Claude Opus 4.7

Anthropic releases flagship large language model, achieving significant improvements in advanced software engineering and visual understanding

large language model

https://www.anthropic.com/news/claude-opus-4-7

Visit Website

In-depth Report

Claude Opus 4.7 is Anthropic's flagship large language model released on April 16, 2026, achieving significant improvements in advanced software engineering, visual understanding, and instruction following. This model supports higher resolution image processing (up to 2576 pixels), and pricing remains the same as Opus 4.6 (input $5/million tokens, output $25/million tokens). However, after the release, user evaluations were polarized, and some users reported that the model experienced regression in performance, mainly reflected in the increase in output tokens and lengthy responses.
Claude Opus 4.7 was released by Anthropic, a star company in the AI field, on April 16, 2026. The new version comes just two months after the last model upgrade, keeping in line with the company's previous update cadence. Anthropic is also developing a more powerful Mythos "myth" model. This model is currently only available for trial use by a small number of top institutions to find solutions to the "AI network catastrophe". It may not be available in the public market in the short term. Anthropic says there is an all-around capability gap between Claude Opus 4.7 and Mythos. Opus 4.7 is a high-end model for developers and programming enthusiasts, focusing on high-end software development capabilities.
Advanced Software Engineering Capabilities: Opus 4.7 delivers significant advancements in advanced software engineering. It performs especially well in complex coding tasks and long-running multi-step workflows. The new version is better at following instructions strictly, can verify whether your answers are correct before outputting, and is more comfortable handling difficult coding tasks that require in-depth reasoning. In professional evaluations, Opus 4.7 showed a 79-point advantage over the second place in the GDPval-AA test, and the hallucination rate dropped by 25 percentage points to 36%. This improvement comes from models that are more inclined to acknowledge knowledge blind spots rather than make up answers. Long text processing capabilities have also been improved. In the MRCR v2 test with 1 million token contexts, the new version showed stronger information retrieval accuracy. Enhanced visual capabilities: The new version has significantly improved visual capabilities and supports higher resolution image processing, up to 2576 pixels (approximately 3.75 million pixels), which is more than 3 times higher than the previous generation. This enables the model to more accurately understand chemical structures, complex technical diagrams, etc., and demonstrate higher taste in tasks such as interface design and data visualization. Command following and safety: Opus 4.7 has significantly improved command following, like a well-trained working dog, able to execute commands more accurately. The model has a built-in network security protection mechanism that can automatically detect and block malicious network security requests. It also supports legitimate network security work such as vulnerability research, penetration testing, and red team drills. Comparison with the previous version: Compared with Opus 4.6, the main improvements of the new version include: visual resolution increased by more than 3 times, instruction following consistency enhanced, long task stability improved, and professional output taste improved.
Pricing for Claude Opus 4.7 remains the same as Opus 4.6, charging $5 per million input tokens and $25 per million output tokens. The same pricing means users can get the enhanced capabilities of the new version at the same cost. However, some users reported that the new version produces more output tokens and more lengthy replies at higher effort levels, and the actual cost of use may be higher than the previous version. Users need to note that prompt words may need to be adjusted when migrating, because the new version has a stronger ability to follow instructions, and prompts originally designed for earlier models may produce different results. Available platforms: Anthropic API (Model ID: claude-opus-4-7), Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, Claude Pro / Max / Team / Enterprise subscription plans
Some professional reviews have given positive reviews of Opus 4.7. In the authoritative benchmark test, Opus 4.7 showed significant advantages, leading in running scores. Improved visual abilities are recognized and appreciated for producing higher quality interfaces, slides, etc. The reduction in hallucination rates is considered an important improvement, and the model is more likely to acknowledge knowledge blind spots rather than make up answers. However, user reviews were polarizing after release. In the ClaudeAI community of Reddit and Hacker News, a large number of users reported that Opus 4.7 experienced serious performance regressions. The main problems include: Lengthy output: The new version has become very "liney" and a lot of useless explanations are added to the reply, resulting in a significant increase in token consumption. Hallucination problem: Although the official claims that the hallucination rate has decreased, users report that calculation-intensive tasks are full of dangerous hallucinations that are not easy to detect. Lazy tendency: Users report that the model has become more "lazy" and unwilling to think deeply about complex problems. Price experience: Although the pricing has not changed, users feel that "the price has increased by 50%" due to the increase in output tokens. A large number of old users have called on Anthropic to "return me 4.6", which shows that there is a gap between the actual experience of the new version and the official promotion.
Anthropic has been in a rare period of glory in recent months. The improvement of the capabilities of Claude Code and Claude Cowork has significantly increased the outside world's evaluation of Anthropic's engineering capabilities. Claude has always been one of the "best writers" in users' minds. Even after the controversy related to the US Department of Defense attracted attention, Claude once topped the list of downloads in the App Store. However, the release of Opus 4.7 has divided opinions on Anthropic. On the one hand, the official benchmark test data is indeed eye-catching; on the other hand, there is a clear gap between the actual user experience and the running scores. This contrast between "running score champions" and "user overturns" is worth pondering.
Gap between marketing and reality: The three major upgrades officially promoted were questioned during actual testing, and there was a gap between user feedback and official promises. Hallucination problem: Although benchmarks show a decrease in hallucination rates, users still encounter subtle error messages in real-world use Cost issue: Although the unit pricing has not changed, the increase in output tokens has led to an increase in actual usage costs. Version rollback demand: A large number of users hope to roll back to Opus 4.6, reflecting a lack of recognition of the performance of the new version.
Who is it suitable for: Developers who need to handle complex long-running tasks, multi-modal application scenarios with high requirements for visual understanding, enterprise-level applications that require strict instruction compliance, professional fields such as legal document analysis and life science patent workflow. Who it is not suitable for: Heavy users who pursue concise and efficient responses, cost-sensitive projects, and old users who expect an experience similar to Opus 4.6. Alternative: If you are not satisfied with Opus 4.7, you can consider using Opus 4.6 or wait for subsequent version updates. Anthropic’s Sonnet model is also a good value for money option.
Claude Opus 4.7 has achieved significant improvements in technical indicators, and the official running scores are outstanding. However, there is a clear gap between the actual user experience and the official promotion. User reviews are highly polarized, with some users even nostalgic for the previous version. For enterprise users, it is recommended to conduct a small-scale test before deciding whether to migrate. For individual developers, you can choose an appropriate model version based on specific usage scenarios.

