Claude Opus 5.5 for Blog Writing: What Changes for Your SEO Drafts

Claude Opus 5.5 costs less per token than Opus 5, reads 1M tokens of context and can't turn off its thinking. Here's what that means for drafting, voice matching and fact handling, and how Claude Skills keep your site's voice consistent.

Claude Opus 5.5 is Anthropic's newest flagship model. It came out on September 22, 2026, costs $4 per million input tokens and $20 per million output tokens, and can read a 1M-token context window. For blog writing, three things matter most. It holds a full brief, your style guide and your research in a single prompt. It is cheaper per token than Opus 5. And its knowledge stops at June 2026, so anything newer has to come from the notes you give it.

YouTube reviews of Opus 5.5 are mostly about coding benchmarks. This post covers publishing instead: drafting, voice matching, fact handling, and how to use Claude Skills to keep every draft sounding like your site.

Why "Claude Opus 5.5" is spiking this week

The model is a month old. The new search interest seems to come from products built on it that launched this week. That's our reading of the dates. Nobody has confirmed it.

  • October 21: Claude Duet. A proactive agent app that runs on Opus 5.5. It's a waitlisted research preview for US Pro and Max users. It uses the skills and memory already in your Claude account, and its background checks count toward your plan limits.
  • October 22: Claude on Google Cloud (Anthropic-managed) went GA, and Fast mode for Opus 5.5 is generally available there too.
  • The 5.5 family is complete. Sonnet 5.5 shipped September 28 and Haiku 5.5 on October 7. So people are now asking which one to use.

Searches for "claude skills" are climbing at the same time. That fits: Duet uses your skills, and Anthropic updated its skill-writing best practices in early October.

Claude Opus 5.5 key facts

SpecOpus 5.5
Release dateSeptember 22, 2026
API model IDclaude-opus-5-5 (Bedrock: anthropic.claude-opus-5-5)
Context window1M tokens
Max output128K tokens (300K on Batch API with a beta header)
Knowledge cutoffJune 2026
Price$4 input / $20 output per million tokens
Fast modeUp to 2.5x faster, $8 / $40 per million tokens
Batch API50% off
ThinkingAdaptive, always on, controlled by effort (default: medium)
AccessPro, Max, Team, Enterprise, API, Claude Code, AWS, Google Cloud, Microsoft Foundry. Not on Free.

What Opus 5.5 changes for blog drafting

Problem: your brief doesn't fit in one prompt.

You have a style guide, three example posts, a keyword list, research notes and a list of internal links. Older workflows cut that down to fit, and the voice got lost along the way.

The fix: 1M tokens of context is enough for all of it, plus a dozen of your best-performing posts as voice samples. You don't have to summarise or cut anything.

Problem: you can't switch thinking off.

Adaptive thinking is always on in Opus 5.5. For a 600-word news brief, that can mean paying for reasoning you don't need.

The fix: use the effort setting instead. Keep it at medium (the default) or go lower for quick rewrites, headlines and meta descriptions. Save max for long explainers where structure and accuracy matter. Effort changes your bill, which the next section shows.

Problem: "40% cheaper" vs "20% cheaper."

You'll see both claims, and both are true:

  • Per token: $4/$20 against Opus 5's $5/$25 is a flat 20% cut.
  • Per workload: Anthropic says typical workloads cost about 40% less, because the model gets tasks done more efficiently.
  • Per task at max effort: MyClaw measured about 119K output tokens per task against roughly 73K for Opus 5. In that test, cost per task came out about even.

For publishers, the takeaway is simple. Your savings depend on your effort setting, not the price sheet. Run your own drafts and compare before you budget.

Here's a rough example. A draft with 20K input tokens (brief plus notes) and 15K output tokens (thinking plus a 1,500-word post) costs about $0.38 on Opus 5.5. The same job costs about $0.19 on Sonnet 5.5. These are illustrative numbers. Your token counts will vary.

Problem: the model doesn't know last week's news.

Opus 5.5's knowledge stops in June 2026. If you ask it about a product, match or policy from after that, it will either say it doesn't know or fill the gap with something plausible.

The fix: give it dated research notes and tell it to use only those for anything recent. Add a rule: "If the notes and your knowledge disagree, trust the notes." This matters most if you run a news portal or cover anything time-sensitive.

Problem: more refusals.

Anthropic tightened its safety classifiers, so Opus 5.5 refuses more often than earlier Opus models. Most cybersecurity tasks get rerouted to Opus 4.8, and biology topics have stricter safeguards. If you run a security or health blog, you'll notice. Anthropic runs verification programs for cyber and life-sciences users.

Two other things are worth knowing. Opus 5.5 is better at resisting prompt injection, which helps when you paste in scraped pages. And its output is watermarked for EU AI Act compliance.

How to test Opus 5.5 on your own SEO drafts

Benchmarks like Terminal-Bench 4.0 (66.4% vs 52.3% for Opus 5) measure coding agents. They tell you nothing about whether a draft sounds like your site. Run a test on your own content instead:

  1. Pick five briefs you've already published. Mix them: one news post, one explainer, one listicle, one comparison, one evergreen guide.
  2. Give each model the same inputs: style guide, research notes, keyword list. Run Opus 5.5 at medium and max effort, and Sonnet 5.5 at default.
  3. Score voice blind. Hide which model wrote what. Ask an editor whether each draft sounds like you.
  4. Check every fact against the notes. Count any claim that isn't in the notes and isn't common knowledge.
  5. Record tokens and cost per draft. The usage figures in the API response tell you more than the price sheet.
  6. Time the edit. Minutes from draft to publish is the number that actually matters to your desk.

If you're an SEO agency, run this once per client voice. A model that nails one brand can still flatten another.

