Chat, Rebuilt from the Ground Up

This is our biggest change to Chat since Conception launched. We rebuilt the core chat pipeline from the request you send to the answer you receive. Most of that work lives below the surface, but it changes what Chat can handle: longer conversations, richer context, files and generated artifacts, and a much more reliable way to stop, continue, rewrite, or branch a conversation.

Meet Claude Fable 5.1

Anthropic's Claude Fable 5.1 is now available in Conception. Anthropic's newest flagship model is built for demanding reasoning and long-horizon agentic work - the kind of tasks that require sustained context, careful analysis, and more than one step to get right. It is available in the model picker on plans that include frontier models, and is fully integrated with Chat's new reasoning, search, attachment, and artifact capabilities.

A Better Way to Search and Answer

Chat now plans its work before it answers. Depending on your question, it can search the web, look through academic sources, or search your workspace. Searches run together, the results are curated before they reach the answer model, and Chat can perform a bounded follow-up search when the first set of sources does not cover the question well enough. The result is less irrelevant context and more focused answers.

Sources are easier to follow, too. Answers can show the search activity behind them, cite the pages they used, and surface source-review information when it is available. Workspace mentions let you point Chat directly at the notes, chats, or databases that matter, so it can work from your own material alongside the open web.

Conversation history is now assembled on the server as real user and assistant turns instead of being copied wholesale from the browser. That makes follow-up questions more dependable, keeps requests smaller, and gives Chat a cleaner foundation for long-running conversations.

Files In, Useful Work Out

Chat can now work with a much broader range of attachments, including images, PDFs, text documents, and spreadsheets. When a model can read a file directly, it receives the appropriate native input; when it cannot, Conception prepares a searchable description or extracted text instead. You no longer need to choose a model based only on whether it can natively open the file you want to use.

Chat can also create files for you. Ask it to calculate, transform, or generate something and it can run code in an isolated environment, then return the result as an artifact. PDFs, Word documents, presentations, spreadsheets, CSVs, Markdown, code, and images can be previewed in place and downloaded when you are ready to use them.

Conversations That Give You More Control

• Rewrite or edit an earlier message without losing control of the conversation. Chat explains when later turns will be affected before it changes them.
• Branch from a previous point to explore a different direction while keeping the original conversation intact.
• Keep alternate answers when you regenerate a response, so a useful answer is not lost just because you wanted to try another model or prompt.
• Stop and recover cleanly. Interrupted or failed responses are recorded honestly, and the next turn can continue without duplicated or half-written messages.
• Pin important chats in your history so the conversations you return to most often stay easy to find.

Reliable Across Tabs, Retries, and Real-World Use

Chat turns now have durable state behind them. If a connection drops, a tab reloads, or a request is retried, Conception can recover the turn instead of starting from an ambiguous partial state. A chat also accepts only one active turn at a time, preventing two open tabs or duplicate requests from interleaving messages or overwriting each other's work.

The same foundation makes long conversations easier to manage: history loads progressively, checkpoints and generated files are cleaned up with the conversation that created them, and usage limits are applied to the work that actually runs.

A Safer Foundation for Your Workspace

Every chat, attachment, mention, and generated artifact is now checked against the workspace and user it belongs to. Uploaded files cannot be reused across accounts, artifact downloads use restrictive browser-safe headers, and potentially active formats such as HTML and SVG are handled as data rather than executable content. Retrieved pages and workspace text are also kept structurally separate from Chat's instructions, strengthening resistance to prompt injection.

This release also brings a new evaluation layer around Chat. We now test answer quality, search coverage, latency, attachment retrieval, artifact access, prompt-injection resistance, and concurrent requests as part of the platform itself. The goal is simple: Chat should be more capable when you use it, and more predictable when things go wrong.