Claude Opus vs Sonnet: Opus 5.5 and Sonnet 5.5 compared for writing.

1. Claude Opus vs Sonnet: the two current models, side by side.

As of October 3, 2026, Claude's developer docs list Opus 5.5 and Sonnet 5.5 as the current Opus and Sonnet. Opus 5.5 is the model the docs tell you to start with when you are unsure; Sonnet 5.5 is described as the best combination of speed and intelligence. Here they are on the specs that matter to a writer, from the models overview and each model's page.

Three things in that table changed in September. Sonnet caught up on knowledge: Sonnet 5 knew the world only to January 2026, five months behind Opus 5.5, and Sonnet 5.5 now matches its June 2026 cutoff. Opus got cheaper: Opus 5 was $5 and $25, and the release notes say Opus 5.5 costs 40% less to run. And Sonnet kept its price: $2 and $10, the same as Sonnet 5, so the price gap between the two is exactly two to one.

What the table cannot hold is behaviour. Opus 5.5 always thinks before it writes, at a default effort of medium on the API. Sonnet 5.5 defaults to high effort and thinks by default, but its lowest setting switches up-front thinking off entirely. That difference showed up in our test (section three): at their defaults, Sonnet thought for longer than Opus in two of the three pairs of runs. If you are choosing among all four current Claude models rather than these two, our guide to the best Claude model for writing covers Haiku and Fable too.

2. Is Opus better than Sonnet? What the published numbers say.

Yes, but the gap is the narrowest it has been. Three published sources, named so you can check them:

• The Sonnet 5.5 launch post reports a GDPval-AA score (a benchmark of real-world knowledge-work tasks) of 1844 for Sonnet 5.5 against 1846 for Opus 5.5, effectively a tie. The same post says Opus 5.5 “remains clearly stronger at complex, open-ended work requiring sustained judgment.”
• Artificial Analysis scored Sonnet 5.5 at 56 on its Intelligence Index on September 28, two points behind Opus 5.5 at max effort and second only to it.
• The Opus 5.5 launch post is the one that talks about writing directly: it says the model puts the most important information up front, and quotes a tester saying “it writes the way I do.” The Sonnet 5.5 post says Sonnet writes more clearly than the previous generation and produces polished documents.

Read together, the honest summary is this: on structured, well-scoped tasks Sonnet 5.5 is now within a rounding error of Opus 5.5, and on open-ended work that needs judgment Opus is still ahead. For writing, that splits cleanly. Drafting a meta description to a rule set is well-scoped. Deciding what a 2,000-word article should argue, what to cut from a draft, and which claim is too weak to keep is judgment. None of the numbers above is a writing benchmark, though, which is why we ran our own test rather than lean on them.

If you are also weighing a model from another vendor, the cross-vendor version of this comparison is Claude vs ChatGPT for writing.

3. The same article brief through both.

On October 3, 2026 we wrote one article brief and ran it three times on each model. The brief asked for a 1,000 to 1,300-word blog post titled “How to write a meta description that gets the click” for the readers of a small bookkeeping software company. It set a two-to-three-sentence lead, exactly five H2 sections in a fixed order, before-and-after examples for a plumber, a dog groomer and an accountant, three facts it was allowed to use (all from Google Search Central's guidance) with an instruction to add no others, no em dashes, and eight banned words and phrases. Each run had the same one-line system prompt, no tools, and each model's API default effort: medium for Opus, high for Sonnet.

Time is the API time for the whole response. Cost is the measured input (about 4,500 tokens: the brief plus the command-line tool's own wrapper) and output at list price, with thinking billed as output. On the hard rules it was a draw: all six drafts used the exact title, all five H2s in order with nothing after the last, stayed inside the word range, used all three examples, and contained zero em dashes, zero banned phrases and zero invented statistics.

The differences were in the softer rules, and they all went one way. Every Opus lead was two sentences, inside the limit. One Sonnet lead ran to four sentences against a limit of three. And two of the three Sonnet drafts added a claim the brief had not given them: one said a good description will not move a page up the results, another that the text appears when someone shares a link. Both are the kind of claim a reader would believe. Neither came from the three facts the brief allowed. We found none of that in the Opus drafts.

Three runs each is a test, not a benchmark, so treat it as a method you can copy. Write the brief the way our brief-first writing process describes (it works the same in Claude), list your hard rules, run it at least three times per model, and score each draft against the rules before you read it for style. Then check every claim the brief did not supply, because that is where both models will surprise you. That last step is exactly what an AI fact checker is for.

4. Speed and price per finished article.

Per token, Sonnet 5.5 costs half as much as Opus 5.5 and generates faster: in our runs it produced about 101 to 108 output tokens a second against 86 to 96 for Opus. Per article, the gap shrinks, because at its default effort Sonnet thought more. Its median run used 2,903 output tokens to Opus's 2,601. The medians:

• Sonnet 5.5: 27 seconds and about 3.8 cents per finished 1,200-word draft.
• Opus 5.5: 30 seconds and about 7 cents per draft, with one outlier run at 52 seconds and 11.9 cents when it thought for 2,877 tokens before writing a word.

So Opus costs a little under twice as much per article, and for a single article the speed difference is a few seconds. The Sonnet 5.5 launch post's own claims point the same way: more than 30% faster output than Sonnet 5 and up to 30% less per task, measured against the old Sonnet, not against Opus. If you need Opus and speed, its API has a fast mode in research preview at up to 2.5 times the output speed, at a premium price.

