Anthropic Releases Claude 4 Model Family With Expanded Reasoning and Agentic Features

Anthropic released its Claude 4 model family on Thursday, introducing two new large language models, claude-opus-4 and claude-sonnet-4, aimed at advanced reasoning tasks, software development, and autonomous agent workflows. The release marks the company's most capable public model lineup to date and arrives as competition among frontier AI developers intensifies.

What Happened

Anthropic made claude-opus-4 and claude-sonnet-4 available through its Claude.ai platform and via API on Amazon Bedrock and Google Cloud Vertex AI. The company described claude-opus-4 as its most intelligent model to date, with particular emphasis on performance in coding benchmarks and extended reasoning tasks. claude-sonnet-4 is positioned as a balance between capability and efficiency, intended for high-volume production deployments.

Both models support an extended context window and are designed to operate within multi-step agentic workflows, allowing them to use tools, execute code, and take sequences of actions with reduced human intervention between steps. Anthropic also introduced a feature it calls "extended thinking," which enables the models to work through complex problems in a structured internal reasoning process before producing a final response.

Background

Anthropic was founded in 2021 by former OpenAI researchers, including CEO Dario Amodei and President Daniela Amodei. The company has positioned its work around AI safety research alongside model development. Its previous Claude 3 family, released in early 2024, included the Haiku, Sonnet, and Opus tiers and was widely used across enterprise software integrations.

The company has raised substantial capital, including a multi-billion dollar investment commitment from Amazon and participation from Google. Amazon has made Anthropic models a central component of its Bedrock managed AI service, and the models are also available on Google Cloud.

The Claude 4 release follows OpenAI's release of its GPT-4o and o-series reasoning models and Google's release of the Gemini 2.5 family, including Gemini 2.5 Pro. Benchmark competition among these three developers has intensified throughout 2024 and into 2025, with each company reporting leading scores on standard evaluations including MMLU, HumanEval, and SWE-bench, a coding-specific benchmark that measures a model's ability to resolve software engineering issues.

Model Capabilities and Benchmarks

Anthropic reported that claude-opus-4 achieved a score of 72.5 percent on SWE-bench Verified, a widely referenced benchmark for agentic coding performance. The company stated the model shows improved performance on long-horizon tasks, meaning tasks that require the model to maintain context and execute reliably across many sequential steps.

claude-sonnet-4 is described as offering faster response times and lower per-token cost compared to claude-opus-4, making it suitable for applications requiring high throughput. Both models support a 200,000-token context window.

The extended thinking feature is available on both models and can be toggled by developers through API parameters. Anthropic stated that extended thinking improves performance on mathematics, science reasoning, and complex instruction-following tasks.

Availability and Pricing

Both models are available immediately through the Claude API. claude-opus-4 is priced at fifteen dollars per million input tokens and seventy-five dollars per million output tokens. claude-sonnet-4 is priced at three dollars per million input tokens and fifteen dollars per million output tokens. Enterprise customers with existing Anthropic or AWS Bedrock agreements can access the models under current contract terms.

Access through Claude.ai is available to Pro and Team subscribers at no additional charge beyond existing subscription pricing.

What Happens Next

Anthropic has indicated that additional models in the Claude 4 family, including a Haiku-tier model optimised for lightweight tasks, are scheduled for release later in 2025.