Claude Opus 4.7 Is Here: State-of-the-Art Coding, xhigh Effort, and a New Cyber Safeguards Tier

Anthropic shipped Claude Opus 4.7 on April 16, 2026, a focused upgrade over Opus 4.6. Pricing is unchanged, but the lift is sharp in advanced software engineering, vision, and instruction following. The launch also ships the first batch of cyber safeguards Anthropic plans to scale up before the eventual broad release of its Mythos Preview model.

Key Takeaways

• Opus 4.7 is generally available today (April 16, 2026) on Claude products, the API, Bedrock, Vertex AI, and Microsoft Foundry
• Pricing matches Opus 4.6: $5 per million input tokens, $25 per million output tokens
• Notable gains on the hardest software engineering tasks; users can hand off work that previously needed close supervision
• New xhigh effort level sits between high and max; Claude Code defaults to xhigh on all plans
• Vision now accepts images up to 2,576 pixels on the long edge -- more than 3x prior models
• /ultrareview slash command in Claude Code; auto mode extended to Max users
• First model to ship the cyber safeguards Anthropic plans to scale up toward Mythos-class models

What Is Claude Opus 4.7?

Claude Opus 4.7 is Anthropic's new flagship general-purpose model, addressable on the API as claude-opus-4-7. It is available on day one across Claude products, the API, Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry. Pricing is unchanged from Opus 4.6 -- $5 per million input tokens and $25 per million output tokens.

Anthropic positions Opus 4.7 as a notable improvement on Opus 4.6 in advanced software engineering, with particular gains on the most difficult tasks. It is less broadly capable than Claude Mythos Preview, but shows better results than Opus 4.6 across a range of benchmarks. The practical message: the Opus tier got meaningfully better at the work you use it for, while the price tag and API surface stayed put.

Software Engineering: Hand Off the Hard Stuff

Opus 4.7 handles complex, long-running tasks with rigor and consistency, pays precise attention to instructions, and devises ways to verify its own outputs before reporting back. Anthropic's framing is that users report being able to hand off their hardest coding work -- the kind that previously needed close supervision -- to Opus 4.7 with confidence.

That is the substantive change for day-to-day use. With Opus 4.6, most teams converged on a tight inner loop: hand the model a hard task, watch it work, intervene when it drifted. With Opus 4.7, the loop loosens -- you can spec the work, send it off, and come back to a finished result more often. The shift is supported by gains on the coding benchmarks Anthropic highlighted: SWE-bench Verified, SWE-bench Pro, SWE-bench Multilingual (with memorization screens applied), Terminal-Bench 2.0 on the Terminus-2 harness, MCP-Atlas, CyberGym, and an internal SWE-bench Multimodal implementation. The Opus 4.6 CyberGym score was also updated from 66.6 to 73.8 to reflect better elicitation.

The other piece of the coding story is /ultrareview, a new Claude Code slash command covered below. It pairs cleanly with the model's improved self-verification: Opus 4.7 is more willing and more able to read through changes carefully and flag issues a human reviewer would catch.

What Is New Around the Model: xhigh, /ultrareview, Auto Mode, Task Budgets

The launch ships four product changes alongside the model itself, all aimed at giving you more control over how much work Claude does on each turn.

The biggest is the new xhigh effort level, which sits between high and max. The effort parameter has been the main lever for trading off tokens against quality, and xhigh gives you a finer-grained option without jumping to max. Claude Code raised the default to xhigh for all plans today, so upgrading to Opus 4.7 inside Claude Code automatically gets you the new midpoint. Anthropic recommends starting with high or xhigh for coding and agentic use, and reaching for max only when the problem warrants it.

The /ultrareview slash command in Claude Code is a dedicated review session that reads through changes and flags bugs and design issues a careful reviewer would catch. Pro and Max Claude Code users get three free ultrareviews to try it. The intent maps cleanly onto how teams already use Claude: write code in one session, hand it to a fresh ultrareview session for a second opinion before opening a PR.

Auto mode -- where Claude makes decisions on your behalf instead of asking for permission on every action -- is now extended to Max users, having previously been limited to higher tiers. The framing in the launch post is careful: auto mode lets you run longer tasks with fewer interruptions and less risk than skipping all permissions outright. Pair it with Claude Code Routines and you have the building blocks for handing off a multi-hour agentic job and inspecting the result later.

Finally, the API gains higher-resolution images and task budgets in public beta. Task budgets let developers guide token spend so Claude can prioritize work across longer runs.

Vision: 2,576 Pixels on the Long Edge

Vision is substantially better in Opus 4.7. The model accepts images up to 2,576 pixels on the long edge -- roughly 3.75 megapixels -- which is more than three times the resolution prior Claude models supported. The practical impact: screenshots, scanned documents, dense slide decks, and UI mockups no longer have to be pre-processed or split before they fit into a single image input.

Higher resolution costs more vision tokens per image, which matters for budget planning. The upside is fewer multi-shot workarounds when the source material is visually dense. Combined with the improvements to instruction following described below, Opus 4.7 is a notably better collaborator on tasks that mix written instructions with screenshots or diagrams.

Vision is one of the benchmark categories where Opus 4.7 shows improvement against Opus 4.6, alongside Office tasks, Document reasoning, Long-context reasoning, Biology, Long-term coherence, and Coding.

