Best Claude Model for Writing (2026): Opus 5.5 Leads, and the Lead Is Not Settled

Short version, before the tables: the newest Opus has the highest Claude score for creative writing, and the lead is not settled. On LMArena's style-controlled creative writing leaderboard, snapshot October 2, 2026, Claude Opus 5.5 at high effort scores 1516. Its confidence interval overlaps Claude Fable 5 at high effort and Claude Opus 4.6 at high effort, and the Opus 5.5 row rests on 1,058 votes against 8,182 and 14,085 for those two. The newer Fable, Fable 5.1, scores below Fable 5. The newer Sonnet, Sonnet 5.5, does not have enough votes to call.

None of that is what the names suggest. Anthropic's current advice is to start with Opus 5.5 for most workloads, and to move up to Fable 5.1 when harder reasoning still falls short. The official use cases are agentic coding, knowledge work, and long-horizon tasks. Creative writing is not among them. The preference votes landed there anyway.

We build Muses, a writing workspace, and we track and integrate frontier models continuously because the quality of a draft depends on which model is behind the cursor. Everything below comes from public sources: the LMArena leaderboard dataset, LiveBench, and Anthropic's model documentation. Where the data runs out, on long-manuscript consistency and on which model refuses least, this page says so rather than filling the gap.

Short answer: which Claude model is best for writing?

If you want one sentence to carry away: in October 2026 the best Claude for a first draft is Opus 5.5 with effort set to high, Fable 5 if you would rather trust the score with thousands more votes, Opus 4.8 when the job is revision and you are paying Opus prices, and Sonnet 4.6 when you write every day. Test that shortlist on your own pages. It is not the release order, and it is not "always buy the newest."

The 2026 Claude lineup for writers

The current family, from Anthropic's model overview. If you are unsure, Anthropic says start with Opus 5.5, and use Fable 5.1 when your own tests at higher effort on Opus 5.5 still fall short.

Still available, and listed by Anthropic as legacy rather than current: Fable 5, Opus 5, Opus 4.8, 4.7, 4.6 and 4.5, Sonnet 5, and Sonnet 4.6. Opus 5 and Opus 4.8 through 4.5 price at $5 and $25, above Opus 5.5. Fable 5 prices with Fable 5.1, at $10 and $50. Sonnet 5 prices with Sonnet 5.5, at $2 and $10. Sonnet 4.6 stays at $3 and $15, which is more than the newer Sonnet, not less. Sonnet 4.5 is the one that has left this list: deprecated on September 30, 2026, still callable, retiring November 30.

One detail before you size a manuscript against a context window. Anthropic changed its tokenizer with Opus 4.7. On the current tokenizer, 1M tokens is roughly 555,000 English words; on the older one, the same 1M held about 750,000 words. Anthropic's pricing documentation puts the increase at approximately 30% more tokens for the same text. Fable 5.1, Opus 5.5, Sonnet 5.5, Fable 5, Opus 5, Sonnet 5, Opus 4.8 and Opus 4.7 all use the new tokenizer. Sonnet 4.6, Opus 4.6, Sonnet 4.5 and Haiku 4.5 use the old one. A 1M window on Opus 5.5 holds meaningfully less of your novel than a 1M window on Opus 4.6, despite the identical headline number.

What the style-controlled leaderboard actually shows

LMArena collects anonymous pairwise human preference votes and fits a Bradley-Terry rating. Style control adjusts for the formatting habits that make responses win votes without being better prose, such as length and heavy use of headers and bullets. The figures below come from the text arena's style_control subset, creative_writing category, in the LMArena leaderboard dataset, snapshot dated October 2, 2026. Ratings and intervals are rounded to the nearest integer from that file.

The effort labels are the part that decides whether the row is what you will actually run.

Opus 5.5's row is high. Its API default is medium. Omitting the parameter does not reproduce this rating. Fable 5, Opus 4.6, Opus 4.7, Opus 4.8 and Sonnet 5 are shown at high, which is the default on those models. Fable 5.1 appears only as max, one step above its default of high, so the out-of-the-box Fable 5.1 is not the row in the table. Sonnet 5.5 appears only as xhigh, above its default of high, from 658 votes. Sonnet 4.6, Sonnet 4.5 and Haiku 4.5 have no effort suffix on the board. Haiku 4.5 does not support the effort parameter at all.