Two adjustments take these numbers toward real life. Nobody ships a first draft, so multiply by the passes you actually run (outline, draft, revision): a finished 2,000-word article on Opus 5.5 lands around 30 cents, on Sonnet 5.5 around 17. And anything that can wait goes through the Batch API at half price. Where the gap does add up is volume: rewriting meta descriptions and title tags across hundreds of pages is where Sonnet, or Haiku, earns its place. For one article a week, pick the better writer and stop optimizing: the model is the cheapest line in the budget.

5. Opus vs Sonnet for SEO content.

For articles meant to rank, use Opus 5.5 for the draft and the final pass and Sonnet 5.5 for the volume work around it. Then accept that past that choice, the model is the least important input. What decides whether a page ranks is the same on both, and neither model supplies it.

Neither has a keyword database, so any search volume either one quotes is a guess; the query has to come from measured data or your own Search Console, which our blog SEO guide walks through. Neither has seen your customers or your numbers, so the first-hand material has to come from you. And both share a June 2026 knowledge cutoff, so anything since then is either missing or invented. The test above showed the subtler version: a confident claim that was never in the brief.

Google is clear that the model is not the issue either way. Google Search Central says it focuses on the quality of content rather than how it was produced, and warns that using generative AI to publish many pages without adding value may violate its scaled content abuse policy. There is more ranking data on that in our AI SEO statistics. What Opus 5.5 changes for search work specifically, from keyword selection to publishing, has its own guide: Opus 5.5 for SEO. The Project, skills and connector setup that works with either model is in how to set Claude up for SEO, and if the goal is also to be the page Claude quotes in its own answers, that is a separate job: getting your pages cited by LLMs.

6. The verdict by writing job.

Which Claude model should you use? Route by the job, not by habit. This is where each model is worth its price for writing and SEO content, based on the published specs, the launch posts and what we saw in our own runs.

The pattern behind the table: Opus where a mistake is expensive and the work is open-ended, Sonnet where the output is short, structured or about to be rewritten. Switch at stage boundaries (outline on Sonnet, draft and edit on Opus), never in the middle of one draft, or the voice changes at the seam. For metadata, compare what you get against a purpose-built meta description generator that checks length against what Google actually shows.

7. Where Haiku and Fable fit.

Haiku 4.5 is the bulk option and nothing more for a writer. At $1 and $5 per million tokens it is the fastest and cheapest current model, which makes it good for hundreds of meta descriptions or summaries through the Batch API. It is a generation older than the two models on this page: its reliable knowledge stops at February 2025, its context window is 200,000 tokens, and the Sonnet 5.5 launch post says Haiku 5.5 will join the 5.5 family in the coming weeks.

Fable 5.1 sits above Opus at $10 and $50 per million tokens with the slowest speed label in the lineup. Claude's docs reserve it for demanding reasoning and long-horizon work, and the Opus 5.5 release notes say Opus now performs at Fable 5.1 level on most work, so for articles it is rarely worth the price. The one writing job where it can be is a single long document built from a large research pack. The full four-model breakdown is in the Claude lineup by writing job, and the ChatGPT equivalent is our best ChatGPT model for writing guide.

A worked example: one week of content, routed by job.

Take a fictional invoicing app, Tallypine, publishing three articles a week. The company and its log are invented; the per-token prices are the real ones from section one, and the per-draft costs are scaled from our measured runs in section three, before retries. Here is one week routed by the verdict table:

The arithmetic: a 2,000-word Opus draft is about 4,000 output tokens plus a 6,000-token brief with sources, roughly 10 cents, and three passes per article make 30. An outline on Sonnet is about 2 cents. A meta description through the Batch API on Sonnet is about a third of a cent. The week comes to roughly $1.45 of model time. Running everything on Opus would add about 45 cents; running everything on Sonnet would save about 40 cents and put the long-form drafts on the model that drifted from the brief in our test.

Now the other column. Tallypine's editor logged about nine hours that week: writing three briefs, checking every claim in three drafts (the step that caught a made-up integration and a pricing figure from last year), the edit, and publishing. That ratio is the real answer to Opus vs Sonnet. Pick the right model per job and you have optimized the cheapest line in the budget; the hours around it are where the cost and the quality both live.

Common mistakes when choosing between them.

1. Judging on one blank prompt. “Write a blog post about X” once in each window is the usual test and the least useful one: both return the average of the web, and one run tells you nothing about variance. Our Opus runs ranged from 27 to 52 seconds on the same brief. Use your real brief, three runs each, and score against your rules.
2. Assuming Sonnet is always faster. It generates faster per token, but at its default high effort it can think longer than Opus does at medium. If speed is the point, turn Sonnet's effort down for short, well-specified jobs, which is what its own docs suggest for latency-sensitive work.
3. Comparing the old pair. Most Opus vs Sonnet advice online predates September 2026, when Sonnet 5 knew the world only to January 2026 and Opus cost $5 and $25. Check the version before you trust the verdict, including on this page (it is dated at the top).
4. Trusting what either model adds. Two of our six drafts slipped in a claim that was not in the brief, and both sounded right. Put the facts in the brief with their sources, and treat anything else as unverified until you open a source yourself.
5. Switching models mid-draft. Finishing a section on whichever model is faster changes the voice at the seam. Keep your style rules in a Claude Project that holds your style guide and switch only between stages.
6. Upgrading the model to fix a volume problem. When output is thin at ten articles a week, the problem is checking time per article, and it scales with people, not models. That is the case for autoblogging with a quality gate rather than without one.