Standard keeps a larger hidden memory and more recent raw context. Prefer it for programming, infrastructure, exact IDs, commands, versions, financial, legal, medical, and other detail-sensitive work. It is also usually sufficient for short chats and can benefit from five-minute cache continuity during an active conversation.

Lite uses a shorter hidden memory and two to three recent raw turns after compaction. Prefer it for casual or long creative chats, especially when resuming after pauses and broad continuity matters more than every exact detail. Lite begins around 5,000 uncovered visible tokens; Standard begins around 16,000. Avoid frequent switching because each mode maintains its own checkpoint.

One more message expected: Standard is usually sufficient. Many future turns after pauses: Lite may reduce repeated context cost. Actual cost depends on output length, cache behavior, provider pricing, and conversation content.

Web search lets Fable use native searches for current information. It can increase latency and cost; check sources for important decisions. It applies to future turns and does not rewrite earlier messages or remove visible citations.

Direct mode performs classic search and is ZDR-eligible where the account supports it. Dynamic filtering uses Anthropic's sandboxed code-execution environment to filter results before relevant information enters context; it can reduce tokens on search-heavy turns but is not ZDR-eligible. Both lets Fable choose either path.

Excluded response inclusion applies only to completed Search results consumed through dynamic filtering. Direct and paused results remain full. Domain and approximate-location controls are optional and server validated. Provider tool choice None disables both Web Search and Web Fetch and can replace the prompt-cache prefix.

Maximum searches per message: Medium 1, High 3, XHigh 5, Max 10. Disabling Web search prevents new searches.

Web fetch reads a specific page or PDF URL already present in the conversation. Web search discovers pages; Web fetch reads a selected source. It can increase input-token usage.

Fable may make up to 2 fetches per message and include up to 8,000 text tokens per fetched resource. The server controls these limits. Web fetch is off by default and applies only to future turns.

Shows a provider-returned summary when available. It never exposes private thinking, signatures, or chain-of-thought. It does not disable the model's internal reasoning and may add visible output or latency.

Effort changes reasoning depth, latency, and potential cost. Output ceilings: Medium 8,192; High 16,384; XHigh 32,768; Max 32,768 tokens. XHigh and Max share an output ceiling but remain different inference-effort choices.

General is a natural helpful assistant. Coding emphasizes rigorous programming, debugging, and review. Creative supports writing and ideation. Precise favors concise, exact answers that separate facts, uncertainty, and assumptions. Changes affect future turns only.