Opus 5.5's first place among Claude models is a point estimate, not a separation. The interval 1497–1536 overlaps Fable 5 (1495–1510) and Opus 4.6 at high (1494–1507). It does clear Opus 4.7 at high (1482–1496) with no overlap, which is a real step past the model this page recommended for long prose in August. It also has company outside the family: Gemini 4 Argon at high effort scores 1519, rank 1, interval 1499–1538, and that interval overlaps Opus 5.5 too. This article is not a cross-vendor comparison. Claude is not alone at the top.

Fable 5.1 does not inherit Fable 5's place. The only published row is max effort, 1469–1494, and Fable 5 at high is 1495–1510. Those intervals do not overlap. That is a reversal of what "newer, and at max effort" would suggest, with the caveat that the settings differ and Fable 5.1's default of high is not on the board.

The Sonnet ordering is quieter than it was in August, and one model in it is leaving. Sonnet 4.5 at 1454 (1448–1460) still clears Sonnet 5 at high, 1435 (1428–1442), with no overlap. Sonnet 4.5 is deprecated and retires November 30, 2026, so that finding is not a model you should adopt now. Sonnet 4.6 at 1449 (1443–1456) also sits clear of Sonnet 5, and the bounds are one point apart, 1443 against 1442. Real, and thin. Sonnet 5.5's interval, 1424–1471, covers Sonnet 4.6, Sonnet 5 and Sonnet 4.5. Six hundred fifty-eight votes cannot settle it.

Opus 4.7 at high, 1482–1496, clears Opus 4.8 at high, 1463–1476, with no overlap. The same is true of the plain rows, if you are looking at the ids without a -high suffix: Opus 4.7 at 1483 (1476–1490) clears Opus 4.8 at 1463 (1457–1470). Among the models that still cost $5 and $25, that older reversal stands. Opus 5.5 at high clears Opus 4.7 at high as well, and costs less.

Preference ratings are one axis. LiveBench, question set 2026-06-25, with model rows read on October 7, 2026, scores language and instruction following separately, and they disagree. The rows below are the effort the published table actually names.

On language, Fable 5 at max effort leads the Claude rows at 90.7. Fable 5.1 at max is 89.5. Opus 5 at max is 88.7. Opus 5.5 at max is 86.3. The newest Opus is not the language leader, and it sits behind the Opus it replaced.

On instruction following the order scrambles. Gemini 3.8 Flash at high effort leads the published list at 81.4. The highest Claude is Fable 5 at max, 75.8, then Fable 5.1 at max, 73.0, then Opus 4.8 at max, 72.0. Opus 5.5's higher published row is xhigh at 67.0, and its max row is 65.7. Opus 4.7 at xhigh is 66.7. Opus 5 at max is 63.8. Sonnet 5.5 is 70.5 at xhigh and 56.8 at max, which is a large swing and another reason not to treat one Sonnet 5.5 number as the model. There is no single number for "good at writing."

The limits are the same ones as before, with a wider spread. Arena ratings measure what anonymous voters preferred in short pairwise comparisons, not what serves your book. Vote counts in the table run from 658 to 24,158. Opus 5.5's rating rests on 1,058 votes. Use the intervals, not the point estimates alone, to narrow a shortlist, then test the shortlist on your own pages.

No, the writing leader is not Anthropic's creative-writing model

Fable 5's name used to invite that assumption. Opus 5.5's place at the top of the writing board invites a different one: that the model Anthropic tells you to start with was tuned for prose. The documentation says otherwise.

The model overview says to start with Claude Opus 5.5 for most workloads, and to use Claude Fable 5.1 for demanding reasoning and long-horizon agentic work, or when your own evals on Opus 5.5 at higher effort still fall short. The Opus 5.5 model page describes it as built for long-running agentic coding and knowledge work. The Fable 5.1 model page describes stronger long-running agentic coding, multistep research, and document work, at the same input and output prices as Fable 5. Creative writing appears in none of those use cases.

