//
Search across all documentation pages
269 pages across 50 sections. 3000 questions total.
No. Claude generates one token at a time, and each new token is chosen using everything written so far, including the tokens it has already produced in that same response.
Because next-token generation involves sampling from a probability distribution rather than always selecting one fixed answer, small variations in wording are normal and expected.
No. A database lookup retrieves a stored, pre-existing record. Next-token prediction constructs new text by predicting likely continuations based on learned patterns, with nothing being retrieved from storage.
Fluent, confident-sounding phrasing comes from the same prediction mechanism regardless of whether the underlying content is accurate. There is no separate certainty check built into basic generation.
Not in a fixed sense. Each token is chosen based on what came before it, so the response takes shape progressively rather than existing as a complete plan from the first token, although extended thinking modes add a visible reasoning pass before the final answer.
It's the set of likelihood scores Claude assigns to every possible next token at a given step, with far more likely tokens scoring much higher than implausible ones.
No. In a plain chat conversation, Claude relies on patterns learned before its knowledge cutoff plus whatever context is in the conversation, unless a research or browsing feature is explicitly enabled.
The core mechanism is shared, but models differ in depth of computation. Faster models like Claude Haiku 4.5 predict quickly with less depth, while models like Claude Opus 4.8 and Claude Fable 5 apply extended reasoning before finalizing an answer.
Because tokens are produced sequentially, one prediction step per token. A longer response requires more prediction steps, not a single larger lookup.
Yes. Since every prediction depends on the exact preceding text, changing your wording changes the context Claude is predicting from, which can shift which tokens seem most likely at each step.
Token. Nearly every other concept on this page, context window, pricing, latency, is defined in terms of tokens.
No. Training (including pretraining and fine-tuning) happens beforehand to produce the model. Inference is what happens every time the already-trained model generates a response to your input.
Because generation is prediction of the next plausible token, not a fact-check against a source. Fluent, confident phrasing can occur regardless of whether the specific fact stated is correct.
No. Plain Claude chat does not automatically retrieve external documents. RAG requires a specific connector, tool, or application built to fetch and inject relevant material before generation.
Zero-shot gives the model no examples and relies on what it already learned during training. Few-shot includes a small number of examples in the prompt itself to demonstrate the exact pattern you want.
Lower temperature makes responses more predictable and repetitive, favoring the single most likely next token each time. Higher temperature allows more varied, sometimes more creative, phrasing by sampling from a wider range of plausible tokens.
It determines how much conversation history, code, or documents can be considered at once. Running out of context window mid-task means older content is no longer visible to the model.
A visible, step-by-step reasoning pass before the model commits to a final answer, which tends to improve reliability on complex, multi-step problems compared to a direct answer with no such pass.
Not necessarily. Parameter count is one factor among training data quality, training method, and reasoning depth. A smaller, well-trained model can outperform a larger, poorly-tuned one on a given task.
Anthropic's training approach for shaping Claude's behavior around a set of guiding principles aimed at being helpful, harmless, and honest, combined with safety testing before a model is released.
Grouping by theme, core mechanics, day-to-day usage, and how Claude is built, keeps related concepts near each other, which makes the list easier to learn from on a first read, not just to look up later.
They're not really comparable on a single scale. Rule-based software is more reliable and auditable within its defined scope; Claude is more flexible across a much wider range of unanticipated inputs. Each is better suited to different problems.
Because Claude generates each response by sampling from a probability distribution over likely next tokens, rather than executing one fixed logic path, so some variation between runs is expected.
Not in the same way you can trace a rule-based system's exact triggered rule. Extended thinking modes expose a visible reasoning process, which helps, but it is not a guaranteed, line-by-line justification the way rule-based logic is.
No, not in the LLM sense. A rule-based chatbot that hits an unhandled input typically returns a fallback message or an error, rather than generating a fluent but incorrect answer.
When correctness must be guaranteed and every decision needs to be traceable and repeatable, such as calculating a bill, checking eligibility against fixed criteria, or enforcing a regulatory rule.
