Claude Opus 5.5: Stronger Coding and Knowledge Work at Lower Cost

For equal volumes of uncached input and output, the savings are 20% compared to Opus 5 and 60% compared to Fable 5.1. As an example, suppose your requests total one million uncached input tokens and 100,000 output tokens. Those tokens would cost $6 on Opus 5.5, $7.50 on Opus 5, and $15 on Fable 5.1.

Caching is an important factor to take into account. When requests reuse cached context, the cache-read rate applies to those tokens, instead of the input price. Opus 5.5 reduces that rate by 60% relative to Opus 5, so applications that repeatedly reuse substantial context have another source of savings. Compared to Fable's prices, cache reads are 20% cheaper. Your overall cost reduction therefore depends on the mix of uncached input, cache activity, and output.

Anthropic expects roughly 40% lower cost per completed task than Opus 5, reflecting lower prices and improved token efficiency. That’s a broader expectation than the 20% base-price reduction, which can be attributed to the increase in capabilities even at lower reasoning settings. The example above shows the savings at equal token volumes, but the cost of completing actual work also depends on how many tokens the model uses and which requests benefit from caching. For that reason, we recommend running your own evaluations against every new model.

Anthropic also expects roughly 30% faster output-token generation than Opus 5, reducing the time required to complete tasks.

Where Opus 5.5 Fits in the Claude Family

For demanding coding and professional work, we recommend teams start with Opus 5.5. The combination of higher reported results and lower pricing makes the choice particularly clear for existing Opus 5 users. The rest of the family still has a place across different uses:

• Opus 5: Opus 5.5 is the natural upgrade for its core workloads. The main integration change is the requirement to keep thinking enabled, which we’ll cover next. Teams should upgrade as soon as feasible unless their own evaluations show lower scores.
• Fable 5.1: Opus 5.5 offers a strong lower-cost alternative on the evaluated coding and knowledge tasks. Anthropic continues to position Fable for the hardest, longest autonomous jobs, although the launch comparisons don’t demonstrate a Fable advantage on that specific category. Model selection here will depend heavily on what your own evaluations and usage report. We recommend you try Opus 5.5 for any workload currently using Fable 5.1, not as a capability increase but as a potential cost reduction.
• Sonnet 5: Its lower prices and speed remain useful for routine production work where its capability is sufficient. Opus 5.5 costs twice as much per base input or output token, so teams should still prefer Sonnet 5 where it's enough.
• Haiku 4.5: Its economical, low-latency role remains attractive for simple, high-volume tasks. Opus 5.5 costs four times as much per base input or output token. It's unlikely that Opus 5.5 will be a good replacement for tasks where Haiku 4.5 is sufficient, simply because it will cost more. However, if you're using Haiku 4.5 on a very high effort setting, you might want to try Opus 5.5 at low effort.

Context capacity is another difference worth considering. Opus 5.5 supports a one-million-token context window and up to 128,000 output tokens, matching Opus 5, Fable 5.1, and Sonnet 5. Haiku 4.5 supports 200,000 and 64,000 tokens, respectively. Opus 5.5 also retains standard input pricing across its full context window, without a premium above 200,000 tokens like Sonnet 5. Opus 5.5 well suited to workloads that require processing large amounts of context without incurring higher input costs at scale.

Adaptive Thinking and the API Changes That Matter

The main difference when switching from Opus 5 is that for Opus 5.5, adaptive thinking is always enabled, whereas for Opus 5, adaptive thinking is an available setting but is disabled by default.

Adaptive thinking means the model decides how much reasoning to use, and you steer it through the effort setting. The available levels are low, medium, high, xhigh, and max. medium is the default. Requests to Opus 5.5 that disable thinking or specify a manual thinking budget return HTTP 400, which is a breaking change compared to Opus 5.

What you need to change depends on your current model’s thinking configuration:

• From Opus 5 or Sonnet 5: Remove any setting that disables thinking. Both already reject manual thinking budgets, so that restriction isn’t new. Their high default effort becomes medium when you move to Opus 5.5. Requests that explicitly select high still use high.
• From Fable 5.1: Always-on adaptive thinking and rejection of manual budgets are the same in Fable 5.1 and Opus 5.5. The default effort changes from high to medium.
• From Haiku 4.5: Thinking was off by default, with manual budgets available when enabled. Moving to Opus 5.5 changes this to always-on adaptive thinking controlled through effort.

