CONFIRMED: Anthropic launches Claude Opus 5.5 with greater capability at lower cost

[{"h":"CONFIRMED: the capability jump now comes with a migration bill","p":"Anthropic launched Claude Opus 5.5 on September 22 as its new leading model. The central change is not only higher performance: the company positions it at 40% lower cost on typical Opus 5 workloads. For production users, efficiency can become an operational advantage, but the package also requires integration review before a broad switch."},{"h":"What changed in real work","p":"Anthropic reports gains in coding, reasoning, tool use and long-running tasks, including migrations and large codebase audits. Announced pricing is $4 per million input tokens and $20 per million output tokens, while cache reads cost $0.20 per million. The relevant metric remains cost per completed task, including latency, human review and failures."},{"h":"The change that can break existing agents","p":"The official technical documentation details changes that do not appear in benchmark rankings. Adaptive thinking is always on and controlled through the effort parameter. Thinking blocks carry model identity and must be returned unchanged in tool loops. Forced tool use with tool_choice any or tool no longer works, requiring auto combined with strict tool use. On the API and Google Cloud, the older computer use tool must also move to the new toolset. Applications that display text between tool calls should review thinking block handling, because that text may arrive empty at the default setting."},{"h":"The impact on Brazilian SMBs","p":"For small and midsize companies, the launch lowers the potential cost of agents for software maintenance, document analysis, internal automation and content production. But savings can disappear if a team merely changes the model ID and discovers later that the agent no longer interprets intermediate responses or tool calls. The safe path is to test a representative workflow, measure cost per outcome and validate streaming, conversation memory, tool use and computer use in staging."},{"h":"Capability requires proportional controls","p":"Anthropic also links Opus 5.5 to stronger safeguards for high-risk uses. That is positive, but it does not replace human approval, logs, permission limits or rollback plans. The editorial counterpoint is simple: cheaper models invite more automation, and automation without governance expands the radius of an error."},{"h":"Editorial thesis","p":"Opus 5.5 matters because it combines frontier capability, lower pricing and a changed technical contract. The advantage is not changing a model name in code, but adapting the surrounding system and proving each task became cheaper without losing reliability."},{"h":"Sources","p":"Anthropic, official announcement
Anthropic, whats new and migration documentation
TechCrunch, independent coverage"}]