Claude Fable 5 Context Window

The 1M window, in numbers

API list prices per million tokens, USD, from Anthropic's pricing docs (verified August 4, 2026). Every value below is read from the site's model-pricing data — change it once, it updates everywhere.

The context window holds everything in the request — system prompt, conversation history, tool definitions, images and documents — plus the output the model generates, including thinking tokens. With max output at 128K, a single request can use roughly 872K of input before hitting the 1M limit.

Rates and specs verified 2026-08-04

What 1M tokens actually costs — worked example

Numbers computed from the site's model-pricing data. A "full-window pass" is 1M input tokens as counted by Fable 5's own tokenizer; output bills on top.

One full-window input pass (uncached)

$11.00

Same pass, batch mode

$5.50

Cached re-read, per M tokens

$1.00

Batch mode halves the input leg but is not real-time — right for scheduled analysis jobs, wrong for interactive work.

The tokenizer tax — why long context costs ~30% more than it looks

Anthropic documents that Fable 5 uses the tokenizer introduced with Opus 4.7, which splits the same text into roughly 30% more tokens than pre-4.7 models (documented range 1.0x–1.35x, weighted toward the top on code, JSON, XML and YAML). The site's model-pricing data uses a single factor:

• Shared by every current model

  Fable 5, Mythos 5, Opus 4.8 and Opus 5 all use the new tokenizer, so switching between them changes nothing. The inflation only bites when you compare against a pre-4.7 model or an old token count.

• Every old token budget is ~30% too low

  Cost estimates written against a pre-4.7 token count understate spend by about 30%. A corpus that measured 1M tokens bills as roughly 1.3M — about $13 per pass at Fable 5 rates — and now exceeds the window.

• Long context is where the tax bites

  On a 2K-token prompt the extra ~600 tokens are a rounding error. On a 1M-token pass the tax is $3 of every $10 — and it compounds across cache writes and every re-read of a large corpus.

Long-context costs, honestly

One full-window pass over a 1M-token corpus, priced per model. The Haiku 4.5 column is the 200K-window reality check. List rates, USD, from the site's model-pricing data.

Assumptions: 1M input tokens as counted by the model (tokenizer inflation deliberately not applied — all three models share the tokenizer, so it cancels in the comparison), 20K output per pass at high effort, no safeguard fallback. "Uncached" = 0% cache hit, "Cached" = 70%. Haiku 4.5 cannot hold the corpus in one request — a 200K window needs about six ~180K chunks, and chunked processing loses cross-chunk reasoning. Caching changes the story: at 70% hit, Fable 5's pass costs $4.70 instead of $11.00.

The biggest lever is caching, not the window size

Cache turns the long-context story around. Same corpus, many requests: write the 5-minute cache once, re-read at $1 per million tokens.

Loading a 1M-token corpus into the 5-minute prompt cache costs $12.50 once, then $1.00 per million tokens on every re-read. For RAG-style workloads that re-scan the same corpus — codebases, document sets, support archives — that is what makes the 1M window affordable: 70% hit ≈ $4.70 per pass instead of $11.00. Batch mode halves the input leg of a full-corpus pass (not real-time).

When you need 1M — and when you don't

The 1M window is a tool for a specific class of jobs. Most real workloads fit in 200K and should stay there.

• Genuinely 1M-needed

  Full-repository code analysis — a large monorepo fits in 1M tokens, not in 200K. Whole-corpus RAG — a small database or document set loaded entire: no chunking, no retrieval to tune. Long agentic runs — a multi-hour agent accumulates context turn by turn; with 128K output there is roughly 872K of room for the trail it leaves behind.

• Not 1M-needed

  Single-document work — a long paper, a code review, a legal filing: 200K is already generous. High-frequency traffic — the per-request cost of a 1M pass ($10 before caching) compounds on volume. Routine conversation — a chat that never approaches 100K has no reason to pay frontier input prices for headroom.

• The 3-question decision

  Does the corpus fit in 200K? If yes, use a 200K model and stop reading. Does the task need the whole corpus in context at once? If not, chunking or retrieval wins. If both answers are yes — does it re-scan the same corpus? If so, cache it and the 1M window is cheap; if it is a one-off full-corpus analysis, that 1M pass is a real, deliberate cost.

The older 128K-window generation (2023–24 era models) cannot hold a 200K corpus, let alone 1M — a 1M corpus chunked into 128K blocks needs about eight passes and loses cross-block reasoning. That class is effectively priced out of long-context work; today the choice is 200K (Haiku 4.5) or 1M (Fable 5, Opus 5, Sonnet 5).

1M context × 30-day retention: the compliance angle

Fable 5 API use carries a mandatory 30-day data-retention window — zero-data-retention agreements do not apply, and the requirement covers everything you send through the 1M window. Teams that process large corpora (source code, customer documents, PII-adjacent records) should treat any long-context workload as data that must be retained for 30 days. If your compliance posture rules that out, that alone decides against Fable 5 for that workload.

Fable 5 API use carries a mandatory 30-day data-retention window — zero-data-retention agreements do not apply, and the requirement covers everything you send through the 1M window. Teams that process large corpora (source code, customer documents, PII-adjacent records) should treat any long-context workload as data that must be retained for 30 days. If your compliance posture rules that out, that alone decides against Fable 5 for that workload.

Go deeper