Claude Opus 4.7 adds high-resolution image support, task budgets, and removes extended thinking

Claude Opus 4.7 is Anthropic's latest generally available model, designed for complex reasoning and agentic coding with support for a 1M token context window and 128k max output tokens.

New features

The model adds high-resolution image support with a maximum resolution of 2576px/3.75MP, up from the previous 1568px/1.15MP limit. This should improve performance on vision-heavy workloads, particularly for computer use and document understanding workflows. Coordinates now map 1:1 with actual pixels, eliminating scale-factor calculations. High-res images consume more tokens, so downsampling is recommended when additional fidelity isn't needed.

Claude Opus 4.7 also introduces a new xhigh effort level for coding and agentic use cases, with high recommended as the minimum for intelligence-sensitive tasks.

Task budgets (beta)

The model adds task budgets, which give Claude an estimate of how many tokens to target for a full agentic loop including thinking, tool calls, and outputs. The model sees a running countdown and uses it to prioritize work. Minimum task budget is 20k tokens, and it's distinct from max_tokens which remains a hard per-request cap.

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"]
)

Task budgets are advisory, not hard caps. If set too low, the model may complete tasks less thoroughly or refuse them entirely. For open-ended agentic tasks where quality matters more than speed, don't set a task budget.

Breaking changes

Extended thinking budgets are removed in Claude Opus 4.7. Setting thinking: {"type": "enabled", "budget_tokens": N} will return a 400 error. Adaptive thinking is now the only thinking-on mode, which in internal evaluations reliably outperforms extended thinking.