[Revised February 9, 2026]

Claude Sonnet 4.5 and Claude Code 2.0: New Model and Development Features

Anthropic's September 29, 2025 release of Claude Sonnet 4.5 marked a major upgrade to its coding-specialist AI, bundled with Claude Code 2.0, an enhanced developer environment. Sonnet 4.5 was hailed as "the best coding model in the world" at launch (^[1]), delivering huge gains in reasoning, mathematics and "using computers" compared to previous Claude models. Claude Code 2.0 introduced a host of new tools – including checkpoints, an IDE extension, parallel agents and automation hooks – that make AI-assisted coding faster, safer and more flexible. Since launch, Anthropic has continued its rapid pace: Haiku 4.5 (October 2025) brought Sonnet 4-level performance at one-third the cost, Opus 4.5 (November 2025) delivered flagship intelligence for agentic workflows, and Opus 4.6 (February 2026) introduced agent teams and a 1M-token context window (^[2]). Together, these updates have transformed complex software development into a true "colleague"-like AI collaboration, with Sonnet 4.5 remaining at $3/$15 per million tokens (^[3]).

Claude Sonnet 4.5: A Leap in AI Coding Capability

Claude Sonnet models are Anthropic's line of AI optimized for software development, offering a balance of speed and intelligence. Sonnet 4.5 delivered state-of-the-art performance on real coding tasks and benchmarks at launch. Anthropic reported 77.2% accuracy on the SWE-bench Verified coding benchmark (200K token context), versus roughly 60% for most competitors (^[4]). With high-compute techniques it can even reach 82.0%. In practical terms Sonnet 4.5 far outperforms earlier models on real-world software tasks: on the "OSWorld" benchmark (testing operating-system usage and application navigation) it scored 61.4%, up from 42.2% for Sonnet 4 just months earlier (^[5]). Sonnet 4.5 remains the recommended model for balanced speed-to-intelligence coding tasks, while the newer Opus 4.6 (released February 5, 2026) now serves as the most intelligent option for complex agentic workflows (^[6]).

• Benchmarks: Sonnet 4.5 achieved 77.2% on SWE-bench Verified (^[4]) and 61.4% on OSWorld tasks (^[5]), dramatically above prior Claude results.
• Extended reasoning: It shows large gains on math (AIME), agentic planning (τ²-bench), finance simulations, etc.
• Focus & autonomy: In Anthropic’s tests Sonnet 4.5 could stay on task for 30+ hours straight without human guidance (^[7]) (^[8]). Engineers watched it build an entire app — including standing up databases, buying domains and even doing a SOC2 security audit — over a full day of operation (^[7]). (For comparison, Anthropic’s previous model could only work reliably for ~7 hours at a stretch (^[8]).)

These improvements make Sonnet 4.5 much more reliable on lengthy, multi-step development jobs. Early customers report drastically better results on complex refactors and bug fixes. For instance, one benchmark showed the model’s code-edit accuracy jump from 91% (9% error rate) with Sonnet 4 to 100% (0% errors) with Sonnet 4.5 on internal tests (^[9]). Put simply, Sonnet 4.5 writes and understands code with far fewer mistakes, even on large codebases.

Sonnet 4.5 also exhibits better domain understanding and alignment. Anthropic calls it "the most aligned frontier model" they've released, with lower rates of hallucination or inappropriate content than earlier versions (^[10]). It was released under AI Safety Level 3 protections, with classifiers to detect dangerous inputs/outputs and enhanced defenses against prompt injection attacks. It has improved capacities in specialized areas – e.g. finance, legal, and medical reasoning – so it can answer enterprise queries more accurately. The model's communication style has been tuned as well: it gives concise, fact-based progress updates and can skip verbose summaries after tool calls to maintain momentum (^[11]). (Developers can still prompt it for more explanation if desired, but by default it avoids needless chit-chat.) For creative outputs, Sonnet 4.5 is solid too – it matches or exceeds prior models on tasks like slide/animation generation and producing polished, well-designed content (^[12]).

Importantly for developers, Sonnet 4.5 introduced new context-management APIs to handle long sessions. A Memory Tool (beta) lets Claude store and recall information outside the active chat window – effectively giving it a persistent "notepad" or knowledge base (^[13]). And Context Editing rules can automatically prune older tool outputs as the token budget fills up (^[14]). Combined, these allow the model to maintain coherence over very long chats (think dozens of files or hours of work) without losing track. There's also a new stop reason flag (model_context_window_exceeded) so applications can detect when the model just ran out of space (^[15]). Both Sonnet 4.5 and the newer Opus 4.6 now support a 1M-token context window in beta, further extending how long sessions can run without losing coherence (^[6]). All of this helps Claude stay on point during extended development cycles.

