What Changed in Claude Opus 4.7 — and Why It Affects Your Prompts
Claude Opus 4.7 is Anthropic's flagship model, released on April 16, 2026. Three changes from Opus 4.6 directly affect how you should write prompts: a new tokenizer that processes text with up to 35% more tokens than before, substantially improved high-resolution vision (3.75 megapixels vs. 1.15MP for 4.6), and a new feature called task budgets for agentic workflows. Pricing remains unchanged at $5 per million input tokens and $25 per million output tokens.
This matters because the same prompt you ran on 4.6 will cost noticeably more on 4.7 if you don't adapt. The model is more capable, but looser prompts now carry a real token overhead. Understanding what changed mechanically is the fastest way to get better outputs at the same — or lower — cost.
What Is a Task Budget, and When Should You Use It?
A task budget is a soft token ceiling you assign to an entire Claude Opus 4.7 agentic loop — covering internal thinking, tool calls, tool results, and final output. You pass it via the API using task_budget: {type: "tokens", total: N} along with the tasks-budgets-2026-03-13 beta header. The minimum is 20,000 tokens; for complex multi-step workflows, Anthropic recommends starting at 64,000 tokens and tuning from there.
Task budgets are not a hard cut-off. Opus 4.7 sees a running token countdown and uses it to self-regulate — it prioritises and compresses gracefully as the budget drains, rather than truncating abruptly at the limit. This makes agentic workflows significantly more predictable and cost-controllable. Without a task budget, a multi-step workflow that could complete in 40,000 tokens can sprawl to 100,000 when the model over-reasons intermediate steps.
Task budgets are currently an API-level feature. Claude.ai users won't see a UI toggle yet — but if you're building automations in Make.com, n8n, or any platform that calls the Claude API directly, this is worth setting up immediately.
How the Improved Vision Changes What You Can Send the Model
Opus 4.7 accepts images up to 2,576 pixels on the long edge, which works out to approximately 3.75 megapixels — a significant jump from Opus 4.6's 1.15MP ceiling. This changes which visuals are actually useful to send the model, because content that was blurry and unreadable at 1.15MP becomes legible and extractable at 3.75MP.
Three workflows that are now practical with Opus 4.7's vision that weren't reliable on 4.6:
--- Full-resolution document screenshots: A scanned invoice, a contract page, or a printed spreadsheet photographed with a phone camera now renders with enough detail for Claude to extract structured data, flag inconsistencies, and read table contents accurately.
--- UI screenshot walkthroughs: Paste a screenshot of a web dashboard or software UI and ask Claude to write step-by-step instructions for navigating it. At 1.15MP, small UI text was frequently misread; at 3.75MP, it is reliably legible.
--- Slide deck review: Anthropic specifically highlighted improved .pptx editing in the Opus 4.7 release notes — the model can now visually verify its own slide layout outputs and catch formatting errors it would have missed at lower resolution.
Why Your Prompts Cost More Now — and the Fix
The new tokenizer in Opus 4.7 processes text with up to 1.35x as many tokens as the previous tokenizer, with the actual ratio varying by content type. Code, structured markup, and verbose conversational instructions tend to see the highest token inflation. This doesn't affect output quality, but it does affect cost — and for practitioners running high-volume workflows, the 35% overhead compounds across thousands of API calls.
The fix is not to reduce context — it's to eliminate filler. The single highest-impact prompt change for Opus 4.7 is removing the conversational warm-up that many practitioners write out of habit. Phrases like "You are a helpful AI assistant who will help me with..." or "Please be thorough and detailed when you..." cost tokens without improving outputs on a 2026 frontier model.
Replace free-text instructions with structured XML blocks. Here is a practical before-and-after:
Before (wasteful on Opus 4.7):
"You are a senior marketing strategist. I need you to help me write a campaign brief for a new product launch. The product is targeted at..."
