Claude Opus 4.7 is Available: Anthropic Raises the Bar on Coding and Unlocks High-Resolution Images

Anthropic has released today Claude Opus 4.7, the new flagship model of the Claude 4 family, available immediately on all Claude products, direct APIs, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. The pricing remains unchanged from Opus 4.6: $5 per million input tokens and $25 per million output tokens. The API string to use is claude-opus-4-7.

What Changes Compared to Opus 4.6

The most evident improvements relate to advanced software engineering, especially on the more difficult tasks. Anthropic describes Opus 4.7 as a model that "plans more carefully, verifies its outputs before reporting results, and maintains consistency on long, multi-step tasks." The most significant data comes from CursorBench: 70% resolution compared to 58% for Opus 4.6. On Rakuten-SWE-Bench, Opus 4.7 solves 3 times more production tasks than its predecessor, with double-digit improvements in Code Quality and Test Quality. CodeRabbit reports over 10% increase in recall on complex pull requests, with no regressions on precision.

Introducing Claude Opus 4.7, our most capable Opus model yet. It handles long-running tasks with more rigor, follows instructions more precisely, and verifies its own outputs before reporting back. You can hand off your hardest work with less supervision.

Claude Opus 4.7 – Claude (@claudeai) April 16, 2026

On the multimodal front, Opus 4.7 accepts images up to 2,576 pixels on the long side, about 3.75 megapixels, more than three times compared to previous Claude models. This feature enables use cases that were previously impractical: computer-use agents reading information-rich screenshots, data extraction from complex technical diagrams, and analysis of chemical structures. XBOW, which uses Claude for autonomous penetration testing, reports a jump from 54.5% to 98.5% on its visual acuity benchmark: a drastic change unlocking an entire class of jobs that were impossible on Opus 4.6.

Another concrete area of improvement is the following of instructions. Opus 4.7 interprets prompts literally, whereas previous models tended to interpret them more liberally or skip parts. The practical consequence is that prompts designed for Opus 4.6 may yield different results: Anthropic explicitly advises conducting some checks on session prompts and system prompts before migrating to production.

New Platform Features

Alongside the model, Anthropic introduces a new effort level named xhigh, positioned between high and max. In Claude Code, the default level has been raised to xhigh for all plans; for coding tests and agent use, Anthropic recommends starting from high or xhigh. The additional reasoning improves reliability on difficult problems but generates more output tokens, especially in subsequent turns of long agentic sessions.

Claude Code introduces the command /ultrareview, a session dedicated to code review that analyzes changes and flags bugs and design issues that a careful reviewer would notice. Pro and Max users receive three ultrareviews for free to try it. The auto mode, which allows Claude to make decisions autonomously reducing interruptions on long tasks, is now extended to Max users. On the API side, task budgets arrive in public beta, a tool to guide Claude's token spending on extended runs.

It is also noteworthy that the tokenizer has been updated: the same input may require 1.0–1.35× more tokens than before, depending on the type of content. Anthropic provides a dedicated migration guide and suggests measuring the difference on real traffic before assuming costs remain identical.

The Cybersecurity Context: Mythos Preview and Project Glasswing

Opus 4.7 fits into a release strategy explicitly conditioned by the capabilities of Claude Mythos Preview, Anthropic's most powerful model that has not been made publicly available. Mythos Preview, announced in early April as part of Project Glasswing, has demonstrated during internal tests the ability to discover and exploit zero-days autonomously across all major operating systems and browsers, including a 27-year-old vulnerability in OpenBSD and one in FFmpeg dating back to 2009. For this reason, Anthropic chose to limit access to a consortium of selected companies.

Opus 4.7 is the first model on which Anthropic tests cybersecurity safeguards before arriving at a potential public release of a Mythos-class model. Its cyber capabilities are intentionally reduced compared to Mythos Preview (Anthropic speaks of active attempts at "differential reduction" during training), and the model includes automatic filters that block high-risk requests in cybersecurity. Security professionals with legitimate uses (vulnerability research, penetration testing, red-teaming) can join the new Cyber Verification Program.

Opus 4.7's safety profile and alignment are similar to those of Opus 4.6. On some measures, such as honesty and resistance to prompt injection attacks, Opus 4.7 improves. The alignment assessment defines it as "largely well-aligned and reliable, though not ideal in every behavior." Mythos Preview remains the best-aligned model according to Anthropic's internal evaluations.

Memory and Professional Work

Opus 4.7 introduces significant improvements in filesystem-based memory use. The model remembers important notes over multiple long sessions and uses them to proceed on new tasks with less context to provide each time. Anthropic reports better results also on GDPval-AA, a third-party evaluation of economically relevant work in finance and legal areas, and on the Finance Agent evaluation. Internally, tests show more rigorous financial analysis, more professional presentations, and more coherent integration across different tasks compared to Opus 4.6.

On the legal front, Harvey reports a 90.9% accuracy on high effort BigLaw Bench, with improved management of ambiguous document editing tasks and correct distinction between assignment provisions and change-of-control provisions, a type of error that has historically challenged frontier models. Databricks reports 21% fewer errors compared to Opus 4.6 on document reasoning tasks.