Frontier models can already work inside a repository for hours. The problem is that every attempt, tool call, and correction appears in both the bill and the waiting time. Claude Opus 5.5, introduced by Anthropic, targets this less glamorous part of agentic work: reaching a useful result with fewer steps, fewer tokens, and less rework.
Anthropic says the model costs 40% less than Opus 5 on typical workloads and generates output more than 30% faster. API pricing is $4 per million input tokens and $20 per million output tokens. Cache reads dropped to $0.20 per million, an important detail for coding agents that repeatedly carry instructions, tool definitions, and project context.
The right unit is the task, not the token
A price table tells only half the story. A model that is cheap per token can still be expensive if it rereads the repository, repeats commands, or repairs incomplete changes. Anthropic's most useful claim is that Opus 5.5 needs fewer steps and tokens to complete long-running work.
In the company's own evaluations, the model leads Opus 5 on agentic coding, computer use, and knowledge work. Early users also report migrations and audits across large codebases, including jobs that ran for many hours. These are useful signals, but they still come from a controlled launch and evaluations selected by the vendor. Your repository remains the benchmark that pays the bill.
Where to test it first
Do not start with an isolated function. Choose a familiar task that crosses files and requires verification: update an integration, fix a regression, migrate an API, or audit a workflow with existing tests. Compare Opus 5.5 with your current model using the same goal and similar limits.
Measure time to a reviewable change, tool calls, token use, and human rework. Record how often the agent needs to reverse course. In agentic work, a correct answer on the third attempt may be much less efficient than a model that costs more per call but gets the change right the first time.
Anthropic also reports stronger prompt-injection resistance and fewer hard-to-reverse actions. That deserves a separate test, especially when an agent reads external documentation, runs commands, or works without supervision for long periods. A safety benchmark does not replace sandboxing, least-privilege access, and review before merge.
Availability and model choice
Opus 5.5 is available in Claude, Claude Code, and the Claude Platform, and it is starting to reach integrations such as GitHub Copilot. Anthropic also increased limits for Pro, Max, Team, and seat-based Enterprise plans. Sonnet 5.5 and Haiku 5.5 are expected in the coming weeks.
This release moves the conversation in a useful direction. Maximum capability still matters, but efficiency per completed task is now part of the product. For teams, the practical question is not whether Opus 5.5 wins another benchmark. It is whether the model delivers a reliable change with less waiting, lower consumption, and less review on the work you actually do.