Claude 3.5 Sonnet Released: Major Gains in Coding, Math, and Vision at Same Price

Anthropic has introduced Claude 3.5 Sonnet, its latest large language model, arriving just days after the company unveiled Claude 3 Opus and Claude 3 Sonnet in March. The new release marks another step in the company’s rapid development cycle and comes with measurable improvements across several standard benchmarks. According to details shared on the company’s blog and covered by Gizmodo, Claude 3.5 Sonnet outperforms its predecessor on multiple academic and real-world evaluations while maintaining the same pricing structure.

The model achieves a score of 59 percent on the GPQA benchmark, which tests graduate-level reasoning across physics, chemistry, and biology. That result places it ahead of GPT-4o and the earlier Claude 3 Opus. On the MATH benchmark, which measures problem-solving ability in competition-level mathematics, Claude 3.5 Sonnet reached 71.1 percent, showing a clear gain over the 60.1 percent posted by Claude 3 Opus. These numbers suggest the new version handles complex analytical tasks with greater accuracy.

Anthropic also highlighted gains in coding performance. The model scored 92 percent on HumanEval, a standard test of Python programming ability, and demonstrated strong results on SWE-Bench, a benchmark that evaluates how well systems can resolve real GitHub issues in large codebases. Early internal testing showed Claude 3.5 Sonnet fixing problems more efficiently than previous versions, which could appeal to software development teams looking for practical assistance.

Beyond raw benchmark numbers, the model introduces several new capabilities. One notable addition is an improved computer use feature that allows the AI to interact with a desktop environment by moving the cursor, clicking buttons, and typing text. This tool remains in public beta and requires careful monitoring because the system can still make mistakes or misinterpret instructions. Anthropic emphasized that users should review any automated actions before relying on them for important work.

The company also updated its Artifacts feature, which now supports interactive previews of code, diagrams, and web applications directly in the chat interface. Developers can generate a React component, for example, and immediately see how it renders and behaves without leaving the conversation. This tighter integration between generation and visualization reduces the steps needed to test ideas and could speed up prototyping workflows.

Claude 3.5 Sonnet maintains the same context window of 200,000 tokens that defined the Claude 3 family. That capacity allows the model to process long documents, code repositories, or extended conversations without losing track of earlier details. The output limit remains at 4,096 tokens for most users, though higher limits are available through the API for specific enterprise plans.

Pricing has not changed from the previous Claude 3 models. Input tokens cost three dollars per million, while output tokens are priced at fifteen dollars per million. This rate structure makes the new model cost-competitive with offerings from OpenAI and Google, especially for applications that require large context windows or frequent coding assistance.

Safety and alignment received continued attention during development. Anthropic applied its Constitutional AI approach, which uses a set of written principles to guide the model’s behavior. The company reported that Claude 3.5 Sonnet shows lower rates of harmful or biased responses compared with Claude 3 Sonnet, though independent testing will be needed to confirm these claims across diverse scenarios. The model still refuses certain categories of requests, such as those involving illegal activities or explicit content, consistent with earlier versions.

Access to the new model rolled out first to users of Claude.ai and the Claude mobile apps. Paid subscribers on the Pro plan gained immediate availability, while free users received limited access with daily message caps. Enterprise customers and developers using the Anthropic API could start integrating the model through standard SDKs in multiple programming languages.

The timing of the release surprised some observers because Anthropic had only recently launched the full Claude 3 family. The quick follow-up suggests the company has accelerated its training and evaluation pipelines, possibly taking advantage of improved infrastructure or more efficient scaling methods. In a statement, Anthropic executives indicated that future models would continue to arrive at a similar pace as research breakthroughs allow.

Industry analysts have begun comparing Claude 3.5 Sonnet directly with GPT-4o and Gemini 1.5 Pro. Early user reports shared on social platforms and developer forums indicate that the new Anthropic model feels more precise when handling technical queries and produces fewer factual errors on specialized topics. However, performance can vary depending on the exact prompt style and the domain of knowledge being tested.

One area where Claude 3.5 Sonnet shows noticeable progress is visual reasoning. The model can now interpret charts, screenshots, and handwritten notes with higher accuracy than its predecessor. This improvement expands the range of tasks the AI can support, from analyzing financial reports to assisting with technical documentation that includes diagrams.