When the input is natural language, the range of valid phrasings is too large to enumerate as explicit rules, or the task benefits from summarizing, drafting, or reasoning over unstructured text.
Yes. A frequent pattern is rule-based logic validating or constraining what reaches or leaves Claude, while Claude handles the flexible, language-heavy part of the task in between.
Yes. Rule-based logic typically runs at a small, fixed compute cost per request. Claude's cost scales with the number of tokens processed and generated, and varies by which model tier you use.
No. A rule-based system changes when a developer edits its rules directly. Claude's behavior for a given task is typically adjusted through prompting, added context, or choosing a different model, or it can shift when Anthropic ships a new model version.
Claude generally handles unanticipated phrasing better, because it generalizes from learned patterns. Rule-based matching typically fails outside the specific phrasings its author anticipated.
Not by default. Claude's knowledge is grounded in training data up to a fixed knowledge cutoff, plus whatever is in the current conversation. Live data access requires an explicitly connected tool or feature.
Not by default in a plain chat conversation. Claude's response is generated from patterns learned during training, plus your conversation's context, unless a browsing or research feature has been explicitly connected in the product you're using.
Because Claude generates new text rather than retrieving indexed pages, there often isn't a single existing source its answer came from in the way a search result is a specific webpage.
Not universally. A rule-based chatbot is more reliable and auditable within its narrow, well-defined scope. Claude is more flexible across a much wider range of inputs, but its generated answers are not automatically verified against a checkable source.
Because it's generating plausible-sounding text based on learned patterns, not retrieving a verified fact from an index. This is called hallucination, and it's a structural byproduct of how generation works.
When you specifically want to find an existing, checkable source, such as a particular article, an official document, or a page you vaguely remember seeing before.
When you want something drafted, summarized, explained, or reasoned through, tasks that involve synthesizing or generating new text rather than locating an existing page.
Usually not well. It typically needs your input to match a pattern the developer anticipated in advance, and it often falls back to a generic "I didn't understand that" message otherwise.
No. Claude's knowledge comes from a training dataset collected up to a fixed knowledge cutoff, and it has no automatic, live connection to current web content by default.
Claude processes and responds to your question using patterns learned from training, producing genuinely useful, context-aware responses, without implying the same kind of comprehension or intent a person has.
Yes. Many products connect Claude to a search or research tool specifically so it can retrieve current, checkable information and then use its generation ability to summarize or reason over that retrieved content.
Start with group A, evaluating fit, then run a small, low-stakes pilot per group B before expanding to broader use.
Group C, verifying output. Grounding Claude in source material and keeping a human review step for high-stakes tasks are the most direct levers on output quality.
Ongoing. Verification habits (grounding, human review, treating specific facts as claims to check) need to run continuously, not just during an initial pilot.
Yes. Pricing, defaults, and capability shift with each release, so a model choice made previously is worth periodically revisiting rather than treated as permanent.
Treating confident-sounding output as automatically verified. Fluent phrasing and factual accuracy come from the same generation process, so specific claims still need checking, especially early in a rollout.
No. Group D specifically recommends matching model tier to task: fast, cheap models for high-volume simple tasks, and deeper-reasoning models for complex, high-stakes work.
A narrow pilot surfaces real failure modes on your actual content and use cases, which is more useful and lower-risk than committing to a broad rollout based on general expectations alone.
Providing the actual source material, a document, dataset, or specific context, directly in the conversation, so Claude can work from what's actually there rather than relying purely on patterns learned during training.
No. Extended thinking tends to reduce certain reasoning errors on complex tasks, but it doesn't add facts the model never learned, and it doesn't replace a human review step for high-stakes output.
Because they define categories of restricted use. Checking them before rolling Claude out for a new use case avoids building a workflow that later needs to be redesigned to comply.
A person. Group E is explicit that responsibility for consequential decisions should stay with a human, with Claude used to draft, summarize, or assist rather than to make the final call unsupervised.