How does context window size vary across Claude's Haiku, Sonnet and Opus tiers?

## The claim being tested The question is really a misconception probe. Price, latency and capability all move together up the Haiku → Sonnet → Opus ladder, so candidates extrapolate and assume context length does too. It does not. In the current Claude generation, all three tiers expose the same standard context window — 200,000 tokens as of mid-2026 — covering everything you send: the system prompt, tool definitions, the full message history, images, and the tokens the model generates in this turn. ## Two limits that are not the same thing **Context window** is the total budget for prompt plus response. Exceed it and the API rejects the request with a validation error rather than silently truncating — an important operational detail, because it means a runaway conversation fails loudly and you must implement your own trimming or summarisation policy. **Maximum output tokens** is a separate, genuinely per-model ceiling on how long one response may be. It is smaller than the window, it does vary between models, and it is the number that changes underneath you when you swap tiers. If a tier migration suddenly truncates your long generations, this ceiling — not the context window — is usually why. **Rate limits** are the third thing people conflate with capacity. Token-per-minute limits are allocated per model on an account, so they constrain throughput, not the size of any single request. ## Extended context Anthropic has offered an extended one-million-token context window as an opt-in beta on Sonnet, enabled with a beta header on the request and subject to premium pricing on requests that run past the standard 200K. Note the shape of that fact, because it is the opposite of the intuition being tested: the long-context option appeared on the **middle** tier, not the top one. Availability, the exact header value and the pricing tiers all change, so in an interview describe the mechanism — per-model opt-in, header-gated, premium priced beyond the standard window — and say you would confirm the current details in the model documentation rather than reciting a header string from memory. ## What to do when the input genuinely does not fit Upgrading the tier is not on this list, and saying so is the point of the answer: - **Retrieve instead of stuffing.** Chunk the corpus, embed it, and put only the relevant passages in the prompt. This is usually cheaper *and* more accurate than a giant context, because the model is not asked to locate a needle in a haystack. - **Summarise the history.** For long conversations and agent loops, periodically replace old turns with a compact summary. Without this, an agent's context grows monotonically until it hits the wall. - **Map-reduce the document.** Process sections independently on the small tier, then combine the section results in one call. - **Use the model that actually offers the bigger window**, accepting the premium pricing, when the task genuinely requires whole-corpus attention — a cross-reference over an entire codebase or contract set, where chunking would destroy the relationships you need. ## Why long context is not free even when it fits Even inside the window, a very large prompt costs real money at the input rate on every single turn, adds latency before the first token appears, and can dilute the model's attention on the instruction you actually care about. "It fits" is not the same as "it is a good idea". Teams that move from stuffing to retrieval usually see cost, latency and accuracy all improve at once, which is why the retrieval answer is the one interviewers are listening for. ## Interaction with caching When a large prefix is stable across requests — a fixed policy document, a tool catalogue, a system prompt — caching that prefix makes long context far more affordable than the raw input rate suggests. That changes the economics of a big-context design without changing the tier, and it is the standard pairing for a document-QA workload where every user question sits behind the same corpus.