Real-World Work: Finance, Memory, Instruction Following

Beyond raw coding benchmarks, Anthropic emphasizes a specific cluster of professional-task improvements. Opus 4.7 is more tasteful and creative when completing professional tasks, and produces higher-quality interfaces, slides, and docs. The launch post calls out two third-party evaluations as anchors: state-of-the-art on the Finance Agent evaluation, and state-of-the-art on GDPval-AA, a third-party evaluation of economically valuable knowledge work across finance, legal, and other domains.

Internal testing showed Opus 4.7 a more effective finance analyst than Opus 4.6 -- producing more rigorous analyses and models, more professional presentations, and tighter integration across tasks. That is the kind of improvement that compounds in agentic settings, where each sub-task feeds the next.

Memory is the other piece. Opus 4.7 is better at file system-based memory: it remembers important notes across long, multi-session work and uses them to move on to new tasks that need less up-front context. This pairs naturally with the file-system-based memory patterns described in the Claude Code context management guide -- the same disciplines apply, and the new model is more reliable at following them.

Instruction following deserves its own callout because it has a direct migration implication. Opus 4.7 is substantially better at following instructions, which means prompts written for earlier models can produce unexpected results. Where previous models interpreted instructions loosely or skipped parts, Opus 4.7 takes them literally. A long-lived prompt that worked on Opus 4.6 by being slightly sloppy may surface edges you did not know were there. Anthropic recommends re-tuning prompts and harnesses on the way in.

The Cyber Safeguards Tier and the Cyber Verification Program

Project Glasswing, announced last week, highlighted the dual-use nature of frontier models for cybersecurity -- legitimate uses like vulnerability research and penetration testing on one side, and serious abuse risk on the other. Anthropic stated then that they would keep Claude Mythos Preview's release limited and test new cyber safeguards on less capable models first.

Opus 4.7 is the first such model. Its cyber capabilities are not as advanced as Mythos Preview -- during training, Anthropic experimented with efforts to differentially reduce these capabilities -- but the launch does ship safeguards that automatically detect and block requests indicating prohibited or high-risk cybersecurity uses. What Anthropic learns from the real-world deployment of these safeguards on Opus 4.7 will help shape the eventual broad release of Mythos-class models.

For security professionals doing legitimate work -- vulnerability research, penetration testing, red-teaming -- Anthropic introduced the new Cyber Verification Program, the supported path for security uses that might otherwise trip the new safeguards. If your team's work falls into one of those categories, joining the program is the right move ahead of any production deployment of Opus 4.7 in your security tooling.

The safety and alignment profile of Opus 4.7 is broadly similar to Opus 4.6, with low rates of concerning behavior across deception, sycophancy, and cooperation with misuse. There are improvements in honesty and resistance to malicious prompt injection attacks, and a modest weakening in one area: a tendency to give overly detailed harm-reduction advice on controlled substances. Anthropic's alignment assessment describes Opus 4.7 as "largely well-aligned and trustworthy, though not fully ideal in its behavior", and notes that Mythos Preview remains the best-aligned model the lab has trained. The full discussion lives in the Claude Opus 4.7 System Card.

Migration From Opus 4.6: Tokenizer and Token Usage

Opus 4.7 is a direct upgrade. Swap claude-opus-4-6 for claude-opus-4-7 and you are running the new model. There are two changes that affect token usage and warrant a brief migration plan.

The first is an updated tokenizer that improves text processing. The same input can map to more tokens on the new tokenizer -- roughly 1.0-1.35x depending on content type. The multiplier varies with the kind of text you send. Code, prose, and structured data all behave differently, so the only way to know your number is to measure on your traffic.

The second is more thinking at higher effort levels, particularly on later turns in agentic settings. This is part of why reliability on hard problems improves, but it does mean more output tokens. You have three knobs to control it: the effort parameter, task budgets in the API, and prompting Claude for conciseness.

Anthropic's own testing reports a net effect that favors the upgrade -- token usage across all effort levels improved on their internal coding evaluation. The recommendation is the same one you would give yourself: measure the difference on real traffic before rolling out widely. Run a small slice of production through Opus 4.7 and compare against the same workload on Opus 4.6. Anthropic published a migration guide alongside the launch, and the System Card covers the safety context.

Pricing and Where to Use It

Pricing is unchanged from Opus 4.6: $5 per million input tokens and $25 per million output tokens. That keeps Opus in the same tier with the same budget shape teams have been planning around.

Availability is broad. Opus 4.7 is live today on Claude products, the API, Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry. The API model name is claude-opus-4-7. If you run Claude inside a cloud you already have a contract with, the rollout should require no new procurement work.

Inside Claude Code, the most concrete day-one moves are upgrading to the new model, accepting the new xhigh default effort level, and trying /ultrareview on a real PR before it ships. Pro and Max users get three free ultrareviews to test the workflow. The desktop app side of the experience is covered in the Claude Code desktop redesign coverage -- Opus 4.7 is the model running underneath that redesigned surface, and the new xhigh default applies there too.

For agentic and scheduled work, the obvious pairing is with the Claude Code Routines launch from yesterday and the broader scheduled-agent comparison. Routines that previously needed close supervision -- triaging alert queues, sweeping stale PRs, drafting docs against a moving codebase -- are exactly the work Anthropic says Opus 4.7 can now handle with confidence. Pick one, swap the model name, raise the effort to xhigh, and watch what changes.