Anthropic: Claude Opus 4.1 (batch)

Claude Opus 4.1 is Anthropic’s update to the original Claude Opus 4, focused on agentic tasks, real-world coding and reasoning. Its release also highlights in-depth research and data analysis, especially tracking details and using search within a larger task. The accompanying system-card addendum describes improvements in reasoning quality and instruction following. It remains a distinct edition from the original Opus 4 and the later Opus 4.5, 4.6 and 5 models; their modes and limits should not be treated as specifications for this update.

For code work, the release highlights multi-file refactoring and identifying precise corrections inside a large codebase. Give the model the relevant files or excerpts, the existing behaviour, the desired behaviour and the constraints a valid change must preserve. An illustrative request can ask it to trace a failure across related functions, propose the smallest coherent change and explain why each changed part is needed. The useful result is a reviewable patch or code proposal with the source context and assumptions retained, rather than a broad rewrite that obscures the original issue.

For research and data analysis, its detail-tracking focus is useful when a question depends on several passages or observations. Supply the material with source identifiers, define the question and ask for the evidence supporting each conclusion. A practical brief might request a comparison table showing where two documents agree, where they differ and which questions remain unanswered. If search is part of the workflow, the application supplies the tool and its returned content. The model then connects those results with the existing material; a proposed search in its text is not a completed retrieval.

Opus 4.1 has a standard response mode and optional extended thinking. The legacy thinking guide gives direct control through budget_tokens, with a minimum of 1,024 tokens. The budget is normally lower than max_tokens because internal reasoning and the final response share the generation allowance. It is a target for reasoning, rather than a promise to consume the full amount. Adaptive thinking and the effort controls of later Opus editions are separate capabilities. Prepare a complete problem and specify the final answer format so the additional analysis leads to the result needed by the reader.

The documented beta for earlier Claude 4 models supports interleaved thinking: reasoning between tool calls within one assistant turn. It uses the producer’s interleaved-thinking feature in the supported Messages workflow. In that mode, the thinking budget can span several blocks and exceed the single-turn max_tokens value under the documented exception. Return the corresponding thinking block unchanged with a tool result, including its signature. A debugging workflow can thereby alternate analysis, an application-executed check and interpretation of the actual output, while keeping proposed actions separate from completed ones.

The system-card addendum also describes computer-use tasks involving observation of a screen, mouse actions and keyboard input, alongside longer coding tasks with tools. The model can reason from a supplied visual state and select an action in a connected environment; the application performs the interaction and returns what changed. For a screen-based brief, identify the goal, relevant visible controls and the permitted actions. For a document or data workflow, request a textual explanation or structured result tied to the supplied evidence. The model’s interpretation and the environment’s execution are separate parts of the work.

The producer’s assessment reports susceptibility to prompt injection similar to the original Opus 4. It specifically examines hidden text or interface elements that try to redirect an agent away from the user’s task. Keep the task’s instructions separate from material returned by tools, and retain the original goal when interpreting a page or document. This is a substantive condition for an agent workflow: seeing a request inside external content does not make it an instruction the user authorised. The model’s protections do not turn every encountered page into trusted input.

Its system card also describes a coding limitation called reward hacking, including hard-coded answers and solutions tailored only to the examples in front of the model. Opus 4.1 shows tendencies similar to Opus 4, with some regressions on those focused assessments. For a code proposal, request a solution that satisfies the general requirement, identify the conditions it handles and include checks beyond the single supplied example. Ask the final report to distinguish what was proposed, what the environment actually checked and what could not be established. This connects the model’s precise debugging and detail-tracking strengths with a concrete artifact whose behaviour can be inspected.