Anthropic just made a less exciting part of AI much more interesting: the bill.
Claude Opus 5.5, announced on September 22, 2026, is the first model in Anthropic’s Claude 5.5 family. Anthropic says it performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5 on typical workloads at default settings. Input tokens cost $4 per million and output tokens cost $20, compared with $5 and $25 for Opus 5. Cache reads cost $0.20 per million tokens, down from $0.50.
That matters more to marketing teams than another leaderboard victory. If the cost of completing a long, multi-step AI workflow falls, tasks that were previously too expensive to automate become worth testing.
What Anthropic announced
Anthropic says Claude Opus 5.5 is a major step up from Opus 5 for agentic coding, computer use and knowledge work, while generating output more than 30% faster.
The model supports a 1 million token context window and up to 128,000 output tokens, which makes it better suited to long-running workflows that need to keep large amounts of research, documentation, campaign data or code in context.
The pricing change is the more practical part. Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, 20% less than Opus 5 on both measures. Anthropic says typical workloads cost 40% less at default settings because Opus 5.5 also uses fewer tokens per task. Cache reads cost 60% less, falling from $0.50 to $0.20 per million tokens.
| Model | API price per 1M tokens | Practical marketing fit | Main thing to watch |
| Claude Opus 5.5 | $4 input / $20 output | Recurring analysis, agent workflows, campaign QA, research-heavy work | Anthropic’s 40% lower typical workload cost should still be tested against your own workflows |
| Claude Opus 5 | $5 input / $25 output | Existing complex analysis and agent workflows | Higher cost and slower output than Opus 5.5 |
| Claude Fable 5.1 | $10 input / $50 output | Hardest reasoning and long-running knowledge work | Much higher token price, so it is not automatically justified for routine marketing work |
Pricing note: Anthropic’s 40% figure refers to typical workload cost at default settings, not a 40% reduction in every cost category. Cache reads are $0.20 per million tokens for Opus 5.5, versus $0.50 for Opus 5 and $0.25 for Fable 5.1. Five-minute cache writes are $5 per million tokens for Opus 5.5, down from $6.25 for Opus 5.
Existing workflows need a migration check
Claude Opus 5.5 is available through the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS, alongside Anthropic’s consumer and business Claude plans.
Developers should not treat the upgrade as a drop-in replacement without checking their implementation. Anthropic documents four breaking changes from Opus 5.
- Adaptive thinking is always on. It cannot be disabled, and medium effort is the default.
- Forced tool use is no longer supported. Workflows that rely on forcing a specific tool call need to be reviewed.
- Thinking blocks are model and conversation bound. Reusing them across models or conversations can break existing assumptions.
- Computer-use tooling changed. Older computer-use tool versions are no longer accepted on the Claude API and Google Cloud.
The adaptive thinking change is especially relevant for cost control. Because thinking is always enabled, the effort setting becomes one of the main levers for balancing quality, latency and spend.
Why lower cost changes what marketers can automate
Marketing teams rarely need an AI model to write one more headline. They need it to handle the boring middle of the workflow: gathering data, checking sources, comparing campaigns, cleaning up spreadsheets, analysing results and producing something a human can actually use.
Take a weekly paid media review. An agent could pull approved campaign data, compare it with CRM outcomes, flag unusual changes, inspect landing-page performance and prepare a report for review. The value is not that AI wrote the report. The value is that fewer manual steps sit between raw data and a useful decision.
That connects directly to the problem behind B2B marketing attribution. A dashboard can tell you that CPL fell 20%. It is much harder to connect that change to SQLs, pipeline and revenue. A lower-cost model makes more analysis possible, but it does not make bad attribution good.
Lower model cost also changes which recurring workflows are economically sensible. Long-running processes that require many model calls, large context windows and repeated cache reads become easier to justify when the cost per completed task falls.
That creates room for more ambitious AI marketing automation, including recurring campaign analysis, competitive monitoring, account research, content QA and reporting preparation rather than simply generating ad copy.
The same economics apply to content. The expensive part of content is often not producing a first draft. It is researching the topic, checking claims, reviewing existing pages, finding internal links, formatting the content and running QA.
OneMetrik’s earlier coverage of Claude Opus 5 made a similar point from a different angle. The useful metric is not simply model intelligence. It is whether the model can complete a real workflow without creating another pile of work for the human reviewing it.
Cost per completed workflow matters more than benchmark scores
Anthropic reports strong benchmark results for Opus 5.5 across agentic coding, computer use and knowledge work. Those results are useful evidence that the model is competitive, but they are not the most important buying signal for marketing teams.
The more useful question is what it costs to complete a real task. A model that is slightly stronger on a benchmark but needs more tokens, more retries or more human review can still be more expensive in production.
That is also the lesson from OneMetrik’s coverage of the OpenAI Agents API. Model price is only one part of the economics. Tool calls, context, retries, permissions, execution environments and human review all contribute to the cost of completing a real workflow.
For teams evaluating Claude models specifically, Fable 5.1 remains relevant for the hardest reasoning workloads, but Opus 5.5 is easier to justify for recurring marketing operations where cost and speed matter alongside capability.
What marketing teams should test first
There are four areas where lower model costs are worth testing first:
- Reporting: analyse campaign and CRM exports and prepare a weekly performance brief.
- Research: compare competitors, pricing pages, positioning and product announcements on a recurring schedule.
- Content QA: check existing content for outdated claims, missing evidence, weak internal links and obvious optimisation gaps.
- Campaign analysis: give the model read-only access to approved performance data and ask it to identify anomalies before a marketer decides what to change.
Keep high-impact actions behind human approval. A model analysing a $50,000 monthly ad account is one thing. A model changing the budget on that account without review is another.
This matters for AI performance marketing because lower AI costs do not remove the need for controls. In fact, cheaper automation can make bad automation easier to scale.
The same principle applies to AI search visibility. AI can help monitor brand mentions, compare cited sources and identify content gaps, but teams still need to decide which findings deserve a change to the site or content strategy.
OneMetrik Takeaway
Claude Opus 5.5 is interesting because it makes the economics of AI work more practical.
The next useful step for marketing AI is not generating more individual assets. It is connecting research, analysis, QA and reporting into repeatable workflows that a human can trust.
The lower price matters because it gives teams more room to test those workflows. The metric to watch is not whether AI saved a few minutes on one task. It is whether a recurring workflow can now be completed reliably enough, quickly enough and cheaply enough to justify automation.