Anthropic Launches Claude Opus 4.7

Anthropic has officially released Claude Opus 4.7, a direct upgrade to its predecessor, Opus 4.6. This update focuses on three key areas: a significant boost in software engineering capabilities, a major enhancement in visual understanding, and the introduction of a new cybersecurity protection framework.

It's important to note that Opus 4.7 is not the most capable model in the current Claude series—that position belongs to Claude Mythos Preview, released last week. However, Opus 4.7 is the first next-generation flagship model designed for broad deployment that has completed full safety verification.

Pricing remains the same as Opus 4.6: API input at $5 per million tokens and output at $25 per million tokens.

Background and Evolution of the Claude Opus Series

The Claude Opus line represents the highest tier in Anthropic's flagship model family. The previous model, Opus 4.6, launched in February 2026, demonstrated strong performance in complex reasoning, code generation, and long-context tasks.

Last week, Anthropic released Claude Mythos Preview, pushing the capability ceiling even higher, but kept it in a restricted release state for safety reasons. The launch of Opus 4.7 can be understood as an engineering trade-off between "peak capability" and "safe, wide deployment"—slightly less powerful than Mythos Preview, but systematically safety-verified and ready for general availability to all users and enterprises.

Benchmark trends show a steady progression across versions. For example, scores on GPQA Diamond have climbed from 81.00 (Opus 4.1) to 94.20 (Opus 4.7). A more dramatic leap is seen in SWE-bench Verified: while scores were nearly flat between Opus 4.5 and 4.6 (80.90 vs. 80.84), Opus 4.7 jumps to 87.60—a breakthrough that is one of the most noteworthy signals from this release.

Core Capability: A Quantum Leap in Software Engineering

The highlight of this update is the marked improvement in software engineering tasks, especially the most complex ones.

Early partner testing reveals substantial gains:

• Cursor: CursorBench pass rate of 70%, up from 58% with Opus 4.6.
• Rakuten: In internal SWE-Bench testing, it solved 3x more production tasks than Opus 4.6.
• GitHub Copilot: On a 93-task programming benchmark, the solve rate increased by 13 percentage points, including 4 tasks that neither of the previous two models could solve.
• Notion: Accuracy on multi-step workflow tasks improved by 14% over Opus 4.6, and tool-call error rates dropped to one-third.

A notable new behavior in Opus 4.7 is that it will self-verify its logic before starting work, proactively checking for correctness in its outputs rather than waiting for user feedback to fix issues. In simple terms, it has learned to "think clearly before acting."

Instruction following has also improved significantly. An important practical note: prompts tuned for Opus 4.6 may yield different results. The older model interpreted instructions more loosely, whereas Opus 4.7 follows them more literally. Developers are advised to re-test and adjust prompts after switching models.

Major Visual Upgrade: Up to 3.75 Megapixel Resolution

Opus 4.7 introduces a foundational upgrade to visual understanding: it now supports image resolutions up to 2,576 pixels on the long edge (approx. 3.75 megixels), which is over 3x the previous limit for Claude models.

This has several practical implications:

• Computer Use scenarios: It can now clearly read UI elements in high-density screenshots. In XBOW's tests, Opus 4.7 scored 98.5% on a visual precision benchmark, compared to just 54.5% for Opus 4.6, unlocking entire categories of previously unusable workflows.
• Science & Engineering Charts: Solve Intelligence reports it can now accurately interpret chemical structure diagrams and complex technical charts.
• OSWorld-Verified (Computer Interface Control): Opus 4.7 scored 78.00, a gain of over 5 points from Opus 4.6's 72.70, ranking second among the 12 models tracked.

This is a model-level change and requires no API parameter adjustments; images sent to the model are automatically processed at higher resolution. Note that higher-resolution images consume more tokens. If extra detail is not needed, down-sample images before sending them.

Cybersecurity & First Differentiated Protection Framework

This section provides important context for the Opus 4.7 release.

