Claude Opus 4.7

More Than an Incremental 4.x Update

Anthropic released Claude Opus 4.7 on April 16, 2026. It remains available as an Active (legacy) model, with retirement not scheduled sooner than April 16, 2027.

The model combines a 1,000,000-token context window, up to 128,000 standard output tokens, text and image input, adaptive thinking, tool use, prompt caching, and batch processing.

Its historical importance comes from API and behavior changes that continued into later Opus models. Claude Opus 4.7 removed manual budget_tokens extended thinking, introduced effort-controlled adaptive thinking, changed tokenization, added better user-facing progress during long agent traces, and became the first Claude model with high-resolution image support.

At the same time, it still behaves differently from Claude 5. Thinking is opt-in rather than always on, forced tool choice is accepted, and several newer orchestration features are absent.

• Model ID: claude-opus-4-7.
• Released April 16, 2026.
• Active legacy model.
• 1M context window and 128K standard output.
• Text and image input with text output.
• Adaptive thinking available but not always on.

Better Fidelity for Screenshots, Documents, and Computer Use

Higher image resolution matters when the information is small, dense, or spatially precise.

UI screenshots may contain tiny labels, status icons, form values, code, or controls. Document images can contain fine print and compact tables. Diagrams may require the model to identify exact regions rather than provide a broad visual summary.

Opus 4.7 increased the maximum image resolution automatically—no beta header or client-side opt-in is required.

Anthropic also changed coordinate behavior. Pointing and bounding-box coordinates returned by Opus 4.7 map 1:1 to actual image pixels, removing the need for a scale-factor conversion.

The tradeoff is token usage. Full-resolution images can consume roughly three times as many image tokens as earlier lower-resolution processing, with Anthropic documenting up to approximately 4,784 tokens per image.

For image-heavy applications, evaluate both accuracy gains and token cost. If fine visual detail does not affect task success, downsampling may remain more economical.

Prompt Caching Matters for Long Agent Sessions

Large agentic workloads often reuse substantial stable context: system instructions, repositories, policies, reference documentation, schemas, or prior material.

Anthropic lists 5-minute cache writes at $6.25 per million tokens, 1-hour cache writes at $10 per million, and cache reads at $0.50 per million.

Batch API input and output receive a 50% discount, making offline evaluations and bulk processing materially cheaper than standard synchronous traffic.

Current Claude Opus 5.5 is cheaper at $4 input and $20 output per million tokens, so migration can reduce the base cost even before accounting for model-quality differences.

When evaluating Opus 4.7, compare cost per accepted result rather than nominal price. Adaptive thinking, high-resolution images, tool calls, long contexts, and retries can all change total task cost.

• Input: $5.00 per 1M tokens.
• Output: $25.00 per 1M tokens.
• 5-minute cache write: $6.25 per 1M tokens.
• 1-hour cache write: $10.00 per 1M tokens.
• Cache read: $0.50 per 1M tokens.
• Batch API: 50% discount on input and output.

Character Counts No Longer Map the Same Way to Cost

Anthropic notes that the Opus 4.7 tokenizer may produce roughly 1× to 1.35× as many tokens as models before Opus 4.7, depending on the content.

That is operationally important. A product may estimate prompt size from characters, apply a fixed safety margin to max_tokens, trigger compaction at a hard threshold, or forecast monthly spend using historical token ratios.

Those assumptions should be revalidated when moving from an earlier Claude model to Opus 4.7.

Use the provider token-counting endpoint for real payloads rather than assuming old character-to-token heuristics remain accurate. Also re-check compaction thresholds and any client-side logic that predicts whether a request fits within the context window.

A Distinctive Transitional Feature Set

The model accepts text and images and generates text. Adaptive thinking is supported but optional. Tool use supports agent workflows, including forced tool-choice patterns that change in newer Opus generations.

High-resolution image support is automatic and handles up to 2,576 pixels on the long edge. The model also returns 1:1 image coordinates for pointing and bounding-box tasks.

Anthropic’s migration documentation notes that Opus 4.7 already returns refusal stop_details, allowing an application to inspect additional refusal metadata when the model declines a request.

The model does not support newer Opus 5.5 capabilities such as mid-conversation system messages or fast mode.