Edit a Long Manuscript with Claude Sonnet 4.6

A novel-wide continuity problem needs evidence from distant chapters. A sentence-level repair usually does not. The cost difference comes from how much of the book you resend each time.

Send the whole manuscript to Sonnet 4.6 once, to find problems; send chapter-sized packets to fix them. On SpicyAPI the price is $3 per million input tokens and $15 per million output tokens. A 400,000-token manuscript (a rough estimate for 300,000 words of English prose, not a tokenizer measurement) plus 5,000 billed output tokens costs $1.275 without caching. A 45,000-token focused packet with the same output costs $0.21. The mistake that gets expensive is repeating the full read for every chapter: 40 chapter edits that each resend the book cost about $50, against $4.80 with 25,000-token packets or $7.20 with 45,000-token packets, each returning 3,000 billed output tokens.

SpicyAPI uses one token-price tier across this route’s 1,000,000-token input window; there is no 200,000-token surcharge. For version choice, see the Sonnet API version guide; for quotation and source-credit checks, see the Opus editing workflow.

Limits and rates

The Sonnet 4.6 endpoint uses model ID anthropic/claude-sonnet-4.6/chat. It accepts text and image_url parts, up to 1,000,000 input tokens, and permits max_tokens up to 64,000 with a default of 4,096. Effort options are low, medium, high and max; the default is medium. It exposes temperature, top_p and top_k but not response_format. For manuscript editing, send extracted text so chapter boundaries and exact wording remain available.

Two ceilings shape every plan. Input is 1,000,000 tokens, but output stops at 64,000 billed tokens ($0.96 at the output rate if all are used), so reading a whole novel does not mean returning one in a single reply. Reasoning tokens count as output: reasoning_tokens is already included in completion_tokens, so do not add it a second time.

SpicyAPI lists Sonnet 4.6 as Uncensored and adds no platform-level content filter, but it can decline or soften a blunt request that comes without context. State the fiction-editing assignment in the first sentence of your system message, as the example below does, and name the passages to preserve. Sonnet 5.5 is the alternative if retaining mature content is the main priority.

Decide what belongs in the request

Suppose your fictional adult protagonist Mara has a left-wrist injury, a brass key stays in an upstairs drawer, and the reason for her return is not revealed until chapter ten. You want to tighten chapter four without changing those facts.

For that local assignment, send chapter four, its adjacent transitions, the three governing facts and a short approved style sample. Sending all ten chapters is useful if you want the model to discover contradictions; it is unnecessary if the author has already identified the exact passage and continuity rules.

For a whole-book audit, send the manuscript with stable chapter and paragraph labels. Ask first for a list of contradictions and supporting passages. Do not combine "find every inconsistency" with "rewrite the entire book" in one output: the detection task can use a large input while returning a short, inspectable report.

Try it in Playground first

1. Open Claude Sonnet 4.6 and switch to Playground.
2. Put the editing instruction in System prompt, and the canon facts plus the conflicting passage from the example below in Message to the model. Ask for a local repair rather than a chapter rewrite, then click Run model.
3. Check that the repair keeps the injury, key location and hidden motive intact.

A bounded editing request

Set SPICY_API_KEY in your environment using a key from API Keys. The following request uses SpicyAPI's Chat Completions interface.

cat > request.json <<'JSON'
{
  "model": "anthropic/claude-sonnet-4.6/chat",
  "messages": [
    {
      "role": "system",
      "content": "Edit the fictional passage for clarity without changing story facts. Return a replacement paragraph followed by a brief explanation of each factual repair. All characters are adults."
    },
    {
      "role": "user",
      "content": "Canon: Mara is 32. Her left wrist is injured; she cannot lift a heavy case with that hand. The brass key remains in the upstairs drawer throughout chapter four. The reason for her return is not revealed until chapter ten. Chapter 4, P12: \"Mara swung the heavy case onto the bench with her left hand. The brass key warmed her palm as she remembered exactly why she had returned.\" Repair P12 without inventing the hidden reason or moving the key. Preserve the action of putting the case on the bench."
    }
  ],
  "max_tokens": 4096,
  "reasoning_effort": "medium"
}
JSON

curl --fail-with-body https://api.spicyapi.ai/v1/chat/completions \
  -H "Authorization: Bearer $SPICY_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $(uuidgen)" \
  --data-binary @request.json

