What changes from Claude Opus 5
Opus 5.5 uses less compute to deliver stronger results. According to Anthropic, it costs about 40% less on typical workloads, generates output more than 30% faster, and communicates decisions more directly.More completed work with less rework
In tests and early-adopter reports, the model sustained engineering sessions for many hours, worked across several repositories, and completed extensive migrations with fewer calls and corrections. One test cited by Anthropic completed a 680,000-line code migration in less than one day.Clearer communication
Opus 5.5 puts the most important information first, reduces unnecessary explanations, and follows provided writing rules more closely. This matters in long sessions: decisions and changes are easier to review before moving to the next step.More caution before acting
Anthropic reports stronger prompt-injection resistance and a lower tendency to take hard-to-reverse actions or exceed defined boundaries. These improvements reduce risk, but do not replace human approval for critical changes.What the benchmarks show
In results published by Anthropic, Claude Opus 5.5 advances in agentic coding, computer use, and professional work. The results below mostly use adaptive thinking at Max effort; Terminal-Bench 4.0 used XHigh.The benchmarks and cost comparisons were published by Anthropic and depend on the harness, tools, safeguards, and effort used. The company notes that small margins between frontier models do not always represent the same difference in real-world use. Validate the model in your workflow before changing a production configuration.
Where it does not lead
The improvement is not uniform. On AutomationBench, GPT-6 Astra scored 41.4%, compared with 40.0% for Opus 5.5. Astra also remained ahead on Terminal-Bench-Science 0.1. Choose Opus 5.5 for its combination of quality, efficiency, communication, and safety—not with the expectation that it leads every evaluation.Where it creates the most value
A software change that must span several systems
The work requires mapping dependencies, editing multiple repositories, running tests, and coordinating stages over many hours. Opus 5.5 was optimized to maintain the plan, delegate subtasks, and reduce correction cycles. Expected result: an end-to-end implementation, executed validations, and documented decisions ready for review.An analysis that must become a spreadsheet, document, and presentation
Dispersed sources must be combined, numbers must be checked, and the recommendation must arrive in a decision-ready format. The model brings research, analysis, and clearer communication together in the same execution. Expected result: consistent executive material with traceable calculations, evidence, and explicit gaps.An agent that works for hours without losing its boundaries
In long-running operations, capability without control increases risk. Opus 5.5 improves self-verification and adherence to defined boundaries, making it more suitable for agents that execute many steps and tools. Expected result: continuous, auditable execution that stops for approval whenever an action exceeds the authorized scope.How to apply it in Tess
1
Choose work that justifies the model
Prefer complex coding, professional research, data analysis, or deliverables spanning multiple sources and tools. Use a lighter model for simple transformations.
2
Provide context and criteria
Include files, repositories, policies, examples, and constraints. Explain the expected result and how it will be validated.
3
Define tools and approvals
Enable only the necessary tools. Require confirmation before publishing, deleting, sending, or changing production environments.
4
Start at the default effort
Medium is the provider default and is designed to balance quality, speed, and cost. Increase effort only when the workflow requires deeper investigation or reasoning.
5
Request a review before completion
Instruct the agent to verify tests, sources, numbers, requirements, and changes before delivering the result.
Three playbooks to get started
1. Multi-repository migration
Input: repositories, architecture, expected behavior, and validation commands.Tools: code, terminal, tests, and technical documentation.
2. Evidence-backed executive analysis
Input: documents, spreadsheets, external sources, and a business question.Tools: search, data analysis, and document generation.
3. Performance audit and remediation
Input: application, metrics, behavioral requirements, and test environment.Tools: code, profiling, terminal, and test suite.
Limitations and best practices
- Adaptive thinking remains active; thinking tokens are billed as output.
- A 1M-token context does not replace information selection. Send only relevant content.
- At XHigh and Max, leave enough output room for thinking, tools, and subagents.
- Safeguards may reroute certain cybersecurity and biology tasks to other Claude models.
- Review code, numbers, citations, and files before publishing or running in production.
- Keep human approval for financial, legal, clinical, security, or other high-impact decisions.