Claude Opus 5.5 for marketers: what changed, what it costs, which jobs it suits

What Anthropic released on September 22, 2026

Claude Opus 5.5 is the first model in Anthropic's Claude 5.5 family, released on September 22, 2026. Anthropic says it performs at the level of Claude Fable 5.1 on most work and, at default settings, costs 40% less than Opus 5 on typical workloads. Its API ID is claude-opus-5-5. The details are in Anthropic's announcement.

Opus 5.5 has a 1 million token context window, writes up to 128,000 tokens in one response and has a reliable knowledge cutoff of June 2026, according to Anthropic's models overview. The same page now tells developers to start with Opus 5.5 for most workloads, and to use Fable 5.1 for demanding reasoning and long agent runs or when Opus 5.5 at higher effort still falls short. The model is on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry, and in the Claude apps on paid plans. Anthropic says Claude Sonnet 5.5 and Claude Haiku 5.5 follow in the coming weeks.

Three changes matter most for marketing work, as Anthropic describes them:

• Writing. Opus 5.5 puts the most important information first, uses less jargon and follows the writing rules you give it. Anthropic says writing was one of the most common pieces of feedback on Opus 5.
• Research that holds up. In an internal test, models wrote a report on a company's quarterly results from a copy of the web, and an automated grader failed any report with an invented figure or quote. 16 of 18 Opus 5.5 reports passed; no Opus 5 or Fable 5.1 attempt did.
• Cheaper long runs. Anthropic says Opus 5.5 uses fewer tokens per task than Opus 5, which, with the lower price, nets out to about 40% lower cost.

For people on a Claude subscription, Anthropic also raised the five-hour usage limits on the Pro, Max, Team and seat-based Enterprise plans.

What Opus 5.5 costs, next to GPT-6

Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens on the Claude API, 20% less than Opus 5's $5 and $25. A read from the prompt cache costs $0.20 per million tokens, 5% of the input price, and the batch API halves both prices to $2 and $10, according to Anthropic's API pricing page.

Claude prices from Anthropic's API pricing page and the Opus 5.5 announcement. GPT-6 prices are the standard, short-context rates on OpenAI's API pricing page. Cached input is a read from the prompt cache; writing to Anthropic's cache costs more than a normal input token, $5 per million for Opus 5.5 at the five-minute duration.

At list prices for input and output, Opus 5.5 costs twice what GPT-6 Sol and Claude Sonnet 5 cost, and 40 times what GPT-6 Luna costs. Anthropic's case is that it finishes a task in fewer tokens and steps, so the bill for a finished job can be lower than the price per token suggests. That is worth checking on your own jobs: a short caption leaves a model little room to save tokens.

How Opus 5.5 scores on the test closest to marketing work

On AutomationBench, a test built by Zapier of business workflows across many connected apps, Anthropic reports Claude Opus 5.5 at 40.0%. That is second to GPT-6 Astra's 41.4% and ahead of every other model either vendor lists. OpenAI describes the test as end-to-end workflows using 47 tools across sales, marketing, operations, support, finance and HR.

Two rows match across the vendors: Claude Opus 5 at 26.9% and Claude Fable 5.1 at 31.4%. Anthropic says its Opus 5 figure comes from Zapier's public leaderboard, and OpenAI says it took competitor figures from publicly available reports. The rows are still not one clean race. Anthropic says the Zapier runs it cites used no fallback model, so a safeguard intervention counted as a failure, which lowers Opus 5.5's score. OpenAI labels its Fable 5.1 figure as run with Claude Opus 5 as a fallback, which it says stepped in on about 40% of tasks. OpenAI names version 1.0.6 of the test and reports GPT-6 Sol at $0.27 per task in its Sol and Luna announcement. Anthropic does not state the effort behind Astra's 41.4%.

Read the ceiling as well as the ranking. Anthropic's chart labels the figure a pass rate, so even the top model passes fewer than half of these workflows when it runs them alone. On GDPval-AA v2.1, a test of professional work across 44 occupations run by Artificial Analysis, Anthropic reports Opus 5.5 at 1846 Elo against 1735 for Fable 5.1 and 1542 for GPT-6 Astra. Every score here is as a vendor publishes it, on a setup it chose, so read the scores as a direction rather than a promise for your account.

Which marketing jobs Opus 5.5 suits

Claude Opus 5.5 suits long, multi-step marketing jobs where a wrong figure or a muddled summary is expensive. It is poor value for short, high-volume jobs that a cheaper model already does well. The table is a starting point, not a rule.

The writing changes are the part most likely to show up in marketing drafts. Because Anthropic says Opus 5.5 follows the writing rules you give it, a written style guide is worth putting at the top of every prompt: banned words, sentence length, how to name the product. On the API, the prompt cache makes that fixed opening cheap to resend.

Keep a person on anything that leaves the building. A pass rate near 40% on business workflows is a reason for small steps and a sign-off, which is why every workflow in overads starts out asking a person before anything goes out.

Opus 5.5 text carries a watermark

Text written by Claude Opus 5.5 carries a watermark: Anthropic says the model ships with the watermarking it adopted to comply with the EU AI Act. How it works is set out in Anthropic's watermark explainer.

