Anthropic has launched its next-generation artificial intelligence models, Claude Opus 4 and Claude Sonnet 4. The company asserts these models establish new industry benchmarks for coding, sophisticated reasoning, and advanced AI agent functionalities.
Anthropic positions Claude Opus 4 as “the world’s best coding model, with sustained performance on complex, long-running tasks and agent workflows.”, reportedly capable of sustained work on intricate, long-duration tasks for up to seven hours. This development significantly advances AI tools for complex problem-solving and software creation, providing developers and businesses with more potent and potentially autonomous systems.
The new models introduce “extended thinking,” a feature enabling them to use tools like web search during prolonged reasoning. They also feature improved memory, creating “memory files” when developers grant local file access, which enhances continuity over extended tasks.
Anthropic also made its Claude Code platform generally available, including new Integrated Development Environment (IDE) integrations. The company released several new API features to support the creation of powerful AI agents. The company says that with Claude 4, they are not just improving benchmarks but building a true collaborative partner for complex work.
However, the increased capabilities of Claude Opus 4 have led Anthropic to implement stricter “AI Safety Level 3” (ASL-3) safeguards. This decision follows internal testing that highlighted the model’s proficiency in advising on biological weapon production.
Jared Kaplan, Anthropic’s chief scientist, acknowledged these potential misuses to TIME, explaining their modeling suggests a concerning potential: “You could try to synthesize something like COVID or a more dangerous version of the flu—and basically, our modeling suggests that this might be possible.”
Kaplan emphasized that while not definitively confirming the model’s risk, Anthropic is opting for caution. This dual strategy of advancing AI power while heightening safety protocols highlights the complex challenges in the AI field.
New AI Capabilities and Performance
Anthropic’s Claude Opus 4 has demonstrated leading performance on several key industry benchmarks. It achieved a 72.5% score on SWE-bench for software engineering and 43.2% on Terminal-bench. Its ability to autonomously work for nearly a full seven-hour corporate workday on complex coding tasks marks a notable step forward.
Mike Krieger, Anthropic’s Chief Product Officer, informed CNBC about the new models’ impact on his writing, stating, “they’ve crossed this threshold where now most of my writing is actually … Opus mostly, and it now is unrecognizable from my writing.”
Claude Sonnet 4 also shows strong coding skills, scoring 72.7% on SWE-bench, and is presented as an optimal mix of high-end capability and practical efficiency. Both models are reportedly 65% less likely to use shortcuts or exploit loopholes compared to the earlier Sonnet 3.7, especially on agentic tasks. Early partners have provided positive feedback.
For example, Cursor describes Opus 4 as “state-of-the-art for coding and a leap forward in complex codebase understanding.” GitHub stated Sonnet 4 “soars in agentic scenarios” and will integrate it as the base for its new coding agent in GitHub Copilot.
These endorsements suggest tangible improvements in real-world applications, from understanding complex codebases to executing multi-file changes with greater precision.
Comparative Benchmark Highlights
While the new Claude 4 models showcase leading scores in coding benchmarks like SWE-bench and Terminal-bench, their performance across other key AI evaluations presents a competitive, nuanced picture.
In graduate-level reasoning, measured by GPQA Diamond, OpenAI’s o3 (83.3%) and Google’s Gemini 2.5 Pro (83.0%) currently set the pace, though Claude Opus 4 (79.6%) and Sonnet 4 (75.4%) are strong contenders, with Opus 4 matching OpenAI o3 when using Anthropic’s ‘high compute’ methods.
For multilingual Q&A tasks (MMMLU), Claude Opus 4 achieves an impressive 88.8%, matching OpenAI o3 at the top of this benchmark. However, in visual reasoning (MMMU validation), OpenAI o3 (82.9%) and Gemini 2.5 Pro (79.6%) maintain an edge over Claude Opus 4 (76.5%).
Similarly, in high school math competitions (AIME 2025), OpenAI o3 (88.9%) and Gemini 2.5 Pro (83.0%) lead in standard evaluations, though Claude Opus 4 shows significant improvement and can reach 90.0% with ‘high compute’ methods, surpassing others. Anthropic notes that these ‘high compute’ approaches, involving more intensive parallel processing, can further elevate Claude 4’s scores across several benchmarks.
Enhanced Developer Tools and Ecosystem
A key component of the announcement is the general availability of Claude Code. This platform now integrates directly into developer workflows via extensions for VS Code and JetBrains.
These integrations allow proposed code edits to appear inline within the IDE. Anthropic is also releasing an extensible Claude Code SDK, enabling developers to build custom agents. An example is the new beta of Claude Code on GitHub, which can respond to reviewer feedback or fix Continuous Integration (CI) errors, a process for automating software building and testing.
Further empowering developers, Anthropic introduced four new API capabilities. These include a code execution tool, a Model Context Protocol (MCP) connector, a Files API, and prompt caching for up to one hour.
The MCP connector is particularly significant. MCP is an open-source framework Anthropic initiated in November 2024 to standardize AI model interactions with diverse data sources. Anthropic’s API enhancements aim to significantly reduce development overhead for creating sophisticated AI agents.
The models are accessible via Anthropic’s API, Amazon Bedrock, and Google Cloud’s Vertex AI. Opus 4 is priced at $15/$75 per million input/output tokens and Sonnet 4 at $3/$15.
Balancing Innovation with Heightened Safety Measures
The introduction of ASL-3 for Claude Opus 4 signifies a critical application of Anthropic’s Responsible Scaling Policy (RSP). This voluntary policy aims to ensure safety measures evolve alongside model capabilities. ASL-3 measures are deemed necessary for AI systems that could substantially increase the ability of individuals with basic STEM backgrounds to create or deploy Chemical, Biological, Radiological, or Nuclear (CBRN) weapons.
The ASL-3 “defense in depth” strategy involves multiple protective layers. Enhanced “constitutional classifiers” scan for dangerous content, specifically targeting queries related to bioweapon creation. Anthropic actively works to prevent jailbreaks, monitors usage, and has a bounty program that led to patching one universal jailbreak.
Cybersecurity for the model’s neural network is also strengthened. These steps are taken because, as Kaplan stated, “We just saw COVID kill millions of people.”
This cautious approach reflects Anthropic’s awareness of profound societal risks. While ASL-3 is a positive development, external AI ethics researchers express concern about the broader challenge of voluntary self-regulation within the competitive AI industry. This comes as Anthropic experiences rapid growth, its annualized revenue reached $2 billion in the first quarter of 2025.