Positioning and performance still come apart, just not in the August shape. In August the model at the top of the writing board was the expensive general model, Fable 5, released June 9, 2026. In October the highest point estimate belongs to the model Anthropic already tells most people to start with, Opus 5.5, released September 22, 2026, at a lower price than Fable and a lower price than Opus 5. That is a useful coincidence. It is still a coincidence. Fable 5.1, released September 1, 2026, is the capability step up in Anthropic's own guide, and on this board it does not write better than Fable 5.

Drafting, editing, and long manuscripts are three different jobs

Treating "writing" as one capability is what produces bad model choices. At minimum it is three.

Drafting rewards whatever people actually prefer reading, which is what the arena measures. For new prose from a prompt or an outline, the creative writing board points at Opus 5.5 at high effort, with Fable 5 and Opus 4.6 at high inside the same interval. If you will not set Opus 5.5 to high, you are not running the configuration that earned 1516.

Editing is a different skill. Revision means following instructions exactly: keep the voice, cut 15%, do not touch dialogue, apply this style sheet. That is instruction following, and LiveBench reshuffles the prose order. Fable 5 at max effort leads the Claude rows at 75.8. Fable 5.1 at max is 73.0. Opus 4.8 at max is 72.0. Opus 5.5 is 67.0 at xhigh and 65.7 at max, behind Opus 4.8 on both published rows, at a lower token price. Which model is "better" depends on which of these jobs you are handing it. The LiveBench rows are max or xhigh, not a matched set of API defaults, so treat the gap as a place to start a test, not as a result you have already replicated.

Long manuscripts depend on context and on consistency across a session, and this is where the public data runs out. Context capacity is documented: 1M tokens, about 555,000 words on the current tokenizer, which is the one Opus 5.5, Fable 5.1 and Sonnet 5.5 use. Consistency is not. No public benchmark measures whether a model's voice drifts across a 60,000-word project, whether it forgets a character's speech pattern by chapter 20, or whether it silently starts flattening your syntax after the fortieth revision. We have opinions from working with these models. Opinions are not measurements, so they are not findings. We are building structured tests for exactly this, and when there are results worth publishing they will be added here.

What a four-turn drafting session actually costs

The following is a worked model, not a measurement. We have not benchmarked these sessions. We are applying Anthropic's published prices to a plausible session shape so the relative costs are visible. Your token counts will differ.

On the older tokenizer, 1,400 output tokens is roughly 1,050 words. On the current one it is closer to 780. Applying published per-MTok rates:

The words-produced column does not compare across the tokenizer change. The same 5,600 output tokens buy about a third more English words on the older tokenizer. Opus 4.6, still on that tokenizer and still priced at $5 and $25, gets about 1,050 words out of the same $0.22 output budget as Opus 5. Read cost against words within a tokenizer generation.

Two adjustments change the picture. The first is prompt caching. Cache reads bill at 10% of the base input price on most of these models, at 5% on Opus 5.5 ($0.20 per million tokens), and at 2.5% on Fable 5.1 ($0.25 per million tokens). Fable 5 stays at the 10% rate ($1 per million). Iterative work on a long draft is close to the ideal case, but only for the part of the prompt that never moves: the system prompt, the source material you loaded once, and the conversation history you append to rather than edit. A rewritten draft pasted back near the top of the prompt no longer matches the cached prefix. The same text appended at the end leaves the stable portion intact. Anything else that mutates the prefix, a timestamp, a reordered note, a changed system instruction, throws the discount away. Anthropic's Batch API is a second lever at 50% off both directions. It is asynchronous, and useless for interactive drafting. The $1.50 and $7.50 figures you may see next to Sonnet 4.6 on a leaderboard are that batch price, not the base rate of $3 and $15.

The second adjustment runs the other way. Thinking tokens bill as output. On Fable 5.1 and Opus 5.5, adaptive thinking is always on and cannot be switched off, so effort is the control for how much of it you pay for. A hard structural problem will produce more thinking than a simple rewrite. The output column above assumes a straightforward drafting turn. Watch your first week of real usage rather than trusting any published estimate, including this one.

The effort setting: the lever most people ignore

Anthropic's effort guide still has the line that deserves more attention than it gets: tuning effort is often a better lever than switching models. Effort is a request parameter. Setting it to the model's default produces the same behavior as omitting it.