Last week, alongside Claude Mythos Preview, Anthropic launched Project Glasswing, an initiative to evaluate AI model risks and benefits in cybersecurity. They announced Mythos Preview would remain in restricted release until new protection mechanisms were validated. Opus 4.7 is the first officially released model to incorporate this cybersecurity protection framework.

Specifically, during Opus 4.7's training, Anthropic deliberately attempted to differentially reduce cybersecurity-related capabilities while deploying a protection framework that automatically detects and blocks high-risk cybersecurity requests. The strategy is to validate this system on a model with relatively lower capability before gradually applying the lessons to enable wider release of models at the Mythos level.

For security professionals with legitimate needs (vulnerability research, penetration testing, red team assessments), Anthropic has opened applications for its Cyber Verification Program.

Companion Features: xhigh Reasoning, Task Budgets, and /ultrareview

Beyond the model itself, this release includes several feature updates:

• New xhigh Reasoning Level: A new xhigh tier sits between high and max, offering finer control over the trade-off between reasoning depth and response latency. In Claude Code, the default reasoning level for all plans has been upgraded from high to xhigh. For testing Opus 4.7 on coding or agent tasks, Anthropic recommends starting with high or xhigh.
• Task Budgets (Public Beta): The Claude Platform API now includes a Task Budgets feature. Developers can specify token consumption allocation strategies for Claude, allowing the model to manage resources and prioritize work effectively during long-running tasks.
• Claude Code's New /ultrareview Command: Triggers a specialized code review that reads code changes line-by-line, flagging bugs and design issues a senior reviewer would catch. Pro and Max users receive 3 free trials. Additionally, Auto mode (automatic permission decision-making) is now also available to Max users.

Benchmark Highlights & Competitive Landscape

Opus 4.7 shows particularly strong rankings in programming and computer control, aligning with Anthropic's focus on software engineering.

Compared to leading competitors, its strengths are clear:

• Dominant in Programming: On SWE-bench Verified, Opus 4.7 scores 87.60, versus Gemini 3.1 Pro's 80.60 (GPT-5.4 data unavailable). It leads significantly on SWE-Bench Pro (64.30 vs. GPT-5.4's 57.70 and Gemini's 54.20).
• Competitive in General Reasoning: On GPQA Diamond, the three models are very close (Opus 4.7: 94.20, GPT-5.4: 92.80, Gemini 3.1 Pro: 94.30). Opus 4.7 has a slight edge on HLE (54.70).
• Weaker in Agent Information Gathering: On BrowseComp, Opus 4.7 scores 79.30, behind Gemini 3.1 Pro (85.90) and GPT-5.4 (82.70). It also trails GPT-5.4 on Terminal Bench 2.0 (69.40 vs. 75.10).

A Note on Pricing: Opus 4.7's standard API pricing ($5 input / $25 output) is notably higher than GPT-5.4 ($2.5/$15) and Gemini 3.1 Pro Preview ($2/$12). For tasks beyond its core coding strength, the choice should be made based on specific task requirements and cost considerations.

Migration Notes: Tokenizer Changes & Token Usage

Switching from Opus 4.6 to 4.7 involves two changes to plan for:

1. Tokenizer Update: The new tokenizer improves text processing, but the same input may map to more tokens. Testing suggests this factor ranges from 1.0x to 1.35x, depending on content type.
2. More Reasoning Output: Opus 4.7 generates more "thinking" tokens at higher reasoning levels, especially in later turns of agent-style tasks.

The good news: Based on Anthropic's internal coding benchmarks, when factoring in task completion quality and token consumption, overall efficiency improves—higher completion rates amortize the token cost. Nonetheless, it's recommended to test with real production traffic before a full switch. Anthropic provides a complete migration guide.

Availability & Pricing

Claude Opus 4.7 is available today across all platforms:

• All Claude products (claude.ai web, mobile, desktop)
• Claude API (model string: claude-opus-4-7)
• Amazon Bedrock
• Google Cloud Vertex AI
• Microsoft Foundry

Pricing remains the same as Opus 4.6: Input $5 / million tokens, Output $25 / million tokens.