Key takeaways
- Opus 4.7 delivers notable improvements over Opus 4.6 in advanced software engineering, with users reporting confidence in handing off their hardest coding tasks without close supervision.
- The model supports image resolution up to 2,576 pixels on the long edge — roughly 3.75 megapixels — more than three times the capacity of prior Claude models.
- A new
xhigh effort level has been introduced between high and max, giving developers finer control over the reasoning-versus-latency tradeoff for difficult tasks. - Pricing holds steady at $5 per million input tokens and $25 per million output tokens, matching Opus 4.6 rates across all platforms.
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. Early-access partners confirmed significant gains: on one 93-task coding benchmark, Opus 4.7 lifted resolution by 13% over Opus 4.6, including four tasks neither Opus 4.6 nor Sonnet 4.6 could solve. On CursorBench, it reached 70% compared to 58% for Opus 4.6, and on Rakuten-SWE-Bench it resolved 3x more production tasks than its predecessor, with double-digit gains in code quality and test quality.
A major leap in vision and multimodal capabilities
The maximum image resolution increase to 3.75 megapixels is particularly significant for computer use agents, screenshot and document understanding workflows, and dense technical diagrams. One life sciences firm noted major improvements in reading chemical structures and interpreting complex patent diagrams, while an autonomous penetration testing company reported a jump from 54.5% to 98.5% on their visual-acuity benchmark. The model's coordinates now map 1:1 with actual pixels, simplifying operations like coordinate-to-image mapping that previously required manual scale-factor adjustments.
Instruction following, memory, and real-world knowledge work
Opus 4.7 is substantially better at following instructions precisely — a change significant enough that prompts written for earlier models may produce unexpected results, as the new model takes instructions more literally where previous versions interpreted them loosely or skipped steps. On the knowledge work front, the model is state-of-the-art on GDPval-AA, a third-party evaluation of economically valuable work across finance and legal domains. It also demonstrates improved file system-based memory, remembering important context across multi-session workflows and reducing the upfront context needed to resume long-running tasks.
Cybersecurity safeguards and the Mythos Preview context
Opus 4.7 is the first model on which Anthropic is testing new cyber safeguards — automatic detection and blocking of requests indicating prohibited or high-risk cybersecurity uses — as part of Project Glasswing, the company's framework for responsible capability deployment. Anthropic deliberately trained Opus 4.7 to carry lower cybersecurity capabilities than the unreleased Claude Mythos Preview, which remains restricted to a handpicked set of organizations including AWS, Apple, Cisco, Google, and JPMorgan Chase for defensive security work. Security professionals seeking to use Opus 4.7 for legitimate purposes such as vulnerability research, penetration testing, and red-teaming are invited to join the new Cyber Verification Program.
Alongside the model launch, Anthropic is releasing task budgets in public beta for API users, giving developers a way to set hard token ceilings on agentic loops so the model can prioritize work and finish gracefully as the budget is consumed — rather than cutting off mid-task. In Claude Code, a new /ultrareview command produces a dedicated review session that flags bugs and design issues a careful reviewer would catch, with three free ultrareviews available to Pro and Max users at launch. Auto mode has also been extended to Max users, enabling longer tasks with fewer interruptions. Opus 4.7 is available today across all Claude products, the Claude API, Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry.
Migration notes for developers
Developers upgrading from Opus 4.6 should account for two changes that affect token usage. First, Opus 4.7 uses an updated tokenizer that may process the same input at roughly 1.0x to 1.35x more tokens depending on content type. Second, the model thinks more at higher effort levels — particularly on later turns in agentic settings — producing more output tokens in exchange for greater reliability on hard problems. Anthropic recommends measuring the difference on real traffic before committing to production, and has published a migration guide covering prompt retuning and effort level adjustments.
References