Medium is a sensible starting point for reasons beyond cost. In Anthropic’s FrontierCode effort comparison, Opus 5.5 scores 54.6% at medium, slightly above its 54.4% result at max. A higher effort setting gives the model more room to work, but it doesn’t guarantee a better result on every task.

Higher effort can be useful for particularly difficult problems. It can also lead to extra checking or adjacent work beyond the request. On the other hand, at low effort, the model can stop short of completing the task. Starting at medium gives you the recommended default, with room to increase effort when the work calls for it. Note: teams should not assume their previous effort setting is still best for this new model. Test Opus 5.5 across a range of effort settings, since it may deliver similar quality to previous models at lower effort settings, reducing costs for your workloads.

Long-running agents also get two new context compaction options, currently in beta. Compaction summarizes older conversation turns so the agent can keep working within its context window. The new options preserve the most recent turns word for word while summarizing older history, and allow the summary to be built in the background while the agent continues working.

This is useful if your application currently shortens conversation history itself. Custom edits to earlier turns can break preserved thinking, while the reasoning information is carried forward between turns. API-managed compaction moves those summarization patterns into the service, reducing the need for your own implementation of history editing. Compaction already existed, but these additions address two common reasons we've observed for building custom handling.

Broader Safeguards and How Fallback Works

Opus 5.5 brings Fable-level classifiers to the Opus family, screening work in cybersecurity, biology, AI research and development, and distillation (reasoning extraction). Anthropic also reports its strongest Opus prompt-injection resistance yet, tied with Fable 5.1 on Gray Swan’s benchmark.

Opus 5 already had classifiers and fallback behavior. The change is the broader coverage and the routes available when a classifier blocks work on Opus 5.5. Fallback lets an eligible request continue on another model:

• Cyber blocks fall back to Opus 4.8.
• Biology and AI R&D blocks fall back to Opus 5.
• Distillation, conventional-weapons, and explosives blocks do not fall back; they're blocked.

Fallback is automatic in Claude applications and Claude Code. On the Claude API, it’s opt-in through the beta Fallback API. This means a task can continue after a block, with a different model handling the affected work. Anthropic's benchmark results include fallback behavior.

Note that Anthropic’s server-side fallback documentation separates the Claude API from partner-operated endpoints such as Amazon Bedrock, where fallback requires client-side handling.

Another important point is that classifier decisions and permitted usage are separate questions. A request may be routed to a model with different safeguards while remaining subject to Anthropic’s Usage Policy. Fallback doesn’t grant permission for otherwise prohibited activity.

Additionally, unlike Fable 5.1, Opus 5.5 adds no model-specific data-retention requirement. Fable falls under Anthropic’s Covered Models requirements, generally including at least 30 days of retention. For Opus 5.5, your existing agreement and platform settings determine the retention requirements.

A Strong Upgrade for Demanding Work

Opus 5.5 improves the case for using an Opus model on substantial engineering and professional tasks. It raises reported performance over Opus 5 and compares favorably with Fable 5.1 on the selected evaluations, at lower token prices for both comparisons.

We recommend starting at medium effort for coding, research, analysis, and tool-using knowledge work. If you’re migrating an existing application to Opus 5.5, compare completion quality and cost on representative tasks and your own evaluations, and account for its always-on adaptive thinking configuration. That gives you a focused way to confirm the benefit while keeping the model choice tied to your specific work.

How Caylent Can Help

As a Preferred Services Partner and Charter Member of the Claude Partner Network, Caylent helps you select and implement Claude models for your applications. Our Anthropic Consulting & Engineering practice supports adoption, enablement, and building agentic applications, including the workload evaluation and cost optimization needed to make a model upgrade useful in production.

Whether you’re moving an Opus 5 application, considering Opus 5.5 for work currently handled by Fable, or building a new agent, connect with us to determine the right model and integration approach for your workload.