User Reviews

Amber.LongX

—

After using Opus 4.7 for a week, my coding ability has indeed improved, but the output is too long, and it is a long paragraph in one sentence, which makes me exhausted.

CarlHarris_2024241

—

Give me back the 4.6! 4.7 Too lazy, unwilling to think deeply about complex issues, and just confused when given a task.

EdvinRøise

—

The measured hallucination is worse than 4.6. There are a lot of subtle errors when doing mathematical calculations. Be careful.

RalphHart_Plus

—

The price has not changed but the output token has almost doubled, and the actual cost has increased by 50%. The scheme is too deep.

Andrea_MooreSr77

—

The programming ability is really strong, SWEBench Pro went from 53.4% to 64.3%, surpassing GPT-5.4.

MIste

—

The visual ability has indeed improved by 3 times, and the pictures can be viewed more accurately. There is nothing to say about this.

EMkin

—

It follows the instructions very accurately. I asked it to skip a certain step and it actually did it. It had automatically made up for it before I changed it.

Eugene_Baker_2022

—

The cybersecurity function is well received and malicious requests are automatically blocked. The security team is overjoyed.

PDiazIII

—

I ran a big project using Opus 4.7, and I ate 100 tokens in one afternoon, which made me feel sad.

Jean821

—

The quality of the generated code is indeed high, but can you not explain a lot every time?

PDiaz_7715

—

The MRCR v2 test performance of 1 million token context is stronger, and long text processing is indeed improved.

Douglas839_m

—

The hallucination rate dropped to this level is already very impressive, at least it will admit that it does not understand.

Aaron.HallZ

—

The visual resolution is mentioned as 2576 pixels. Looking at the technical chart, it is finally clear and I am touched.

GAbak

—

4.7 is released, the score looks very strong, but in practice emmm...you know.

DorothyYoung007

—

The gap between him and Mythos is still too big. Mythos is his biological son, 4.7 drinks soup.

JacobPerez_77

—

So delicious! Your programming ability has gone up a level, and you can make a lot of money with this upgrade.

xMircoFreitas_dev

—

After testing, GDPval-AA is 79 points ahead of the second place, which is a lot better.

WhaleAlertHall

—

Opus 4.7 yyds! The new model is so powerful and can handle complex tasks easily.

DrNurdanAkgül_88

—

I started using it on the day it was released, and overall I am satisfied with it, except that the token consumption is higher than 4.6.

Claude Opus 4.7

In-depth Report

User Reviews

Similar Products