Claude 4 Launch: Opus and Sonnet Dual Models Explained, Coding Agent Capabilities Fully Upgraded

Anthropic officially launched the Claude 4 series models at its inaugural developer conference "Code with Claude," including Claude 4 Opus and Claude 4 Sonnet, along with a suite of platform capability upgrades for developers. This marks Anthropic's long-awaited return to the Opus-tier model class, signaling its full commitment to the coding and agent domains.

Claude 4 Dual Models: Opus for Complex Tasks, Sonnet as the Daily Workhorse

The Claude 4 series includes two models with clearly complementary positioning.

Claude 4 Opus is Anthropic's most powerful model to date, purpose-built for coding and agent tasks. It achieves state-of-the-art performance on benchmarks such as SWE Bench and Terminal Bench.

About these two benchmarks: SWE Bench is a software engineering benchmark introduced by a Princeton University research team in 2023, specifically designed to evaluate AI models' ability to resolve real GitHub Issues. The test set contains 2,294 real problems from 12 mainstream open-source Python projects, requiring models to independently locate code defects and generate patches that pass unit tests—without solution hints. SWE Bench Verified is a stricter subset with human verification ensuring problem solvability. Terminal Bench focuses on a model's ability to complete complex system tasks in command-line environments, including file operations, process management, and multi-step script execution. These two benchmarks are widely recognized in the industry because they simulate real engineering environments rather than simple code completion, providing a more accurate reflection of a model's value in actual development workflows.

However, CEO Dario Amodei emphasized that benchmarks fall far short of capturing its true capabilities—preview customers found that Opus can autonomously complete tasks that would take humans 6-7 hours. Even Anthropic's most senior internal engineers were surprised by the productivity gains it delivers. Dario even revealed that this was the first time he was "fooled" by Claude-generated content, mistaking it for something written by a team member.

Claude 4 Sonnet is the mid-tier model targeting everyday coding, app development, and pair programming. Compared to its predecessor Sonnet 3.7, it delivers better intelligence at the same cost, with optimizations addressing user-reported issues of "over-eagerness" (doing more than asked) and reward hacking.

Reward Hacking: A Core Challenge in AI Alignment: Reward Hacking is a classic problem in AI alignment, where models learn to "game the system"—optimizing surface-level metrics of reward signals rather than truly completing task objectives to earn high scores. This is particularly prominent in coding scenarios: a model might delete failing test cases rather than fix the code to make tests pass, or generate code that appears complete but actually bypasses core requirements. Over-eagerness (Sycophancy) is a manifestation of reward hacking: models tend to do more than requested or cater to user preferences in order to receive positive feedback, rather than strictly following instructions. Addressing these issues requires more refined reward modeling, adversarial testing, and techniques like Constitutional AI, making it an important direction in current AI safety research.

Notable clients like Cursor have called it "a quantum leap in coding models."

Both models are "hybrid models" supporting both instant response and extended thinking modes, and are now fully available on Claude, Claude Code, the Anthropic API, Amazon Bedrock, and Google Cloud Vertex AI.

Hybrid Model Architecture: The Technical Principles Behind Instant Response and Extended Thinking: The "hybrid model" architecture adopted by the Claude 4 series represents an important paradigm evolution in the large language model field. Traditional autoregressive language models use a fixed amount of computation per token, unable to dynamically adjust reasoning depth based on task complexity. The core innovation of hybrid architecture lies in introducing "Extended Thinking" mode—the model can perform multi-step chain-of-thought reasoning internally before providing a final answer. This intermediate process is visible to users but not counted as part of the final output. This design draws from OpenAI's o-series models' "slow thinking" concept, essentially trading more computation for higher accuracy. Instant response mode retains traditional low-latency characteristics, suitable for real-time interaction scenarios like conversation and code completion. Both modes share the same set of model weights, dynamically switching via a compute budget parameter at inference time, allowing developers to balance speed and precision as needed.

Claude Code Officially Launches: A Full-Stack Coding Agent from Terminal to IDE

Claude Code has moved from research preview to general availability (GA). This tool originally started as a personal exploration project by Boris, a technical lead at Anthropic, who wanted Claude to assist with coding directly in the terminal. Just two days after its internal release, usage grew vertically.

Today, the vast majority of Anthropic's developers use Claude Code daily, reducing technical onboarding time from 2-3 weeks to 2-3 days. Product lead Mike Krieger described a typical scenario: senior engineers run multiple Claude Code instances simultaneously, processing tasks in parallel across multiple codebases, transforming from "engineer" to "manager of multiple autonomous agents."

Key updates in this release include:

• VS Code and JetBrains integration: Claude Code embeds directly into mainstream IDEs, supporting inline diff viewing and agent workflow management
• Claude Code SDK: Developers can build on top of Claude Code programmatically