Anthropic Tests Message Threads to Stop Resending Chats, ChatGPT Folds Work into Chat, and a Two-Sentence Claude Code Rule Draws Over-Blocking Complaints

Three stories from the past few days are about the plumbing around a chat, not the model inside it: who stores the conversation, how products and quotas are bundled, and how a single prompt rule can change what a tool will do. We could not open the original articles from our environment, so details come from search summaries and secondary coverage, and we flag where sources are thin.

1. Message Threads: Stop Resending the Whole Conversation

Anthropic launched a public beta of message threads in its API, according to The Next Web. Instead of resending the full history on every turn, an agent sends only the new messages. Anthropic reportedly says this cuts the wait before Claude replies in long sessions by about a third. Coverage also notes that Claude still reads the whole conversation on every turn, and that prices and rate limits are unchanged, so the gain is latency and payload size, not model compute. We did not find Anthropic's own documentation, so check it for field names, retention and limits.

Most chatbots today are stateless on the wire: the client or your server keeps the transcript and replays it. That is simple but gets slow and heavy for long agent runs with big tool outputs. Server-held threads move that state to the provider, which raises questions you should answer before adopting it: how long is a thread retained, can you delete it for a user's data request, and what happens if you need to switch providers mid-conversation? A common safeguard is to keep your own copy anyway:

// Keep your own transcript even if the provider holds the thread.
async function send(convo: Convo, userMsg: Message) {
  convo.transcript.push(userMsg);            // your source of truth
  const reply = await provider.threads.append(convo.threadId, [userMsg]);
  convo.transcript.push(reply);
  return reply;
}
// If the thread expires or you change vendors, replay convo.transcript.

2. ChatGPT Retires Classic and Merges Chat and Work

On November 2, 9to5Mac reported that OpenAI is retiring the ChatGPT Classic app and will combine the Chat and Work tabs in a new app, starting with a test on a small group of users. No date for the wider release was given in the coverage we saw. Separately, OpenAI reportedly said that from December, Plus, Pro and Business subscribers will share one usage allowance across ChatGPT and Codex, with instant chat exempt from the quota. Secondary reports say most users will not notice, but heavy users may hit limits more often.

For builders, this is the clearest sign yet that the chat window and the agent workspace are becoming one surface. A user who starts with a question may expect the same thread to run a long task, which means your quota and billing model must handle both. If you meter chat and agent work separately today, decide whether a shared budget is better for users than two meters that run out at different times, and show remaining allowance before a long task starts, not after it fails.

3. A Two-Sentence Rule, and the Cost of Over-Blocking

Anthropic added two sentences to Claude Code's instructions for its Opus 5.5 model to stop it from recreating specific existing video games, according to coverage citing Anthropic staff member Addy Osmani on X on October 30. Developers say the rule also catches legitimate work; one user reported Claude refused to help reverse engineer software on an LG smart TV. We have not seen Anthropic's full wording or any response to those reports.

The takeaway is not about this particular rule. A narrow policy written in natural language is applied by a model that generalizes, so it will fire on neighbors of the case you meant. If you add a restriction to a system prompt, test it against near misses before shipping: a small set of allowed requests that look similar to the blocked ones, and a count of how many get refused.

What to Take Away

Conversation state, product bundling and prompt-level policy are all being adjusted outside of model releases. Keep your own transcript, expect quotas to be pooled, and measure refusals on legitimate requests whenever you edit a system prompt.

Suggested visual: a diagram contrasting a replayed transcript with an append-only thread. Suggested code sample: a refusal-rate test that runs 20 near-miss prompts before and after a system prompt change.

— Maya