【Introduction】Anthropic officially launches Claude Opus 4.7, with core upgrades focused on complex task execution, high-definition visual understanding, and more stable long-chain workflows. For average users, the most immediate changes are better instruction adherence, enhanced image comprehension, and outputs closer to final deliverables—while also noting that token consumption will increase significantly.

Last night, Anthropic officially released Claude Opus 4.7 and defined it as the most powerful available Claude model currently accessible at scale.

While not as explosive in performance as the previously leaked next-generation Claude Mythos Preview, Opus 4.7 is vastly superior to the previous widely usable Opus 4.6—achieving comprehensive dominance except for a slight dip in Agentic search capability.

The official keywords for this upgrade: complex tasks, stronger vision, more robust long-chain execution, and reduced need for human intervention.

Anyone still using large language models for document writing, screenshot analysis, presentation creation, or material compilation cannot avoid the transformative experience brought by Opus 4.7.

The biggest highlight of this update is the dramatic enhancement in visual capabilities—jumping from approximately 50% score on tests in Opus 4.6 to near-perfect performance in Opus 4.7!

This fills AI’s largest remaining visual gap and may have already quietly crossed the most critical threshold for replacing human labor.

GPT-5.4 Thinking evaluates the impact of Claude Opus 4.7's release on knowledge workers as follows:

The Key to This Upgrade

Lies in Complex Task Completion

Anthropic has centered the core improvements of Opus 4.7 on advanced software engineering and sustained long-duration task execution.

Users can now delegate high-complexity coding tasks previously requiring close supervision to Opus 4.7, which adheres strictly to instructions and proactively verifies outputs before delivering results.

In the API release notes, Anthropic refers to Opus 4.7 as the strongest general-purpose model currently available, specifically designed for complex reasoning and agent-driven coding scenarios.

The focus of large model competition is shifting from "how well it answers" to "whether it can complete the task." Being able to craft a polished response is no longer sufficient.

Can it clean up a lengthy document? Can it synthesize scattered materials into a deliverable artifact? Can it maintain focus for tens of minutes or even longer without deviation? These are what truly determine whether it can genuinely shoulder real-world workloads.

This emphasis is clearly visible in the official launch highlights of Opus 4.7.

Programming Is Just the Appetizer

SWE-bench Multilingual measures a model’s ability to fix real GitHub issues across multiple programming languages.

Opus 4.7 scores 80.5%, compared to Opus 4.6’s 77.8%—a gain of 2.7 percentage points.

On the surface, this appears to be a routine iteration. But the data on the right side of the same chart is far more telling—and will be discussed later.

Long Tasks Within 1M Tokens

GraphWalks is an OpenAI benchmark testing long-context capabilities, where a directed graph is encoded entirely within a 1M-token context window, challenging the model to traverse the graph.

Two evaluation methods: one is Parents (given a node, find all its direct predecessors), and another is BFS (Breadth-First Search) from a starting point to identify nodes reachable at a specific depth—a strict test for multi-step agent workflows.

In the Parents 1M test, Opus 4.7 improves from 71.1% to 75.1%—a standard 4-point gain.

In the BFS 1M test, Opus 4.7 surges from 41.2% to 58.6%—an impressive 17.4-point leap.

Let’s examine another scenario.

Vending-Bench 2 simulates operating an automated vending machine, measuring decision coherence over extended workflows.

Opus 4.6 ends with $8,018; Opus 4.7 achieves $10,937.

Same vending machine, same time window—Opus 4.7 earns 36% more.

The Agent’s Eyes Have Been Upgraded

ScreenSpot-Pro evaluates an agent’s screen localization accuracy.

Given a high-resolution desktop screenshot of professional tools like VSCode, Photoshop, or AutoCAD along with a natural language instruction, the model must locate specific UI elements. On high-resolution screens, target UI elements often occupy only 0.07% of the total image—demanding extreme visual precision.

Under low resolution without tool use, Opus 4.6 scores 57.7%, while Opus 4.7 reaches 69.0%—a 11.3-point improvement.

Switching to high resolution, Opus 4.7 achieves 79.5% without tool calls, and jumps to 87.6% when tool invocation is enabled.

In some benchmarks (e.g., XBOW), Opus 4.7’s visual performance nearly doubles compared to Opus 4.6—rising from 54.5% to nearly perfect 98.5%!

This creates a chasm in computer use capabilities between Opus 4.7 and Opus 4.6.

Returning to the earlier programming chart.

SWE-bench Multimodal is evaluated using Anthropic’s internal test harness.

It tests frontend JavaScript bug fixing, incorporating visual assets such as UI screenshots and design mockups—requiring the model to integrate both image and code context.

Scores jump from Opus 4.6’s 27.1% to Opus 4.7’s 34.5%—a 7.4-point increase.

The key upgrade in Opus 4.7’s programming capability lies in its ability to understand screen content. With upgraded eyes, the mind can handle more complex tasks.

GPT-5.4 and Gemini 3.1 Pro Couldn’t Keep Up

After self-comparisons, let’s now assess how Opus 4.7 stacks up against established rivals.

GDPval-AA is an evaluation based on Artificial Analysis’s dataset derived from OpenAI GDPval.

It covers 44 knowledge worker professions and 9 major GDP sectors, with tasks drawn from real deliverables by senior professionals (average 14 years of experience). The AA version runs models in agent loops, using blind pairwise comparisons to calculate Elo ratings.