A valid repair lets Mara use her right hand and removes the key from her palm, while leaving the hidden motive unstated. It should not solve the injury by declaring it healed or replace the heavy case with an invented lightweight bag; those changes evade the assigned continuity constraints.

If your output is cut off, reduce the requested deliverable or raise max_tokens; sending more background will not create more output room. If it finishes but changes the reveal, repeat the exact reveal constraint beside the affected passage and ask for a local correction.

For a full manuscript, add "stream": true and "stream_options": {"include_usage": true} to the same request body. A 400,000-token read can run past the roughly 100-second limit on non-streaming calls and return no answer; streaming avoids that, and the final chunk then carries the token counts you need for the cost check below. Join the delta.content pieces in order to rebuild the report, and ask for findings only (paragraph ID, quoted passages, rule broken), at low effort with max_tokens around 16,000, because reasoning counts inside max_tokens. The chat guide covers streaming and recovery.

Let an Agent prepare the call

For repeated manuscript passes, ask your coding Agent to use the SpicyAPI Skill: “Read anthropic/claude-sonnet-4.6/chat, then adapt the example to my input; show the request before running it.” The CLI lookup is spicyapi models get anthropic/claude-sonnet-4.6/chat --json; the MCP lookup is spicyapi_model_get with model: "anthropic/claude-sonnet-4.6/chat". Chat calls use a compatible client or the chat API; the media task commands do not generate chat replies.

What an editing session costs

One request is rarely the whole bill. Each pass resends the packet, so the number to estimate is packet size times passes, plus output.

Build a 25,000-token packet from the chapter (10,000), its neighbours (4,000), the canon sheet (6,000) and instructions plus style sample (5,000). Forty packets cost $3.00 in input and $1.80 in output under the table’s assumptions. Allow more billed output for rewrites at high or max effort.

For agent-driven edits, cap passes and name the files in scope; a blanket “check everything” rule can multiply reads beyond the editing assignment.

What a usable chapter-four edit looks like

The sample is small, but it contains three different mistakes. The injured wrist is a physical continuity error. The key in Mara's palm is a prop-location error. Explaining her return is a reveal-timing error. Naming these separately lets you repair them without changing the scene's purpose.

An illustrative human-written replacement is:

Mara eased the case onto the bench with her right hand and waited for the ache in her left wrist to settle. The upstairs drawer would have to wait. So would the question that had brought her back.

This is not a Sonnet result. It preserves the heavy case and injury, leaves the key upstairs and keeps the motive undisclosed. Depending on the chapter's voice, the last sentence may still draw too much attention to the mystery. In that case, ask for a quieter ending tied to the next physical action, rather than asking for an entirely new scene.

The distinction matters when editing adult fiction. A model can produce grammatically clean prose while changing who initiates an interaction, resolving a conflict too early or making a guarded character suddenly candid. Supply the relationship state for this chapter, not the ending you hope to reach eventually. “They reconcile in chapter ten” should be labeled future canon, so chapter four can preserve the distance that makes the reconciliation meaningful.

For a local line edit, ask for one replacement and a compact explanation. For a developmental edit, ask why the scene's decision feels unearned, with references to the preceding dialogue. Those tasks need different evidence. A wordy rewrite cannot compensate for failing to understand what the scene changes.

Keep a whole-book audit from becoming a whole-book rewrite

When searching a full manuscript, a useful finding has two locations and a reason they conflict. “The key's location is inconsistent” leaves you doing the search yourself. “Chapter 3 P19 locks the key in the drawer; chapter 4 P12 puts it in Mara's palm without a retrieval” gives you an edit to consider.

