Claude Opus 5.5: New features, pricing, benchmarks, and comparison with Opus 5

Lisa Ernst · 26.09.2026 · Artificial Intelligence · 11 min.

Claude Opus 5.5 has been available since September 22, 2026, and replaces Opus 5 as Anthropic's strongest Opus model after only about two months. The most striking change is not just the higher benchmark performance: regular API prices are dropping from $5 to $4 per million input tokens and from $25 to $20 per million output tokens. Anthropic estimates the cost for typical token-based workloads to be around 40 percent less compared to Opus 5, as Opus 5.5 is also supposed to handle tokens more efficiently.

This overview shows what's actually new in Opus 5.5, how price, context window, and benchmarks change compared to Opus 5, which migration pitfalls developers need to be aware of, and for whom a switch makes sense. Benchmark figures are classified as manufacturer values and not treated as universally valid rankings.

Source: ctan.org – uantwerpendocs

Quick Overview

• Release: Claude Opus 5.5 was released on September 22, 2026, as the first model in the new Claude 5.5 family.
• API Price: $4 per million input tokens and $20 per million output tokens; Opus 5 costs $5 and $25 respectively.
• Cache Reads: $0.20 per million tokens instead of $0.50 with Opus 5 – a reduction of 60 percent.
• Context: The context window remains up to 1 million tokens, with a regular maximum output of 128,000 tokens.
• Benchmarks: Anthropic reports particularly significant gains compared to Opus 5 in Terminal-Bench, CursorBench, AutomationBench, and Terminal-Bench-Science.
• Migration: The model name changes to claude-opus-5-5; there are also changes in Thinking, Tool Choice, and Computer Use that may affect existing integrations.

What is Claude Opus 5.5?

Opus 5.5 is Anthropic's new flagship model for demanding coding, agent, and knowledge work. Anthropic positions it for long, multi-step tasks in large codebases, complex tool usage, and professional workflows. For those who want to understand Claude fundamentally, the Zerlo overview What is Claude AI? provides the most important basics about Anthropic, models, and usage possibilities.

Technically, much remains familiar: Opus 5.5 processes text and images as input and generates text as output. According to the current Claude platform documentation, the context window is 1 million tokens, and the regular maximum output is 128,000 tokens. The reliable knowledge cutoff is stated as June 2026. For events after this date, the model will therefore continue to require up-to-date external data sources or web/tool access.

What's new in Opus 5.5?

1. Lower token costs and higher efficiency

The direct list price drops by 20 percent compared to Opus 5: input drops from $5 to $4 per million tokens, output from $25 to $20. The difference is larger for prompt caching. Cache reads cost $0.20 per million tokens with Opus 5.5 instead of $0.50 with Opus 5.

Nevertheless, Anthropic speaks of approximately 40 percent lower costs for typical token-based workloads. The reason: not only is the price per token lower, but the model is also expected to require fewer tokens for many tasks. This 40 percent is therefore not a blanket discount guarantee. The actual difference strongly depends on how much input, output, and cache usage a specific workflow generates.

2. Adaptive Thinking is always active

With Opus 5.5, Thinking can no longer be completely disabled. Instead, the Effort parameter controls how much computational effort the model should expend for a request. The default value for Opus 5.5 is medium; for Opus 5, the default according to the model documentation was high. This is important for cost and latency comparisons, as two models at different effort levels cannot be directly compared.

3. Stronger focus on coding and agents

Anthropic particularly emphasizes long-running software tasks. Published tests examined large code migrations, refactorings, terminal tasks, and multi-tool workflows. Such examples partially originate from Anthropic or early test customers and should therefore be understood as manufacturer or tester results – not as a guarantee that every real-world project will achieve the same time savings.

Source: ctan.org

Opus 5.5 is particularly aimed at developers who automate long coding and agent workflows. Crucial here are not only benchmark points but also token consumption, tool usage, and reliability across many work steps.

4. Enhanced safety and behavior checks

Anthropic reports that Opus 5.5 performs better than previous models in its own automated behavior testing and is more resistant to prompt injection than Opus 5. At the same time, the company emphasizes that limitations still exist. The model also launches with a similar class of additional protection mechanisms as Fable 5.1, including for cybersecurity, biology, and protection against model distillation. For productive systems, this does not replace your own access control, input validation, or control of critical actions.

Claude Opus 5.5 vs. Opus 5: Technical Specifications

Claude Opus 5.5: Pricing in Detail

For developers, the price change is one of the most important innovations. The following table shows the regular Claude API list prices. Cloud providers may have their own billing details; taxes, regional surcharges, or special inference options are not included here.

A simple calculation example shows the basic effect: 10 million input tokens plus 2 million output tokens cost a total of $80 with Opus 5.5 without caching. With Opus 5, it would be $100. This corresponds exactly to the 20 percent reduction in list prices. For cache-intensive agents, the difference can be larger.

Additionally, for Opus 5.5, there's a Fast Mode with up to 2.5x speed. It costs $8 per million input tokens and $40 per million output tokens. Therefore, those optimizing for cost rather than minimal latency should stick with the normal mode. The Batch API can reduce input and output costs by 50 percent for suitable asynchronous tasks.

