Claude Opus 5.5 Cuts Costs and Adds Safeguards for Autonomous AI
Claude Opus 5.5 cuts costs and adds safeguards for autonomous AI. The model is available through Anthropic, Amazon Web Services, Google Cloud and Microsoft Azure, with lower prices for standard API use and cached input reads.
Early testers used Opus 5.5 for engineering tasks that ran for hours. Developers should review the model’s safeguards, reasoning requirements and security-access rules before deployment.
What does Claude Opus 5.5 do for long-running work?
Anthropic designed Opus 5.5 for codebase migrations, software audits, financial analysis, data collection and workflows that involve several applications. The target is sustained work that may run for hours with minimal supervision, rather than a single prompt and response.
In one early test, Opus 5.5 audited and fixed a 200,000-line codebase in less than three hours. Opus 5 took more than 20 hours to complete the same task and used 2.5 times as many tokens, according to Anthropic.
Early testers also reported improvements in retaining context, delegating work to other agents and checking results. Their accounts indicated that the model needed fewer prompts, tool calls and corrections to complete tasks.
Anthropic reports gains in coding, computer use and professional knowledge work. Its benchmarks and customer tests indicate that Opus 5.5 can complete many tasks with fewer tokens than Opus 5, though performance may vary according to the task, tools and effort settings.
What happened during an overnight engineering test?
Sean Heintz, staff software developer at Clio, described assigning a large engineering task across six repositories and letting the model run overnight without supervision. According to his account, it stayed on task for more than 18 hours while defining how the services communicate and how each should apply that pattern.
Heintz said Opus 5.5 reached milestones faster than Opus 5 and required minimal reworking. He also said its code comments were shorter and more useful than the longer, prose-heavy comments produced in the earlier test.
What did code-review testing show?
Carl Bennett, CIO at Deloitte Consulting, said the lowest effort setting caught 72% of known bugs in code reviews, compared with 56% for Opus 5 at high effort. Help Net Security reported that Bennett said Opus 5.5 produced fewer false alarms and a fraction of the output.
In US consulting analysis, Bennett said low thinking effort matched higher thinking settings on half the output and passed quality checks. Those results may indicate that more lower-effort deployments could produce client-ready work with less output, although Anthropic notes that performance varies by task, tools and effort settings.
How does thinking mode affect local business projects?
Opus 5.5 cannot run with thinking switched off. Every request goes through a reasoning process, and developers can adjust the effort applied to the task.
The model also includes preserved thinking, a safeguard intended to prevent API users from editing Claude’s prior context. Anthropic says this makes it harder to extract and copy the model’s capabilities through large-scale distillation attacks. Preserved thinking applies to Opus 5.5 API accounts created on or after August 31, 2026.
Anthropic says Opus 5.5 is easier to understand, places key information earlier in its responses and follows writing instructions more closely.
What safeguards apply to cybersecurity and research?
Opus 5.5 includes safeguards for cybersecurity, biology and attempts to copy the model. Most cybersecurity tasks are rerouted to Opus 4.8, a more restricted model path, under Anthropic’s safeguards.
Anthropic has expanded its Cyber Verification Program into three access tiers for vetted security professionals. The tiers cover Claude Opus 5.5, Claude Sonnet 5.5 and Claude Mythos 5.1, according to Anthropic. The program allows security teams to apply for the access level that suits their work and includes access to Anthropic’s most capable models, including new models released in the future.
Organizations whose biological research is impeded by the safeguards can apply for Anthropic’s Life Sciences Verification Program.
For coding work, Anthropic added a classifier that screens coding-agent actions before execution. It also introduced an open-source sandbox that security teams can audit and code-review features designed to catch vulnerabilities before changes are merged.
The model has stronger defenses against prompt-injection attacks. In an evaluation of attempts to cross containment boundaries, Opus 5.5 tried to circumvent its assigned limits about 85% less often than Opus 5 or Claude Mythos 5.1, according to Anthropic. Anthropic said every attempt was low severity and self-reported.
METR and Frontier Design evaluated Opus 5.5 before release. Anthropic acknowledges that current evaluations cannot identify every potential failure before deployment.
What does Claude Opus 5.5 cost?
Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. Those rates are lower than Opus 5’s $5 per million input tokens and $25 per million output tokens.
Cached input reads cost $0.20 per million tokens, down from $0.50. The lower cache price could benefit coding agents and other systems that consult the same instructions, files or conversation history repeatedly.
The model produces output more than 30% faster than Opus 5. A separate fast mode is available through Claude Code and the Claude Platform, offering up to 2.5 times the speed for $8 per million input tokens and $40 per million output tokens.
Anthropic is also increasing five-hour usage limits for Pro, Max, Team and seat-based Enterprise customers. Subscription users receive a rate-limit reset that they can save and use later.
Evaluate the full cost of each workflow, including input, output, cached reads, effort settings, tools and repeated context use, before choosing a configuration.
How should a local business assess Opus 5.5?
A local business should first identify work that genuinely runs long enough to benefit from the model, such as researching customer questions across several sources, auditing a business application or coordinating a multi-step digital workflow.
The business should also calculate the full run cost, including input, output, cached reads and any use of the fast mode. Consider the task’s complexity and compare permitted effort settings. In US consulting analysis, Bennett said low thinking effort matched higher thinking settings on half the output and passed quality checks.
Finally, the business should keep a person responsible for checking outputs and actions, particularly when a workflow touches customer data, cybersecurity, biology or other sensitive material.
What to do now
- List one repeatable, extended task. Choose work such as customer research, data collection or a software audit that can run across several steps without a person intervening throughout.
- Compare effort settings. Run the task at more than one permitted effort level and check the quality, output length and completion time against the result the business needs.
- Calculate the complete API cost. Include input tokens, output tokens, cached reads and fast-mode usage before deciding which configuration suits the workflow.
- Assign a reviewer. Decide who will inspect long outputs and sensitive actions before the workflow reaches customers or changes business systems.
- Check access rules. If the work involves cybersecurity or biological research, review Anthropic’s Cyber Verification Program or Life Sciences Verification Program before deployment.
FAQ
How much does Claude Opus 5.5 cost compared with Opus 5?
Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. Cached input reads cost $0.20 per million tokens, down from $0.50.
Can Claude Opus 5.5 be used without thinking mode?
No. Opus 5.5 always uses a reasoning process. Developers can adjust the effort applied to each task, but they cannot switch thinking off.
How fast is Claude Opus 5.5 compared with Opus 5?
Claude Opus 5.5 generates output more than 30% faster than Opus 5. A separate fast mode offers up to 2.5 times the speed at $8 per million input tokens and $40 per million output tokens.