Require only contradictions supported by passages in the supplied manuscript. For an absent handoff, the model should label a missing explanation rather than insist the action is impossible. A later reveal may explain it. Review those findings against your intended plot before asking for replacements.

This two-pass approach also changes the bill in a concrete way: the expensive full-input request returns a short findings list, while later revisions carry the affected scene and its governing facts. Keep a full manuscript in subsequent calls only when those local facts are insufficient. The cost examples below show how repeated full-input calls and actual cache reads affect that choice.

Calculate caching without double-counting input

Divide input into mutually exclusive categories: uncached input U, cached reads C, and cache-write tokens W. With output O:

Cost = (3U + 0.30C + 3.75W + 15O) / 1,000,000.

Do not count cached-read tokens as ordinary input too. Consider a 400,000-token packet with a stable 320,000-token manuscript prefix, 80,000 tokens of changing material and 5,000 output tokens:

The write costs 25% more than ordinary input and a read costs 10% of it, so a prefix that is read even once after being written already comes out ahead (320,000 tokens: $1.296 for write plus one read against $1.92 for two uncached reads). Over six passes, one write and five reads of the stable prefix gives $1.515 + 5 x $0.411 = $3.57, against $7.65 uncached.

The $3.57 scenario requires one cache write and five reads. This hosted schema exposes no cache-control field to select Anthropic’s native five-minute or one-hour lifetime. Keep unchanged canon and manuscript material first and changing instructions last; changing an earlier section can reduce the matching prefix. Budget uncached, then compare the usage counts and the settled task cost with the estimate to see which categories applied. A short follow-up such as “fix paragraph 12” still carries all the history you resend as input.

Keep the bill down without starving the edit

Ask for findings before rewrites: a list of affected paragraphs identifies a small editing scope without paying to regenerate unaffected chapters. Request replacement passages rather than the original plus replacements when you already have the original locally. Keep a focused factual packet for local edits and reserve the whole manuscript for questions that need global evidence.

Do not shrink the packet past the point where the answer becomes guesswork. In the Mara example, dropping the key-location fact saves a few tokens and makes the error undetectable. The goal is the smallest complete evidence set for the question.

If cost is the main constraint and the packet is routine, compare neighbours before committing a big job. Sonnet 5.5 on SpicyAPI is $2 input and $10 output per million, a third lower per token than 4.6, with the same 1,000,000-token window and a 128,000-token output ceiling (5.5 endpoint). Anthropic estimates about 30% more tokens with the newer tokenizer, depending on the content. If this book and edit both increase by 30%, 400,000 input tokens become 520,000: $1.04 of input plus about $0.065 of output for the same edit, $1.105 against $1.275, roughly 13% lower rather than a third, before any extra thinking. Recount your tokens before budgeting. Choose 4.6 when you need its temperature, top_p and top_k controls or an existing prompt tuned on it; the version guide compares the contracts. Opus 4.6 costs $5 input and $25 output per million and permits 128,000 output tokens. Extra reply capacity alone does not justify its higher price when a cheaper route meets that limit; choose it when it resolves a specific editing failure.

FAQ

Is the full million-token window charged at a premium?

No. The bill rises when a request contains more input, produces more output or is repeated, not when it crosses 200,000 tokens.

Can it rewrite the entire book in one reply?

The hosted output ceiling is 64,000 tokens, including billed reasoning. Request a manuscript-wide findings report, then bounded chapter rewrites; preserve your original manuscript separately from the proposed replacements.

Is Sonnet 5.5 automatically cheaper for the same manuscript?

Its per-token rates are lower, but the tokenizer, reasoning and reply length can change the billed counts. The 13% saving above is an explicit scenario. Choose 4.6 when your integration needs its sampling controls or forced tools, and compare measured task bills when deciding on a migration.