Claude Opus 4.7: a new tier of AI-assisted engineering

On April 16, 2026, Anthropic announced the general availability of Claude Opus 4.7 — the most complete Opus model released to date. On top of being a direct evolution of Opus 4.6, 4.7 delivers concrete leaps in software engineering, computer vision and execution of long tasks with more context. For anyone living the day-to-day of DevOps, SRE and Platform Engineering, this release is more than a cosmetic update — it's a real change of tier.

What actually changed in Opus 4.7

Most of the gains come across three axes:

• Software engineering: +13% on a 93-task internal coding benchmark versus Opus 4.6, and 3x more tasks resolved in production on Rakuten-SWE-Bench. On CursorBench it jumped from 58% to 70%.
• Computer vision: support for images up to 2,576 pixels on the long side (approximately 3.75 megapixels) and 98.5% accuracy on a visual acuity benchmark — against 54.5% for Opus 4.6.
• Long-context reasoning: better precision on complex instructions and stronger execution of tasks with multiple chained steps.

On the legal side, the model reached 90.9% accuracy on BigLaw Bench, reinforcing the usefulness of Opus 4.7 in domains that demand terminological precision.

Operational news: xhigh and task budgets

One of the most interesting changes for anyone automating pipelines with Claude Code is the introduction of a new effort level: xhigh, positioned between high and max. This level gives more granular control over the trade-off between intelligence and cost — especially useful in complex coding tasks, where max is often overkill.

In Claude Code, the default is now xhigh. It's important to keep in mind that, at this effort level, Opus 4.7 can cost 20 to 30% more than Opus 4.6 at maximum effort. Another change is the tokenizer update: it improves text processing, but for some contexts the same input produces more tokens — which can impact spend unexpectedly.

Anthropic also launched, in public beta, the task budgets feature, which guides the model on how many tokens it can spend per task, and the /ultrareview command in Claude Code, designed for deep code reviews.

Pricing: the same, but with an asterisk

Per-token pricing stays the same as Opus 4.6:

• $5 per million input tokens
• $25 per million output tokens

In practice, however, two factors move the total cost of your operation:

1. New tokenizer: the same prompt can produce more tokens in certain contexts.
2. Higher effort by default: the model tends to think harder on difficult problems, producing more output tokens.

If your organization is already a heavy Opus user, monitor the spend report during the first days and review your spend controls before the bill surprises you.

Access and defaults by license

Standard licenses continue to use Claude Sonnet 4.6 as the default model. Premium licenses now use Claude Opus 4.7 by default across all environments: Cowork, Claude Code and chat. This applies automatically on Claude.ai, the Anthropic API, Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry.

For teams that manage access centrally, Anthropic recommends using server-managed settings to define default and available models per team — avoiding users running expensive workloads without oversight.

Why this matters for DevOps, SRE and Platform Engineering

The benchmark numbers are pretty, but what actually changes in operations?

• More reliable infrastructure agents: 3x more tasks resolved in production means agents that execute runbooks, perform remediation and operate Kubernetes with far less human intervention.
• Code review with fewer false positives: the CursorBench jump and /ultrareview turn Claude into a viable PR reviewer, not just autocomplete.
• Multimodal observability: with better vision, dashboards, architecture diagrams and incident screenshots can be fed directly into the model for triage.
• IaC automation: better long-context results mean the model can reason over an entire Terraform stack or a Helm chart without losing the plot.

Risks and caveats

The safety profile of 4.7 is similar to 4.6 — Anthropic itself describes the model as "largely well-aligned and reliable" in its alignment evaluation. Even so, a few points deserve attention before pushing to production:

• Review guardrails and tool-use policies, especially for agents with access to real infrastructure (clouds, clusters, secrets).
• Implement rate limits and task budgets from day one. xhigh can become a financial nightmare inside uncontrolled loops.
• Automate human checks at critical points: production deploys, merges to main, changes to IAM policies.
• Keep observability on the agents: tokens consumed, decisions taken, commands executed — everything needs to be auditable.

How CloudScript can accelerate your adoption

Adopting a model like Opus 4.7 in real DevOps and SRE environments isn't just a matter of plugging in an API key — it's rethinking pipelines, guardrails, FinOps and observability to get real value out of the capability gains. That's exactly where CloudScript comes in:

• Platform Engineering: we design Internal Developer Platforms that expose Claude as a native capability to your teams, with centralized cost and access policies.
• DevOps & SRE: we integrate Claude-based agents into CI/CD pipelines, remediation automation and operational runbooks — with rollback, approvals and end-to-end observability.
• FinOps for AI: we implement task budgets, per-team spend dashboards and proactive alerts so your Anthropic bill never becomes a month-end surprise.
• AIOps: we use Claude and other LLMs to correlate logs, metrics and traces during incidents — measurably reducing MTTR.

If you want to test Opus 4.7 in production with safety and cost predictability, talk to the CloudScript team. We assess your current stack, design the adoption plan and deliver the automation running — with SLOs and guardrails, not just promises.

Source and credits

Article based on the official announcement of Claude Opus 4.7:

Introducing Claude Opus 4.7 — Anthropic

Benchmark data, pricing and availability taken directly from Anthropic's official communication on April 16, 2026. Analysis and contextualization for the Brazilian market by the CloudScript Technology team.