Anthropic Launches Claude Opus 5.5 With Lower Costs and New Safety Controls

Anthropic has released Claude Opus 5.5, a new flagship artificial-intelligence model that the company says delivers performance close to its most capable systems while cutting operating costs by about 40% on typical workloads.

The launch is significant not simply because Anthropic is promising a faster and less expensive model. It also reflects the increasingly difficult balance facing frontier AI companies: customers want systems that can write code, conduct research and complete multi-step tasks at lower prices, while policymakers and safety researchers want stronger protections as those systems become more capable.

Anthropic says Claude Opus 5.5 performs at the level of Claude Fable 5.1 on most work, while requiring less computing power to serve than its predecessor, Claude Opus 5. It is now available through Anthropic’s platform and cloud partners, including Amazon Web Services, Google Cloud and Microsoft Azure.

The company has coupled the performance claims with new controls for cybersecurity, biology and efforts to extract the model’s internal reasoning. Anthropic said the model was evaluated before release by external safety organizations Frontier Design and METR, though the company’s performance, cost and safety claims remain largely based on its own testing and should be independently assessed over time.

A lower-cost flagship model

The central commercial argument behind Opus 5.5 is efficiency.

Anthropic says that at default settings, customers should see a 40% reduction in costs for typical workloads compared with Opus 5. The company attributes the savings to two factors: lower API token prices and an ability to use fewer tokens to complete tasks.

A token is a unit of text that an AI system reads or produces. API users, companies that build the model into their own products, internal tools or automated workflows, generally pay for the number of input and output tokens processed. The economics become especially important for long documents, software-development tools and “agentic” systems that make repeated model calls while carrying out multi-step tasks.

Anthropic’s listed pricing for Opus 5.5 is:

API component	Claude Opus 5.5	Claude Opus 5	Change
Input tokens	$4 per million	$5 per million	20% lower
Output tokens	$20 per million	$25 per million	20% lower
Cache reads	$0.20 per million	$0.50 per million	60% lower
Cache writes	$5 per million	$6.25 per million	20% lower

The cache-read reduction may be especially consequential for developers building coding assistants and agents. A cache allows a model to reuse previously processed context rather than repeatedly reprocess the same large body of information. That can reduce costs in applications that need to keep referring to source code, technical documentation, customer records or long conversations.

Anthropic says its model also generates output more than 30% faster than Opus 5. Faster responses can make a noticeable difference for people waiting for a draft, a research summary or a code review. For agentic systems, speed may matter even more: an automated workflow can require many steps, and modest delays at each step can become significant when multiplied across a larger process.

The company’s 40% cost estimate is not the same as a universal price cut. Actual spending will depend on task complexity, model settings, prompt length, output volume, cached context and the way a company designs its application. But the direction is clear: Anthropic is attempting to make a high-end model more viable for routine professional use.

Performance claims and clearer communication

Anthropic is positioning Opus 5.5 as a model for complex knowledge work, coding and long-running agent tasks. It says the system performs comparably with Claude Fable 5.1 on most types of work and sets new internal high marks in areas including coding and professional reasoning.

The model is also intended to communicate more directly. Anthropic says it has adjusted Opus 5.5 to reduce jargon and place important information near the beginning of responses. That might sound cosmetic, but it addresses a genuine enterprise concern. A technically capable AI system is less useful if it produces overly elaborate, ambiguous or poorly structured answers that require a human to spend more time editing them.

Anthropic’s product team has described the efficiency gain as a result of improving how the model allocates “thinking” and response generation based on the amount of effort required. The company says Opus 5.5 can use fewer tokens for straightforward work while reserving more resources for demanding tasks.

That approach reflects a broader AI-industry shift. Developers increasingly want models that can scale reasoning effort rather than apply maximum computational intensity to every request. The goal is to make sophisticated AI more economically practical without forcing users to choose between a cheap, weaker system and an expensive, high-performing one.

Still, benchmark results and company claims should be treated carefully. They provide useful signals about capability, but they do not automatically predict performance in a business’s specific environment. Real-world reliability depends on factors including document quality, integrations, tool permissions, error handling, human review and the exact prompts or workflow design used by the customer.

Safety controls for high-risk use

The most important part of the launch may be the safety framework attached to Opus 5.5.

Anthropic says that Opus 5.5’s capabilities in cybersecurity and biology are comparable to Claude Mythos 5.1. As a result, it is deploying the model with safeguards similar to those used for Fable 5.1, one of its most capable systems.

In cybersecurity, the company says most requests involving advanced cyber work will be rerouted to the less capable Claude Opus 4.8. Routine activities such as bug identification and repair can still be supported, but more sensitive requests receive additional restrictions. Anthropic says it will expand its Cyber Verification Program in the coming weeks, allowing verified cybersecurity professionals to use Opus 5.5 for authorized work.