Claude Subscription and API are Separate

Opus 5.5 is available in Claude products for Pro, Max, Team, and Enterprise users. However, usage of the Claude API is billed separately based on API consumption and is not automatically included with a Claude Pro subscription. This separation is important for companies, as chat usage and productive API workloads can develop independently in terms of cost.

Benchmarks: How much faster or better is Opus 5.5?

Anthropic publishes several of its own comparison values for Opus 5.5. In the following selection, only Opus 5.5 is compared with its direct predecessor, Opus 5. Percentage values are absolute hit or success rates; GDPval-AA is a point value.

Important: These are benchmark results published by Anthropic. Test configurations vary by benchmark; unless otherwise stated, Anthropic uses high or maximum effort settings for Opus 5.5. Safety mechanisms and fallbacks can also influence the results. A benchmark is therefore an indication of capabilities, but not a substitute for testing with your own prompt, dataset, tool setup, and cost profile.

For a broader comparison with another major model provider, the Zerlo article Gemini vs. Claude is also helpful. However, such comparisons should always consider which specific model generation and test date is being referred to.

Migration from Opus 5 to Opus 5.5

Users of Opus 5 only in the Claude app generally need to pay little attention. For API integrations, however, there are several changes that should be tested before a production switch.

1. Adjust Model ID: The API uses claude-opus-5-5.
2. Thinking no longer switchable: With Opus 5.5, Adaptive Thinking is always active. Applications should control computational effort via 'effort' instead of completely disabling Thinking.
3. Check Tool Choice: Forced variants such as 'any' or a fixed tool selection can lead to a 400 error with Opus 5.5. Anthropic recommends, among other things, 'auto' in conjunction with strict tool usage or structured output.
4. Update Computer-Use Tool: For Opus 5.5, the newer Computer Toolset is intended for the Claude API and Google Cloud. For Amazon Bedrock, different specifications may apply for compatibility reasons.
5. Do not rewrite Thinking blocks: Applications that process Conversation History themselves should treat Thinking blocks based on their type and pass them on unchanged.

Source: ctan.org

For existing API workflows, it is worthwhile to conduct targeted regression testing before switching models, especially for Tool Choice, Computer Use, structured outputs, long conversations, and custom cost control.

Where is Claude Opus 5.5 available?

According to Anthropic, Opus 5.5 is available to end-users in Claude for Pro, Max, Team, and Enterprise. Developers can use the model via the Claude Platform or Claude API. Additionally, Anthropic names Amazon Web Services, Google Cloud, and Microsoft Foundry as supported platforms.

Anthropic has also increased the five-hour usage limits for Pro, Max, Team, and seat-based Enterprise plans with the introduction of Opus 5.5. However, how much usage is actually possible still depends on the respective plan, the task, and token consumption.

Who benefits from switching from Opus 5?

For new API projects, there is little reason to start with Opus 5 if there are no special compatibility requirements. Opus 5.5 is cheaper per token, more current, and consistently achieves better values than Opus 5 in the benchmarks considered here in Anthropic's published tests.

For existing coding and agent systems, the switch is particularly interesting if cache reads, long contexts, and many tool calls constitute a large part of the costs. Here, the combination of lower list prices, significantly cheaper cache reads, and potential token efficiency can have a greater impact than the mere 20% price reduction.

For production integrations with hardcoded tool control, the migration should not be done solely by replacing the model name. Changes in Thinking and Tool Choice can cause behavioral changes or API errors. A short test catalog with typical and critical workflows is more sensible than an untested rollout.

Limitations and Important Context

• Vendor Benchmarks: The most important comparative figures come from Anthropic. Independent tests may yield different results depending on prompting, tooling, and evaluation.
• Costs are workload-dependent: The often-mentioned saving of around 40 percent is an estimate for typical token-based tasks, not a guaranteed saving for every API call.
• Knowledge Cutoff: The reliable knowledge cutoff is June 2026. For current events after this date, external information retrieval is necessary.
• Opus remains premium: Despite the price reduction, Opus 5.5 is designed for demanding tasks. For simpler or very high-volume workloads, cheaper Claude models may be more economical.
• Safety mechanisms alter behavior: Safety systems can block requests or direct them to other processing paths. This should be considered in benchmarks and for automated agents.

FAQ

Conclusion

Claude Opus 5.5 is not a minor incremental update. Compared to Opus 5, Anthropic combines higher published benchmark scores with lower API prices, significantly cheaper cache reads, and a stronger focus on long coding and agent tasks. The 1-million-token context window is retained; the main new features are efficiency, model behavior, and API details related to Thinking and Tools.

For new demanding Claude projects, Opus 5.5 is therefore the obvious Opus version. For existing Opus 5 systems, the switch is also worthwhile, but should be tested as a true migration due to the changes in Thinking, Tool Choice, and Computer Use. The benchmark advantages are promising, but the result in your own workflow – including costs, latency, and error rate – remains crucial.