After (structured, token-efficient):
<task>Write a campaign brief</task>
<product>[One-sentence product description]</product>
<audience>[Target segment, 2 lines max]</audience>
<output>500 words. Structure: Objective / Audience / Key message / Channels / Success metrics. Plain text, no preamble.</output>
The XML version uses fewer tokens, is easier for the model to parse, and produces more structurally consistent outputs — because each block signals the role of its content explicitly.
Prompting Techniques That Work with Opus 4.7 — and What to Stop Doing
Opus 4.7 is a frontier reasoning model that thinks internally before responding. Techniques designed to coax reasoning out of earlier models — writing "let's think step by step," adding phrases like "this is extremely important," using emotional pressure to improve output quality — have no effect and consume tokens that the model would otherwise use for actual reasoning.
What works well with Opus 4.7 in 2026:
--- XML-tagged context blocks — Use <instructions>, <context>, <examples>, and <output_format> to separate your prompt components. Claude's instruction-following on Opus 4.7 is significantly better when content roles are labelled rather than implied.
--- Few-shot examples inside clear tags — Two to three complete input-output examples communicate format, tone, depth, and edge-case handling simultaneously. This is more efficient than writing rules that describe what you want.
--- Explicit output format constraints — Specify word count, structure, what to exclude, and what to avoid. Opus 4.7's instruction-following precision means format constraints stated directly in <output_format> are respected with high fidelity.
--- Prompt chaining for multi-step work — Break complex deliverables into sequential prompts where each output feeds the next. This gives you quality checkpoints, keeps token budgets per step manageable, and is easier to debug than one large prompt that tries to do everything.
A Copy-Paste Template That Works for Most Tasks
This template applies the XML structured approach and adapts cleanly across writing, analysis, research, and data extraction tasks. The structure stays constant — you fill in the contents for each use case.
Try This Prompt:
<instructions>
You are analysing [document type or data] to produce [specific output type].
Be direct and specific. No preamble. No meta-commentary.
</instructions>
<context>
[Paste your document, data, or relevant background here]
</context>
<task>
[Your specific request — one clear, unambiguous instruction]
</task>
<output_format>
[Specify format, length, and structure. Example: "A bulleted list of 5–7 key findings, each 1–2 sentences. Plain English. No headers."]
</output_format>
The <output_format> block is the one practitioners most frequently omit — and it's the single biggest driver of inconsistency. Without a format constraint, Opus 4.7 will choose a format that may be longer and more conversational than your use case requires.
Three Common Mistakes When Switching from Opus 4.6 to 4.7
Most practitioners who upgrade to Opus 4.7 without adjusting their workflow encounter three predictable problems. Knowing them in advance takes less than five minutes to prevent and can save significant cost at volume.
--- Carrying over verbose system prompts from 4.6 without trimming. System prompts written for earlier models often include redundant instructions because earlier models needed more explicit handholding. Audit your system prompts and remove instructions that restate the obvious or repeat themselves. On Opus 4.7, those extra tokens cost money without contributing to output quality.
--- Not using task budgets on agentic workflows. Without a task budget, Opus 4.7 running in an agentic loop has no self-regulation signal. This leads to over-verbose reasoning chains and unpredictable token usage across steps. Set a task budget before deploying any multi-step Claude workflow at volume.
--- Assuming the new vision is automatic. Opus 4.7's 3.75MP vision applies only when you explicitly pass images in the correct API format — as an image block, not as base64 embedded in text and not as a description of an image. If your workflow pastes image URLs into a text field, the vision capability is not being triggered.
懂AI，更懂你 UD相伴，AI不冷。這些不是小技巧——它們是讓 Opus 4.7 真正為你工作的基礎配置。
Benchmark Your AI Proficiency Level
Understanding Claude Opus 4.7 is one signal of AI maturity. The UD AI IQ Test gives you a full-picture benchmark across models, prompting techniques, and workflow design — and we'll walk you through every step of building your AI capability from there.