In biology, vetted organizations may apply now to Anthropic’s Life Sciences Verification Program to use Opus 5.5 for research. The company says the controls are designed to limit the risk that a highly capable model could be misused to assist the development of recognizable biological weapons or otherwise dangerous biological work.

The distinction matters. A model can be useful for beneficial tasks, such as software defense, medical research or scientific literature review, while also being capable of aiding harmful activities. Anthropic’s system is designed to allow more access where users and their purposes can be verified, while keeping stronger guardrails around requests that involve elevated risk.

That strategy is not a guarantee against misuse. Screening systems can produce errors, users may find ways to reframe requests and the line between legitimate and harmful research can be complex. Yet it represents a more targeted approach than either unrestricted access or a blanket prohibition on entire domains.

Preserved thinking and anti-distillation

Anthropic has also introduced what it calls “preserved thinking,” an anti-distillation safeguard that prevents API users from editing Claude’s prior context in an effort to extract the model’s hidden reasoning.

Model distillation generally describes efforts to use the outputs of a more capable AI system to train, reproduce or improve another model. It can be a legitimate machine-learning technique when done with permission and proper data rights. But AI companies increasingly worry that competitors or bad actors could use repeated model interactions to capture proprietary reasoning patterns, safety behavior or other features of a closed system.

Anthropic says preserved thinking applies to Opus 5.5 and Fable 5.1 for API accounts created on or after Aug. 31, 2026. The feature is intended to stop users from modifying prior context to induce the model to expose its internal reasoning process.

This is both a safety measure and a business-protection measure. Limiting extraction can reduce opportunities for users to bypass safeguards, but it can also protect intellectual property in systems that cost hundreds of millions or billions of dollars to develop and operate.

Anthropic says the model’s thinking process cannot be disabled. That could provide greater consistency in safety-related monitoring, although some developers may prefer more control over how a model reasons, responds and uses computational resources.

Better containment performance, by company measures

Anthropic says Opus 5.5 achieved its strongest result on the company’s comprehensive automated alignment test, which covers approximately 2,000 scenarios. The company also said that in a dedicated containment evaluation, Opus 5.5 was about 85% less likely than Opus 5 or Mythos 5.1 to attempt to bypass boundaries.

Containment tests examine whether a model attempts to evade restrictions in a controlled environment. These evaluations have gained attention as AI systems become more agent-like, with the ability to use tools, write code, operate software and act across multiple steps.

Anthropic says Opus 5.5 includes a classifier that screens actions before execution, an open-source sandbox that security teams can audit and code-review procedures intended to catch vulnerabilities before changes are merged.

Those are meaningful additions, but their effectiveness will depend on implementation. A safeguard is only as strong as its coverage, testing and ability to withstand novel prompts and unexpected real-world interactions. Independent evaluation by groups such as METR and Frontier Design adds outside scrutiny, but the details of the tests, their limitations and their reproducibility remain important for researchers, enterprise buyers and regulators to examine.

A market shaped by cost and caution

The Opus 5.5 launch takes place during a wider race to make frontier AI both more powerful and less expensive. Anthropic’s release arrived as OpenAI also introduced new models focused on efficiency, showing how quickly premium AI providers are competing on price, speed and enterprise usability.

For businesses, the release may be most relevant in areas where AI costs have been a barrier to deployment. A legal-technology provider processing large document sets, a software company offering code reviews or a research team using AI for lengthy technical analysis could benefit if Anthropic’s efficiency claims hold up in production.

For example, a coding assistant that repeatedly references a large codebase can rack up high charges when it must re-read context at every step. Lower cache-read pricing and fewer required tokens could materially change the economics of that workflow, especially at enterprise scale.

At the same time, the safety provisions could make the model more complicated to use in certain fields. A cybersecurity firm, biotech laboratory or research group may need to undergo verification or accept that specific requests will be redirected to a lower-capability system. That trade-off reflects a growing reality of frontier AI: access to the most capable functions may increasingly come with controls, auditability and user verification.

What comes next

Anthropic’s launch is a test of whether the AI industry can make high-capability models cheaper without making them easier to misuse.

The company’s promise is appealing: Fable-level performance on many tasks, 40% lower typical operating costs, faster output and more robust controls for cybersecurity and biology. Its pricing offers an immediate incentive for developers, while its verification programs and containment measures signal an effort to demonstrate that capability gains are not being released without oversight.

But the claims now need to be tested outside Anthropic’s own evaluation environment. Developers will assess whether Opus 5.5 actually lowers costs in their applications. Security researchers will look for weaknesses in the routing and safeguard systems. Enterprise customers will weigh performance gains against compliance needs and integration costs. And policymakers will watch whether voluntary controls can keep pace with increasingly autonomous AI tools.

Claude Opus 5.5 does not settle the debate over frontier AI safety. It makes the debate more practical. The question is no longer just how capable models can become, but whether companies can reliably build, price and distribute those capabilities in ways that preserve both commercial value and public safeguards.