That default is no longer one value. Opus 5.5 defaults to medium. Every other current model that supports effort defaults to high. Haiku 4.5 does not support the parameter. Fable 5.1, Fable 5, Opus 5.5, Opus 5, Opus 4.8, Opus 4.7, Sonnet 5.5 and Sonnet 5 take five levels, low through max. Opus 4.6 and Sonnet 4.6 take max and do not list xhigh.

For writing, effort maps onto something recognizable. Higher effort buys more planning before the first sentence: better structural decisions, more consistent handling of a long brief, more attention to what you asked for. It also costs more and takes longer. Lower effort produces faster turns with less deliberation behind them.

The practical setting depends on which model you picked, because the leaderboard rows are not all defaults.

On Opus 5.5 the creative-writing rating is the high row, one step above the default. The move that matches the board is stepping up, not leaving the parameter off. Step down to low only for quick rewrites you have already decided to throw away, and step up to xhigh for structural work: reorganizing a chapter, fixing a section that keeps collapsing, working through a brief with many constraints. Anthropic's own note on this model is that a request which omits effort runs one level lower than it did on Opus 5.

On Fable 5, Opus 4.7, Opus 4.8 and Sonnet 5, high is the default, and it is the row in the table above. The move there is still stepping down for a fast first pass, and stepping up when the structure is the problem. Anthropic notes that lower effort on Fable 5 still performs well and often beats xhigh on earlier models.

Fable 5.1's only creative-writing row is max, above its default of high. Sonnet 5.5's only row is xhigh, above its default of high. Running either at the default means you are not running the configuration in the table. We have no measured comparison of writing quality across effort levels, so treat all of this as a starting point to test.

Refusals, tone, and the "Sonnet got colder" complaint

Two things are frequently conflated when writers complain about a model: refusing a request and being cold about fulfilling it.

On tone, the community complaint that Sonnet 5 reads as drier than Sonnet 4.5 is still anecdote. What earns it a second mention is that the October board points the same way: Sonnet 5 at high sits 19 points below Sonnet 4.5, 1435 against 1454, with no interval overlap. Sonnet 4.5 retires on November 30, 2026, so the model you would switch back to is on a clock. Sonnet 4.6, which stays, also sits clear of Sonnet 5, with bounds one point apart. Anecdote plus a matching number is not proof. It is more than vibes, and it is not a reason to adopt a model that is already deprecated.

On refusals, the documented fact is wider than it was in August. Anthropic states that Claude Fable 5.1, Claude Fable 5, Claude Opus 5.5, Claude Opus 5, and Claude Sonnet 5.5 include safety classifiers that can decline a request. Opus 5.5's classifiers cover a broader set of categories than Opus 5's, including biology, cybersecurity, and reasoning extraction. Beyond that, no public benchmark measures refusal rates on creative writing tasks, so a ranking of which model refuses least would be invented. We are not going to invent it.

What does help is framing. Say the work is fiction and name the form. Give the genre and the intended readership. Explain what a difficult scene is doing in the story: what it costs the character, what it sets up. Ask for the level of detail the scene actually needs rather than the maximum. These are the things a human editor would want before reading a hard chapter. They are the context that makes the request legible.

Which Claude for which writer

Budget first. Sonnet 4.6 at $3 and $15 per million tokens. On the creative writing board it sits just clear of Sonnet 5, which costs $2 and $10. If the bill matters more than that thin edge, Sonnet 5 is the cheaper measured option. Sonnet 5.5 is the current Sonnet and the same price as Sonnet 5, and 658 votes cannot tell you whether it writes better or worse. Haiku 4.5 is cheaper again and ranks 133rd.

Novelists and long-form nonfiction. Opus 5.5, with effort set to high if you want the configuration on the board. It clears Opus 4.7 at high with no interval overlap and costs $4 and $20, against $5 and $25 for Opus 4.7. Budget for the new tokenizer when you plan how much manuscript to keep in context. Fable 5 is the upgrade if you want the score that rests on 8,182 votes rather than 1,058, at two and a half times the token price of Opus 5.5. Among the models that still cost $5 and $25, Opus 4.7 still clears Opus 4.8.

Daily content and high-volume drafting. Sonnet 4.6 as the default, with Opus 5.5 at high reserved for the pieces that matter. At roughly $0.13 per drafting session, a piece a day for a month is a rounding error against the time it saves. The model choice matters much less than whether you actually revise the output. If all you want is a first pass to react to, describe it in our AI text generator and it opens as a draft in your workspace, with no API setup at all.