Opus 4.7 scores 1753, Opus 4.6 scores 1619, GPT-5.4 scores 1674, and Gemini 3.1 Pro scores 1314.

Opus 4.7 outperforms GPT-5.4 by 79 points and Gemini 3.1 Pro by 439 points.

OfficeQA Pro is an enterprise-level reasoning benchmark developed by Databricks, using corpus from nearly 100 years of U.S. Treasury bulletins—89,000 pages of PDFs and 26 million numerical entries. Models must precisely locate documents, parse tables and text, and perform cross-document analytical reasoning.

Here, Opus 4.7 achieves 80.6%, while Opus 4.6 scores only 57.1%; GPT-5.4 and Gemini 3.1 Pro trail at 51.1% and 42.9%, respectively.

In other words, Opus 4.7 is 1.6× faster than GPT-5.4 and 1.9× faster than Gemini 3.1 Pro.

The Most Explosive Leap: Biology

Turning to the final chart: Structural Biology, molecular reasoning.

Opus 4.6 scored just 30.9%. Opus 4.7 jumps to 74.0%.

From 30% to 74% in a single version iteration—a 2.4× improvement.

This represents the most dramatic leap across any benchmark tested.

What Average Users Notice First

Are Three Key Changes

The first change: stronger instruction following.

Anthropic states that Opus 4.7’s instruction-following capability has improved dramatically—where past models often loosely interpreted or missed details, Opus 4.7 is far more inclined to execute each step precisely.

The trade-off: older prompts may produce unexpected results, requiring users to rephrase their inputs.

For ordinary users, this directly reduces prompt engineering mysticism—writing requirements, formatting specs, and setting constraints becomes far more effective.

The second change: Claude sees images in greater detail.

Opus 4.7 supports images up to 2576 pixels in longest edge—approximately 3.75 million pixels, more than triple the capacity of prior Claude models.

Officially highlighted use cases include dense screenshots, complex charts, fine-grained structural diagrams, and tasks requiring pixel-level reference.

In real-world applications, this translates to accurately interpreting densely packed data screenshots, identifying product prototype details, extracting information from intricate flowcharts, and minimizing detail loss when reading high-resolution posters or reports.

The third change: output is closer to a production-ready deliverable.

Anthropic notes that Opus 4.7 demonstrates superior aesthetic sense and creativity in interface design, slide decks, and document creation.

It performs better in file-system-based memory retention, remembering key annotations across multiple rounds and sessions, reducing redundant background explanations.

For those frequently refining materials, organizing projects, or repeatedly revising the same content, this improvement feels far more tangible than raw benchmark gains.

This Release

Safety Is Equally Prioritized

Just a week prior, Anthropic unveiled Project Glasswing, explicitly addressing risks and rewards of frontier models in cybersecurity.

Opus 4.7 is the first model publicly deployed under this new framework. Officially, its cybersecurity capabilities are weaker than Mythos Preview, and it includes automated detection and interception of high-risk cyber requests as safety guardrails.

Compliance and security researchers may apply to join the new Cyber Verification Program.

From a safety assessment standpoint, Opus 4.7’s overall safety profile closely resembles Opus 4.6—stronger in honesty and resistance to malicious prompt injection, but slightly weaker in certain subcategories.

Anthropic concludes that Opus 4.7 is “reasonably reliable and trustworthy” overall, though still falling short of ideal.

This signals that Anthropic did not present this release as a cost-free, full-scale leap forward.

Who Benefits Immediately

And Who Should Proceed Cautiously

The first beneficiaries are clear: developers, analysts, legal professionals, researchers, and anyone who frequently handles documents, spreadsheets, and presentation materials.

Early feedback from official testing shows many partners consistently report: more stable long workflows, stronger error recovery, and significant improvements in document reasoning, code review, data analysis, and long-context tasks.

Places requiring caution are already documented in the official release notes.

Higher-resolution images consume more tokens—users should compress images if they don’t require such detail.

Opus 4.7 also features a new tokenizer (Tokenizer). The same input may generate approximately 1.0 to 1.35 times more tokens, and output tokens under high Effort will also increase.

For average users chatting directly in the Claude app, this primarily affects usage quotas and response experience.

For users and team customers leveraging APIs like Llama and Hermes Agent, this translates into a tangible cost variable.

Fortunately, pricing for Opus 4.7 remains consistent with Opus 4.6 and 4.5—no price hike—but the cost itself is already prohibitively high...

The Signal Anthropic Is Sending

Is Now Clear

The Opus 4.7 release makes Anthropic’s current strategic bets unmistakably clear: long-task execution, visual understanding, tool orchestration, and low-supervision delivery are being bundled into the next phase’s dominant battleground for large models.

Alongside the launch, Xhigh Effort (thinking level between high and max), Task Nudgets public beta, and /ultrareview in Claude Code all revolve around this direction.

Beyond the official announcement, Claude also published a 232-page system card detailing further insights—limited space prevents full coverage here.

For average users, the most direct impression of Claude Opus 4.7 is that once instructions are clearly given, it is more likely to get things right, see images in finer detail, and produce outputs ready for immediate use.

Large models are moving from mere conversation to actual productivity—this step takes a giant leap forward.

The strongest productivity model capable of truly excellent work has evolved from Opus 4.6 to Opus 4.7.

Source: Singularity Weekly