John StephensonAnthropic just released Claude Opus 4.7, and the headline capability isn't a benchmark number. It's behavior: the model now routinely checks its own work before reporting back to you.
Self-verification — identifying a logical error during planning, writing a test for its own code, or feeding generated audio through a speech recognizer to confirm accuracy — is the difference between an AI tool that requires supervision and one you can actually hand off work to. According to Anthropic, users are now doing exactly that with their hardest coding problems.
Opus 4.7 is available today across all Claude products, Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry. Pricing is unchanged: $5 per million input tokens, $25 per million output tokens. Access via claude-opus-4-7 in the Claude API.
0%
CursorBench score vs 58% for Opus 4.6
3x
Higher image resolution up to 2,576px long edge
3x
More production tasks resolved vs Opus 4.6 (Rakuten)
What Actually Changed
Four areas of meaningful improvement over Opus 4.6. Click each to expand.
Instruction Following
Opus 4.7 takes instructions literally rather than loosely. This is a genuine upgrade but prompts tuned for Opus 4.6 may behave differently. Re-tune any that relied on the model interpreting loosely or skipping steps.
Vision at 3x Resolution
Images up to 2,576px on the long edge — more than 3x the prior limit. Computer-use agents, diagram extraction, and pixel-level tasks all benefit. Higher-res images do consume more tokens.
File-Based Memory
Significantly better at filesystem-based memory across multi-session work. It remembers context from earlier sessions and starts new tasks with less up-front setup — key for long-running autonomous workflows.
Real-World Work Quality
Tested as a finance analyst, Opus 4.7 produced more rigorous analyses and tighter task integration. It scores top-tier on GDPval-AA covering finance, legal, and other high-value knowledge work.
Benchmark Comparison — Opus 4.7 vs Opus 4.6
Source: Anthropic, April 2026. CursorBench and BigLaw Bench (Harvey) shown.
Capability Profile
How Opus 4.7 scores across six capability dimensions. The biggest gains are in coding precision, vision, and instruction-following accuracy.
Opus 4.7 vs Opus 4.6 — Capability Radar (BizHacker.io Analysis)
Scores are relative assessments based on Anthropic benchmark data, April 2026. Out of 10.
What's New Beyond the Model
New xhigh effort level
Sits between "high" and "max" — finer control over reasoning vs latency. Claude Code raises its default to xhigh for all plans. For coding and agentic tasks, start at high or xhigh.
Task budgets in public beta
Developers can set token budgets to guide spend across longer runs. Essential for managing costs on complex autonomous workflows. Available now in the Claude API.
/ultrareview in Claude Code
A new slash command producing a dedicated review session. It reads through changes and flags bugs and design issues. Pro and Max users get three free ultrareviews.
Auto mode extended to Max users
Auto mode lets Claude make decisions on your behalf — fewer interruptions, lower risk on long tasks. Now extended to all Max plan subscribers.
Cyber safeguards deployed
Opus 4.7 ships with automatic detection and blocking of prohibited or high-risk cybersecurity requests — the first deployment of the safeguard system Anthropic will use for broader Mythos-class model availability.
The Cybersecurity Context
Opus 4.7 lands in an unusual moment. Last week Anthropic announced Project Glasswing — a defensive cybersecurity initiative built around Claude Mythos Preview, a more powerful model with advanced cyber capabilities. Anthropic committed to keeping Mythos Preview limited and testing safeguards on less capable models first.
"Opus 4.7 is the first such model: its cyber capabilities are not as advanced as those of Mythos Preview. We are releasing Opus 4.7 with safeguards that automatically detect and block requests that indicate prohibited or high-risk cybersecurity uses."— Anthropic, April 16, 2026
Security professionals who need Opus 4.7 for legitimate work — vulnerability research, penetration testing, red-teaming — can apply to Anthropic's new Cyber Verification Program.
What Early Partners Are Saying
Anthropic published feedback from 28 early-access partners. A few that stand out for business leaders:
Early Access Partner Feedback
Pricing and Migration
Pricing is unchanged from Opus 4.6. Two token-related changes are worth planning for:
FactorDetailImpact
Updated tokenizer
Same input = 1.0–1.35x more tokens
Higher cost for identical prompts
More thinking
More output at higher effort levels
More tokens, higher reliability
Effort control
New xhigh between high and max
Finer cost vs quality tradeoff
Task budgets
API beta — guide token spend
Prioritize work across longer runs
Anthropic's own testing found the net effect favorable. A migration guide is at platform.claude.com/docs.