Editors and rewriters. Fable 5, on the LiveBench instruction-following gap, 75.8 at max effort against Opus 5.5's 65.7 at max and 67.0 at xhigh. If the budget stays at Opus prices, Opus 4.8 at max effort, 72.0, is the published row that still leads Opus 5.5. Opus 4.8's API default is high, and LiveBench does not publish a high-effort row for it, so the 72.0 is one step above what you get by omitting the parameter. If your job is applying a style sheet, holding a voice steady, and making the cuts you specified, start there and check it on a chapter you already know well.

Most of the friction in a writing day is not the model. It is the draft in one tab, the research in another, the outline in a third, and no version history over any of it. That is what we built Muses to fix: a writing workspace where the draft, the notes, the source material and AI-assisted revision live in one project, with the history to walk a bad revision pass back. The models behind it are chosen and kept current by the platform, so none of the per-model tuning above is something you have to configure yourself. Try it free.

Model deprecation: protect your writing workflow

Model recommendations expire. Several of the picks above are older models, and one model that used to be a finding in this article now has a retirement date.

Sonnet 4.5 was deprecated on September 30, 2026, and retires on November 30, 2026. The recommended replacement is claude-sonnet-5-5. It still outscores Sonnet 5 on the creative writing board. That is not a reason to build a workflow on it. It is the only model discussed here with a fixed retirement date.

Haiku 4.5 is still listed as active. The deprecations page gives a floor of not sooner than October 15, 2026, and it does not list a deprecation. Anthropic's rule for a publicly released model is at least 60 days' notice before retirement. A retirement on October 15 would already have required that notice. It is not on the page, so October 15 has not become a scheduled retirement. It remains the high-volume pick with the shortest published floor, and the next floor among models this page still mentions is Opus 4.5, not sooner than November 24, 2026.

The rest have more room. Sonnet 4.6 runs to February 17, 2027. Opus 4.6 runs to February 5, 2027. Opus 4.7 runs to April 16, 2027. Opus 4.8 runs to May 28, 2027. Sonnet 5 runs to June 30, 2027. Fable 5 runs to June 9, 2027. Opus 5 runs to July 24, 2027. Fable 5.1 runs to September 1, 2027. Opus 5.5 runs to September 22, 2027. Sonnet 5.5 runs to September 28, 2027. Those are "not sooner than" dates, not retirement announcements.

If you call the API directly, pin the exact model ID. Anthropic notes that every Claude model ID is a pinned snapshot, including the dateless IDs used from the 4.6 generation onward, so claude-opus-5-5 names one specific model and will keep naming it. Writers using Claude through an application do not have that control, which is a reason to check what your tool is actually calling before you conclude a model has changed under you. When these models turn over, this page gets updated.

How we keep this page current

Last verified: October 7, 2026.

We review this page monthly and check three things each time: the model lineup and IDs, published pricing, and the leaderboard snapshot. When a rating moves enough to change a recommendation, the recommendation changes with it, and the update date at the top of the page moves. The August version of this page said the newest Claude was not the best Claude for writing. The October snapshot no longer supports that sentence, so the sentence changed.

Sources, all public:

• LMArena leaderboard dataset — text arena, style_control subset, creative_writing category, snapshot dated 2026-10-02. Bradley-Terry ratings from anonymous pairwise human preference votes. Integers in the table are rounded from that file.
• LiveBench — question set 2026-06-25. Language and instruction-following category averages, read October 7, 2026, with the effort named on each row.
• Anthropic model overview — context windows, output limits, pricing, thinking modes, default effort, knowledge cutoffs, tokenizer note.
• Anthropic pricing — per-MTok rates, prompt caching multipliers, Batch API discount.
• Choosing the right model and Effort — effort levels and the effort-versus-model-switching guidance.
• Refusals and fallback — which current models ship safety classifiers.
• Model deprecations — lifecycle status, retirement dates, notice policy.

If you want the craft side rather than the model side, our guide to getting better at writing covers the draft-distance-feedback-revise cycle that any of these models slots into, and our reading of On Writing covers where the machine should and should not be in the room.