Education professionals have started exploring how the model might help with tutoring and content creation. Because the system can break down complex concepts into step-by-step explanations, it could serve as a supplementary tool for students studying mathematics, physics, or computer science. Teachers have also experimented with using the model to generate practice problems and grading rubrics, though they stress the need for human oversight to catch occasional inaccuracies.

Creative professionals have mixed reactions. Some writers and designers appreciate the model’s ability to offer structured feedback on drafts or suggest alternative approaches to visual layouts. Others worry that increased AI fluency might reduce demand for certain entry-level creative tasks. These concerns echo broader discussions happening across many industries as language models grow more capable.

Anthropic continues to position itself as an organization focused on responsible development. The company maintains a public transparency commitment and publishes regular updates on its safety research. While the new model represents a performance jump, the firm reiterated that it still falls short of the hypothetical frontier systems that could pose significant societal risks. This measured stance contrasts with some competitors who have used more dramatic language when announcing their latest offerings.

Developers interested in building applications with Claude 3.5 Sonnet can access detailed documentation on the Anthropic website. The API supports streaming responses, function calling, and integration with external tools. Several popular frameworks have already added support for the new model, allowing teams to switch from earlier versions with minimal code changes.

Early benchmarks also include results from the MMMU test, which evaluates multimodal understanding across academic subjects. Claude 3.5 Sonnet scored 68.3 percent, placing it near the top of current publicly available models. This result indicates progress in connecting textual reasoning with image interpretation, a skill set that matters for applications ranging from medical image analysis to architectural design review.

The release has prompted fresh speculation about when Anthropic might introduce a Claude 4 series. Company representatives have avoided specific timelines but acknowledged that research on larger models continues. Observers expect any future version to build on the architectural foundations established in the Claude 3 and 3.5 families while incorporating lessons learned from real-world deployment.

Business users have taken particular interest in the improved coding abilities. Companies that rely on large software teams see potential for faster bug resolution and more consistent code reviews. However, most experts recommend keeping a human developer in the loop to verify logic, security implications, and alignment with project standards. Over-reliance on any single model could introduce subtle errors that accumulate over time.

The updated computer use feature has generated the most discussion among early testers. By allowing the model to control a virtual desktop, Anthropic has moved closer to creating an AI assistant that can perform multi-step digital tasks without constant human guidance. Current limitations include occasional misclicks and difficulty with interfaces that change rapidly. The company plans to refine this capability based on user feedback and hopes to expand the range of supported applications in future updates.

Educational institutions have begun piloting the model in controlled environments. Some universities are testing whether Claude 3.5 Sonnet can provide consistent feedback on student essays or programming assignments. Initial results suggest the AI can identify common mistakes and offer targeted suggestions, though it sometimes misses nuanced arguments or creative solutions that fall outside typical patterns.

As organizations integrate the new model into their workflows, questions about data privacy and compliance remain central. Anthropic offers enterprise plans with enhanced security controls and the option to process information within private VPC environments. These features address concerns from sectors such as healthcare, finance, and government that must meet strict regulatory requirements.

The rapid succession of model releases from Anthropic reflects a broader trend across the AI industry. Companies are competing not only on absolute performance but also on the speed with which they can translate research findings into production systems. This acceleration creates both opportunities and challenges for users who must evaluate new offerings while maintaining stable applications.

Claude 3.5 Sonnet represents a solid step forward without introducing radical changes to the underlying philosophy that guides Anthropic’s work. The model delivers measurable gains on established tests, adds practical new features, and remains available at familiar price points. For developers, researchers, and businesses already familiar with the Claude platform, the upgrade offers immediate benefits with relatively low switching costs.

Looking ahead, the AI community will watch closely to see how competitors respond. OpenAI, Google, and Meta each have upcoming releases planned, and the pace of progress suggests that benchmark records may continue to fall throughout the year. For now, Claude 3.5 Sonnet sets a high standard for balanced performance across reasoning, coding, and multimodal tasks while preserving the safety focus that has defined Anthropic since its founding. The coming months will reveal how effectively organizations can incorporate these capabilities into products and services that deliver genuine value to end users.