• Nothing is added to the text and there are no hidden characters. The watermark is a pattern in which word the model picks when several are equally good, and Anthropic says readers cannot tell the difference.
• Anthropic applies it worldwide, not only in the EU, and says it carries no information about the person or organisation using Claude.
• Light editing probably will not remove it, and a complete rewrite will. Short passages, factual sentences and light proofreading can carry too little of it to detect.
• Checking text needs Anthropic's detection API, which is in private preview for eligible organisations such as regulators, media, fact-checkers and researchers.
• Image files Claude produces get a C2PA content credential in their metadata instead, Anthropic says.

The watermark does not change who owns the output, Anthropic says. It means whether Claude drafted a post may be checkable later by someone with detection access. Whether a post must be labelled as AI-made is a separate question set by each platform and jurisdiction, and this is not legal advice.

What changes for teams building on the API

Teams that call Claude Opus 5.5 from their own tools meet four breaking changes from Opus 5, according to Anthropic's what's new page.

• Thinking cannot be switched off. A request that disables it, or sets a manual thinking budget, returns a 400 error. The effort setting is the control instead.
• Forced tool use is not supported. Setting tool_choice to any or to a named tool returns an error, so say in the prompt when a tool applies.
• Thinking blocks are tied to the model and the conversation, so a conversation that switches to most other models continues without the earlier reasoning.
• On the Claude API and Google Cloud, the older computer_20251124 computer use tool is refused.

Two behaviour changes need no code change but can surprise you. The default effort is medium, down from high on Opus 5, so set it explicitly and re-test. And the short notes the model writes between tool calls now arrive as thinking blocks, so an app that streams them to users goes quiet until it sets a display value that returns them.

Using Claude with your overads workspace

overads does not run on Claude: the text AI inside overads runs on Gemini. Claude can still work in an overads workspace as a client of the overads MCP server, which offers posts, workflows and ad metrics as tools an assistant calls with a workspace API key. The MCP server and API keys are included on paid overads plans.

In Claude Code, adding the server is one command in a terminal. Claude Desktop reaches it through the mcp-remote package, with the key in a header. The MCP setup page has both snippets, plus ones for Cursor, VS Code, Codex, Windsurf, n8n and Zapier.

What Claude can do depends on the key's scopes. A key with connections:read, posts:read and analytics:read only reads. Add posts:write and media:write to let it draft and schedule, and workflows:run to let it run automations. A tool outside the key's scopes is left out of the list the model sees.

Posts go out through Publish to X, LinkedIn, Bluesky, Mastodon and Pinterest. When a post is held for approval, the create call still succeeds, the post waits as pending, and the server tells the model to say so rather than retry. Holding every scheduled post in a workspace for approval is a workspace setting with no switch in the app, so it is off unless overads has turned it on for your workspace. The scheduling and approvals guide covers the rules, and every tool is listed on the MCP tools page.

What the overads tool list adds to each Claude request

Connecting the overads MCP server adds a fixed block of tool definitions to the requests Claude sends, and on Opus 5.5 that block costs a few cents uncached and a fraction of a cent from the cache. Anthropic bills the tools a request carries, meaning each one's name, description and schema, as input tokens, according to its API pricing page.

We measured the overads tool list on September 25, 2026, from the tool definitions the server sends, written out as the Claude API's tools parameter. With every scope, a key sees the 23 tools that are always on, which come to 15,598 bytes of JSON. A read-only key sees 14 of them, 9,212 bytes. Tools that read data derived from a platform stay off unless a deployment enables them, and are not counted here. The costs below count one token per byte, a deliberately high estimate: tokenizers pack several characters into most tokens, and Anthropic's token-counting endpoint gives the exact figure for your client.

Uncached at $4 per million input tokens, from the cache at $0.20. The first request that writes the list to the cache pays $5 per million instead. Anthropic adds its own tool-use instructions, 286 tokens on Opus 5.5, whenever a request carries tools, and some clients add a few bytes, such as a prefix on each tool name.

A narrower key is cheaper as well as safer. The read-only list is about 40% smaller, and a read-only key cannot create or edit a post or run a workflow. Give an assistant the write scopes only on the key it uses to schedule.

Where overads fits

overads does not run on Claude or sell access to it. The text AI inside overads runs on Gemini, and using Opus 5.5 takes your own Claude plan or API account. What overads adds is the workspace Claude works in: your connected channels, a publishing queue with approvals, saved workflows and read-only ad reporting, reachable over MCP.

Three limits apply whichever model you use. overads holds no write access to any ad account, so neither it nor Claude can change a campaign through it; Ads manager reads Meta, Google and Snapchat and proposes changes that a person applies. Publish is live on X, LinkedIn, Bluesky, Mastodon and Pinterest only. And overads does not check what Claude says about your brand: Signals is built for Gemini and Perplexity, and its AI-answer scan is switched off on the live product.

If you only need Claude to write copy, the Claude apps do that on their own, with no overads account. If you want an agent that edits live campaigns, it needs the ad platform's own tools and write permissions, which overads does not have.

What to try this week

• Give Opus 5.5 one long job you already check by hand, such as a monthly client report, and compare its figures against the source line by line.
• Write your house style rules down once and put them at the top of every prompt; Anthropic says Opus 5.5 follows them.
• Create a read-only overads API key, add the MCP server to Claude Code or Claude Desktop, and ask for last month's ad spend summary before you grant any write scope.
• If you build on the API, set the effort level explicitly rather than relying on the new medium default, and remove any request setting that turns thinking off.