What This Means For Your Business
4 Action Items for Business Leaders
1Re-tune your prompts. Workflows built on Opus 4.6 should be tested on 4.7 before full rollout. Improved instruction-following means loose or ambiguous prompts now behave more literally.
2Measure token costs on real traffic. The new tokenizer and increased reasoning can increase token usage 1.0–1.35x. Use task budgets and the effort parameter to manage costs. Start with
xhigh for coding tasks.3Evaluate vision-dependent workflows. The 3x resolution upgrade is significant for document extraction, diagram analysis, and screenshot processing. Computer-use agents will perform noticeably better.
4Plan for autonomous multi-session work. Improved file-based memory and self-verification make Opus 4.7 more suitable for long-horizon autonomous tasks. This is a meaningful threshold for handing off complex engineering work.
Frequently Asked Questions
Common questions about Claude Opus 4.7
What is Claude Opus 4.7?
Claude Opus 4.7 is Anthropic's latest flagship AI model, released April 16, 2026. It delivers major improvements in advanced software engineering, vision, instruction following, and long-running autonomous tasks. It is available across all Claude products and via the API using the model string
claude-opus-4-7.
How much does Claude Opus 4.7 cost?
Pricing is unchanged from Opus 4.6: $5 per million input tokens and $25 per million output tokens. However, the updated tokenizer means the same input text may consume 1.0–1.35x more tokens than before, so real costs may increase slightly depending on content type. Use task budgets and the effort parameter to manage spend.
What is the xhigh effort level?
xhigh is a new effort level that sits between 'high' and 'max', giving finer control over the tradeoff between reasoning depth and latency. Anthropic has raised the default effort level in Claude Code to xhigh for all plans. For coding and agentic tasks, starting at high or xhigh is recommended.
Is Opus 4.7 a direct upgrade from Opus 4.6?
Yes — Opus 4.7 is a direct drop-in replacement. The main migration consideration is that prompts written for 4.6 that relied on loose interpretation may now behave differently. Anthropic recommends re-tuning prompts and measuring token costs on real traffic before full rollout. A migration guide is at platform.claude.com/docs.
What is the Cyber Verification Program?
Anthropic has launched a Cyber Verification Program for security professionals who need Opus 4.7 for legitimate offensive security work — including vulnerability research, penetration testing, and red-teaming. Opus 4.7 ships with safeguards that automatically detect and block prohibited or high-risk cybersecurity requests by default. Verified professionals can apply for elevated access.
What is /ultrareview in Claude Code?
/ultrareview is a new slash command in Claude Code that produces a dedicated review session. It reads through your code changes and flags bugs and design issues that a careful reviewer would catch. Pro and Max Claude Code users receive three free ultrareviews to try it out.
What vision improvements does Opus 4.7 include?
Opus 4.7 accepts images up to 2,576 pixels on the long edge — roughly 3.75 megapixels, more than three times prior Claude models. This is a model-level change so images are processed at higher fidelity automatically. Higher-resolution images consume more tokens, so users who don't need the extra detail can downsample before sending.
Where is Opus 4.7 available?
Claude Opus 4.7 is available across all Claude products (claude.ai, Claude Code, Claude Cowork), via the Claude API, Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry. The API model string is
claude-opus-4-7.
Will my existing Opus 4.6 prompts work with Opus 4.7?
They will work, but they may behave differently. Opus 4.7's improved instruction-following means it interprets prompts more literally. Prompts that previously relied on the model filling in gaps, skipping sections, or interpreting loosely may now produce more literal or unexpected results. Anthropic recommends a structured testing phase before full production rollout.
What is the difference between task budgets and the effort parameter?
The effort parameter (low, medium, high, xhigh, max) controls how much the model reasons before responding — it's a per-request quality control. Task budgets are a separate mechanism that guides Claude's total token spend across a longer autonomous run, helping prevent runaway costs on complex multi-step workflows. Both are available via the Claude API.
John Stephenson
