What is Claude Opus 4.7, and what is new in it?

Yesterday, Thursday, April 16, 2026, Anthropic released Claude Opus 4.7, the newest model in its top "Opus" line and the successor to Opus 4.6 from February. It is available to everyone who can use Opus today: in the Claude apps, through Anthropic's API, and on the Amazon, Google and Microsoft cloud platforms. The price is unchanged.

Anthropic says the biggest gains are on the hardest software tasks, the kind people previously had to supervise closely. CNBC's report highlights another angle: Anthropic itself calls Opus 4.7 less broadly capable than Claude Mythos Preview, the model it announced last week and is keeping for a small group of security partners.

What Anthropic released

A token is a small piece of text, often part of a word; you pay per token sent and received.

Several updates arrived with it. On the API, task budgets (in public beta) let developers guide how many tokens Claude spends across a long job. In Claude Code, Anthropic's coding agent, a new /ultrareview command runs a dedicated review of code changes, and the default effort level is now xhigh.

How it improves

Harder coding, longer tasks. Early-access companies quoted by Anthropic report gains over Opus 4.6. One describes a 13% higher success rate on its own 93-task coding test, including four tasks that both Opus 4.6 and Sonnet 4.6 had failed. Others say it works through tool failures instead of stopping, and checks its own results before reporting back.

Sharper eyes. Opus 4.7 sees images at much higher resolution. That helps when Claude reads dense screenshots, technical diagrams or chemical structures. One security testing company reports its visual-accuracy score rising from 54.5% with Opus 4.6 to 98.5%.

Literal instruction following. Earlier models sometimes read instructions loosely or skipped parts. Opus 4.7 follows them as written. That is usually good, but old prompts that relied on loose reading may now give surprising results.

Effort levels. "Effort" sets how much the model thinks before answering. More effort means better answers on hard problems but more time and tokens. The new xhigh level sits between high and max; Anthropic suggests starting at high or xhigh for coding and agent work.

Memory across sessions. It makes better use of notes saved in files, so long projects need less re-explaining.

Two changes affect cost. A new tokenizer, the component that splits text into tokens, can turn the same input into roughly 1.0 to 1.35 times as many tokens. And the model thinks more at high effort. Anthropic says its own tests show a net improvement but advises measuring on real work.

The safety angle

Last week, with Project Glasswing, Anthropic promised to keep Mythos Preview restricted and to trial its new cyber protections on weaker models before anything else. Opus 4.7 is the first of those. During training, Anthropic experimented with ways to reduce its cyber skills selectively, and it now ships with filters that detect and block requests suggesting prohibited or high-risk hacking. Security professionals with legitimate needs, such as penetration testing (authorized attempts to break into a system to find weaknesses), can apply to a new Cyber Verification Program.

Anthropic reports a safety profile similar to Opus 4.6: low rates of deception and of going along with misuse, better honesty and better resistance to prompt injection (hidden instructions planted in content to hijack a model), but slightly weaker behavior in one area, giving overly detailed harm-reduction advice about controlled substances.

A worked example: a prompt written for a literal reader

Because Opus 4.7 takes instructions at face value, spell out exactly what you want and what you don't:

Review the three attached chapters of my novel for continuity errors only:
names, dates, places and timeline. Do not comment on style or suggest
rewrites. List each problem as: chapter, page, the conflict, and which
version you think is correct. If you find none in a chapter, say so.

An older model might have added style advice anyway. Opus 4.7 should stick to continuity and say plainly when a chapter is clean.

Why it matters

Opus 4.7 shows Anthropic running two tracks: a public model that keeps improving, and a stronger model held back for defenders. The safeguards tested here are meant to inform when, and whether, models of the Mythos class can be offered more widely. CNBC notes that the Glasswing launch has already prompted meetings between the Trump administration, technology CEOs and bank leaders about AI security risks.

For everyday users and developers, it is a straightforward upgrade at the same price, with better coding, vision and long-task reliability.

How should you adjust your prompts for Claude Opus 4.7?

Be explicit. Since the model follows instructions literally, state the scope, the format and what to leave out. If earlier prompts said "you may also" or listed optional extras, check whether you still want them.

Choose the effort level deliberately. For quick questions, lower effort saves time and tokens; for complex code or analysis, try high or xhigh. Developers can set task budgets so long jobs do not overspend.

Finally, check costs on real work. The new tokenizer and deeper thinking can raise token counts even though the price per token is the same. Sending smaller images when fine detail is not needed also helps, since high-resolution pictures use more tokens.