for Pro, Max and Ultra subscribers, according to multiple outlets. Accounts link through Plaid, so Claude never sees bank passwords; users can unlink at any time. Coverage says Claude cannot move, manage or invest money, and that the feature carries no ads or sponsored recommendations. Anthropic reportedly says conversations inside the Personal Finance hub are excluded from model training, while chats outside it follow normal privacy settings. Outlets differ on details such as history length and the number of linkable accounts, so check Anthropic's help article for those. Coverage frames this as a response to OpenAI's Plaid-based finance preview for Pro users in May.

The design choice is the interesting part. Sensitive data in, no side effects out. If your assistant touches money, health or HR records, split tools by blast radius: reads are cheap to grant, writes should be separate, narrow and confirmed.

const tools = {
  get_transactions: { sideEffects: false, run: getTransactions },
  // Writes are a different tier: explicit, narrow, user-confirmed.
  schedule_transfer: { sideEffects: true, requiresConfirmation: true, run: scheduleTransfer },
};

async function callTool(name: keyof typeof tools, args: unknown, confirmed = false) {
  const t = tools[name];
  if (t.sideEffects && !confirmed) return { status: 'needs_confirmation', name, args };
  return t.run(args);
}

2. Companion Chatbots Meet Vendor Risk

An October 24 case tracker lists a Chapter 11 filing by Character.AI in Delaware bankruptcy court on October 12, with pending suits paused while a sale process proceeds. We found only that single tracker, not a company statement or court filing, so this is unconfirmed. Separately, Google and Character.AI were reported in January to have settled several teen-harm suits.

If it holds up, the practical lesson is about dependency. Teams that build on a third-party character or companion platform inherit its legal exposure and its solvency. Keep conversation data exportable, avoid hard-wiring one vendor's persona format, and have a plan for what users see if the backend disappears.

3. Metered Limits Replace Hard Counts

OpenAI's help center says ChatGPT Plus and Pro users have no fixed limit on how many ChatGPT Sites they can create, but usage and storage limits still apply across all Sites. Third-party guides describe a single plan-specific beta cap combining Site count, storage and traffic, with exact values unpublished. Hitting it reportedly blocks new Sites or keeping a high-traffic Site public, while editing continues. We could not confirm the exact date this wording changed, and a third-party analysis dated October 3 said the limit had neither been a flat 10 nor fully removed.

For builders, an opaque shared budget is harder to explain than "10 projects". If you meter agents or generated artifacts, expose current usage, warn before the cliff, and make clear what still works once a user hits it.

Conclusion

Powerful assistants are being shipped with narrower permissions, companion-style products are becoming a legal and financial liability for the companies running them, and limits are shifting from counts to budgets. The common thread is operational honesty: say what the system can touch, who stands behind it, and what it will cost the user to keep using it.

Suggested visuals: a read-versus-write tool tier diagram, and a usage-meter mockup showing a warning at 80%.

— Maya