Claude Opus 4.7: Features, API Changes, and Migration Notes

Claude Opus 4.7 was released by Anthropic on April 16, 2026, as an upgrade for agentic coding, visual work, and long-running professional tasks. It introduced high-resolution image input, an xhigh effort level, beta task budgets, and several API changes that require attention when migrating from Opus 4.6.

Opus 4.7 is no longer Anthropic's newest Opus model: Opus 4.8 and Claude Opus 5 were released later. This article remains useful for teams maintaining Opus 4.7 integrations or evaluating a controlled migration. New projects should compare the currently available models before choosing one.

Claude Opus 4.7 specifications

Item	Claude Opus 4.7
API model ID	claude-opus-4-7
Context window	1 million tokens
Maximum output	128,000 tokens
Base API price	$5 per million input tokens and $25 per million output tokens
Thinking mode	Adaptive thinking
Notable additions	High-resolution vision, xhigh effort, and task budgets in beta

Pricing excludes prompt-caching variations and platform-specific charges. Consult Anthropic's current API pricing before estimating production costs.

What Claude Opus 4.7 changed

High-resolution image input

Opus 4.7 increased the maximum supported long edge for high-resolution image processing from 1,568 to 2,576 pixels. This is useful when small interface controls, dense documents, diagrams, or fine visual details would be lost through aggressive downscaling. High-resolution support is automatic; it does not require a beta header.

The extra detail has a cost. Anthropic notes that high-resolution images can consume substantially more visual tokens than standard-resolution input. Resize images before sending them when the task does not require pixel-level detail. This reduces both latency and token charges. See the Claude vision documentation for supported formats and image-sizing behavior.

A new xhigh effort level

The effort setting lets developers trade reasoning depth for latency and token use. Opus 4.7 added xhigh between the existing high and maximum settings. Anthropic recommends it as a strong starting point for many coding and agentic workloads, but that is not a substitute for application-specific evaluation.

Use a lower effort level for simple classification, extraction, or routine transformations. Reserve higher levels for tasks where additional reasoning produces a measurable improvement. Compare quality, latency, input size, output size, and tool-call count on representative prompts before changing a production default.

Task budgets for agent loops

Task budgets are a beta feature that tells Claude roughly how many tokens remain for a complete agentic loop, including thinking, tool calls, tool results, and the final answer. The model sees a countdown and can pace its work so it finishes more gracefully as the budget is consumed.

A task budget is advisory; it is not the same as a hard response cap. max_tokens still limits output for a single API request, while task_budget helps the model allocate work across an agent loop. Use both controls when you need predictable request boundaries and better model awareness of the overall scope.

from anthropic import Anthropic

client = Anthropic()

response = client.beta.messages.create(
    model="claude-opus-4-7",
    max_tokens=128000,
    output_config={
        "effort": "high",
        "task_budget": {
            "type": "tokens",
            "total": 128000
        }
    },
    messages=[
        {
            "role": "user",
            "content": "Review the codebase and propose a refactor plan."
        }
    ],
    betas=["task-budgets-2026-03-13"]
)

The beta header and schema can change while the feature remains in beta. Verify the example against Anthropic's task budget documentation before deploying it.

Breaking API changes from Opus 4.6

These changes matter for applications that call the Messages API directly. Test them in a staging environment before switching the model ID.

Extended-thinking budgets are not supported

Opus 4.7 rejects the older extended-thinking configuration that used thinking.type: "enabled" with budget_tokens. Replace it with adaptive thinking and use effort to steer how much work the model performs.

# Opus 4.6 style: do not send this to Opus 4.7
thinking = {
    "type": "enabled",
    "budget_tokens": 32000
}

# Opus 4.7
thinking = {
    "type": "adaptive"
}
output_config = {
    "effort": "high"
}

For Opus 4.7, adaptive thinking is not automatically enabled when the request omits the thinking field. Send the adaptive configuration when the task requires it.

Non-default sampling parameters are rejected

Requests to Opus 4.7 return a 400 error when temperature, top_p, or top_k is set to a non-default value. The safest migration is to omit those parameters and use a clear prompt plus the effort setting. Even on older models, setting temperature to zero never guaranteed byte-for-byte identical output.

Thinking display behavior changed

Do not build a user interface that depends on private reasoning text. If your product needs visible progress, provide status updates based on tool execution, application state, or a supported summarized-thinking option. Treat the response schema documented for your exact model and SDK version as authoritative.

Token counts may differ

A model migration can change tokenization and therefore both cost and prompt fit. Re-run token counting on representative production inputs, leave sufficient room for the expected response, and review any context-compression thresholds in your agent framework. Do not assume a prompt that barely fit a previous model will have identical token usage.

Capability improvements and practical limits

Agentic coding and tool use

Anthropic positioned Opus 4.7 as an improvement for sustained coding and agent tasks, including following work through validation rather than stopping after a plausible patch. Real results still depend on the harness: the agent needs appropriate repository context, safe tool permissions, a working test environment, and explicit completion criteria.

Evaluate the model on your own issues rather than relying on a single public benchmark. Measure task success, regressions, human review time, tool errors, and total token cost. A stronger model can still perform worse when instructions are ambiguous or essential tools are unavailable.

Documents, slides, and visual verification

The higher image resolution supports screenshot inspection, document understanding, charts, and slide-layout checks. For office-file workflows, render the final pages or slides and let the model inspect those images in addition to the document structure. Human review remains necessary for important financial, legal, or customer-facing files.

File-based memory

Opus 4.7 can work with an agent's notes, plans, or other file-based state over multiple steps. This is not automatic long-term memory. The application must decide what to store, where to store it, and when to retrieve it. Keep sensitive data out of uncontrolled memory files, apply retention rules, and make stored conclusions traceable to their source.

Migration checklist

1. Change the model ID only in a test environment first.
2. Remove extended-thinking budgets and non-default sampling parameters.
3. Enable adaptive thinking where it adds value, then tune effort with representative tasks.
4. Recalculate token use, latency, and cost for typical and worst-case inputs.
5. Verify streaming, tool calls, retries, and error handling against the current SDK.
6. Test high-resolution images only where the extra visual detail is useful.
7. Add a task budget only after confirming the beta contract and required header.
8. Run regression evaluations and retain a rollback path.

Should a new project still choose Opus 4.7?

Use Opus 4.7 when an existing system has been evaluated against it, compatibility is important, or a provider specifically exposes that model. For a new integration, compare it with current Claude models on the same workload. Anthropic introduced Claude Opus 5 after Opus 4.7, so “latest model” claims about 4.7 are now outdated.

Teams choosing between chat access and an API should also read Claude Pro versus the Claude API. For broader development options, see the best AI tools for programming. Whichever model you use, pin the model ID, monitor release notes, and treat migrations as software changes that require testing rather than simple name replacements.