Claude Skills: keep your voice consistent with Opus 5.5

Pasting your style guide into every chat stops working once you have more than one writer, or more than one site. Claude Skills fix that.

What a Claude skill is

A skill is a folder built around a SKILL.md file, with optional scripts and reference files. Claude loads it automatically when a task matches. Skills launched in October 2025 and work on every plan, including Free. Built-in skills cover Excel, Word, PowerPoint and PDF, and Notion, Figma, Atlassian and others offer partner skills.

How to set up a voice skill

  1. Turn on code execution: Settings > Capabilities (Team and Enterprise admins do this in org settings).
  2. Create a folder with a SKILL.md at the top. Give it a clear name and a description such as "Use when drafting or editing blog posts for [your site]."
  3. In SKILL.md, write your voice rules: person (you/we), sentence length, banned phrases, header style, how to handle numbers and dates.
  4. Put reference files one level below SKILL.md: voice-samples.md with three to five of your best posts, and a sources-policy.md.
  5. Zip the folder and upload it under Customize > Skills.

Follow the October best-practice update

  • Test the skill on every model it'll run on. A skill tuned on Opus 5.5 may behave differently on Sonnet 5.5 or Haiku 5.5.
  • Keep reference files one level deep. Don't nest folders inside folders.
  • Start long files with a contents list so Claude can find the right section.
  • End with a checklist that includes a verification step. For example: "Before finishing, confirm every date and number appears in the research notes."

The verification step does the most work. It turns "write in our voice" into "write in our voice and check yourself." Duet reads the skills in your account, so a well-built voice skill carries over to that too.

Opus 5.5 vs Sonnet 5.5 vs Haiku 5.5 for writing

ModelPrice (input/output per M)Best writing job
Opus 5.5$4 / $20Long explainers, comparisons, tricky voice matching, multi-source synthesis
Sonnet 5.5$2 / $10Daily news posts, standard SEO drafts at volume
Haiku 5.5From $0.10 / $0.50Titles, meta descriptions, FAQs, tagging, rewrites
Fable 5.1$10 / $50Anthropic says Opus 5.5 performs at about this level on most tasks, at well under half the price

Many desks will end up with a split. Sonnet 5.5 for volume, Opus 5.5 for the posts that need to rank on hard queries, Haiku 5.5 for metadata. Affiliate sites doing comparison content will probably lean on Opus. Solo bloggers paying per token may find Sonnet good enough.

Developers: what breaks when you switch from Opus 5

  • Thinking can't be disabled.
  • Forced tool use returns an error.
  • Thinking blocks are tied to the model and conversation that produced them.
  • The computer_20251124 tool is rejected.
  • Text between tool calls now comes back inside thinking blocks, which are empty at the default display setting. If your pipeline parsed that text, check it.

How WriteOut fixes it

A better model doesn't solve the hard part of trend publishing: knowing what to write and getting it out first. WriteOut spots Google searches just starting to rise in your niche. It builds dated research notes so the model isn't guessing past its cutoff, drafts in your site's voice, and publishes to WordPress, static sites or any CMS once you approve. You choose the topic, check the draft and ship it.

Learn more →

Frequently asked questions

What is Claude Opus 5.5?

Claude Opus 5.5 is Anthropic's flagship model and the first in the Claude 5.5 family. It was released on September 22, 2026. It has a 1M-token context window, up to 128K output tokens and a June 2026 knowledge cutoff. Adaptive thinking is always on and you control it with an effort setting. It costs $4 per million input tokens and $20 per million output tokens.

Is Claude Opus 5.5 free?

No. Opus 5.5 isn't on the Claude Free plan. You can use it on Pro, Max, Team and Enterprise plans, with paid plans starting at $20 a month. You can also use it through the API, Claude Code, AWS, Google Cloud and Microsoft Foundry. Claude Skills do work on the Free plan, but with other models.

Is Opus 5.5 cheaper than Opus 5?

Per token, yes. It's 20% cheaper, at $4/$20 against $5/$25 per million tokens. Anthropic says typical workloads cost about 40% less. But one independent test at max effort found Opus 5.5 used far more output tokens per task, so cost per task came out roughly even. Your effort setting decides how much you actually save.

What are Claude skills?

A Claude skill is a folder with a SKILL.md file and optional scripts or reference files. Claude loads it automatically when a task matches. Publishers can use skills to store voice rules, sample posts and fact-checking checklists, so drafts stay consistent. You upload a skill as a ZIP under Customize > Skills, with code execution turned on.

What is Claude Duet?

Claude Duet is a proactive agent app built on Opus 5.5. It launched on October 21, 2026, as a waitlisted research preview for US Pro and Max users. It uses the skills and memory in your Claude account. Its background checks count toward your plan's usage limits, so heavy use can eat into your other Claude work.

Why does Opus 5.5 refuse more requests?

Anthropic gave Opus 5.5 stricter safety classifiers than earlier Opus models. Most cybersecurity tasks are rerouted to Opus 4.8, and biology topics have stronger safeguards. If you publish security or life-sciences content, you may get more refusals. Anthropic runs verification programs for professional users in those fields.

Sources

  1. Introducing Claude Opus 5.5 · Anthropic
  2. Opus 5.5 overview · Claude Platform Docs
  3. What are skills · Claude Help Center
  4. Claude Skills best-practices update · MindStudio
  5. Opus 5.5 Review · MyClaw
  6. Anthropic Claude Duet · MacRumors
  7. Anthropic opens Claude Duet, an Opus 5.5 agent, to US Pro and Max · AI Weekly
  8. Claude updates Oct 2026 · Releasebot

Keep reading.

All posts →

WriteOut is in early access.

Leave your email and the site you'd use it on. We're letting people in one site at a time.