Claude Code 2.0: Enhanced Development Environment

Alongside Sonnet 4.5, Anthropic revamped Claude Code (their AI-powered coding CLI and IDE) to version 2.0. At launch, the core of Claude Code was Sonnet 4.5 (replacing Opus 4.1), though as of February 2026 it defaults to Opus 4.6 for even stronger agentic performance. The 2.0 release added major workflow features. In Anthropic's words, these let Claude Code handle "longer, more complex development tasks" with confidence (^[16]). Key tools introduced in 2.0 include:

• Automatic checkpoints: Claude Code 2.0 now auto-saves your workspace before each AI-made change, and you can instantly rollback if something goes wrong (^[17]). Pressing Esc twice (or the /rewind command) restores the previous state. You can choose to revert only the code, only the conversation, or both, so you always have an “undo” that spans multiple hours of AI work (^[18]). This safety net lets developers experiment boldly (large refactors or experimental features) without fear of irrevocable damage.

• Native IDE integration: A brand-new VS Code extension (beta) embeds Claude Code directly into the editor (^[19]). In VS Code (or compatible editors like VSCodium), a sidebar panel shows Claude’s proposed changes in real time with inline diffs. This is like having an AI doing code review live as it writes. At the same time, the terminal interface got polish: improved status display and a searchable prompt history (Ctrl+R) make it easier to track progress and re-use past queries (^[20]). Developers no longer have to constantly switch between terminal and IDE; changes appear smoothly in either environment.

• Subagents (parallel AI threads): One of the most powerful features is the ability for Claude Code to spawn subagents – specialized AI "assistants" that work in parallel. Each subagent has its own context window and tool set, optimized for a particular task (e.g. front-end coding, back-end logic, testing, etc.) (^[21]). This means the main agent can delegate subtasks to these agents and they all run at once. For example, Claude might spin up one subagent to implement an API endpoint while another writes its unit tests. Because they operate concurrently, overall development time can drop dramatically. (Anthropic's Agent SDK blog notes that subagents "enable parallelization: you can spin up multiple subagents to work on different tasks simultaneously" (^[22]).) With the February 2026 release of Opus 4.6, subagents evolved further into Agent Teams – a new experimental feature where multiple Claude Code instances coordinate as a team, with one session acting as team lead and others working independently in their own context windows (^[23]).

• Hooks (automation triggers): Claude Code 2.0 added user-defined hooks – scripts or commands that automatically run at specific points in the workflow (^[24]). For instance, after every code edit the system could run a linter, formatter (prettier, gofmt, etc.), or test suite, all without a prompt. Hooks turn best-practice maintenance tasks into enforced rules. Common use cases include auto-formatting files after every change, running tests before submitting a PR, sending notifications when Claude awaits input, or blocking any edits to production-critical files (^[25]). Since the initial release, hooks have expanded to include prompt-based hooks and agent-based hooks that use a Claude model to evaluate conditions, enabling judgment-based automation alongside deterministic rules. By encoding these as hooks, teams gain deterministic control over the AI's behavior (ensuring consistency) rather than leaving it to chance whether the model remembered to run tests.

• Background tasks: (Relatedly, Claude Code 2.0 can now launch and manage background processes. This lets the AI start long-running services – like local dev servers, build watchers, or CI pipelines – and then continue coding in the foreground. In effect, Claude can “keep the server running in the background” while it works on other tasks.)

• Agent SDK: Underlying Claude Code's features is the Claude Agent SDK (^[26]) (originally named Claude Code SDK, renamed to Agent SDK as the tool expanded beyond coding into general-purpose agent building (^[27])). This exposes the same tools, context-management systems and permissions framework that power Claude Code, now usable by developers to build custom agents for any domain. For example, one could build financial analysis agents, security auditing agents, deep-research agents, or helpdesk bots, all leveraging the parallelism, checkpoints and agent teams. In short, Anthropic is "giving developers the building blocks we use ourselves" (^[26]), so that the advanced workflows of Claude Code become available to everyone.

In practical terms, Claude Code 2.0 transforms the coding experience. The addition of checkpoints and undo means you can let the AI make sweeping changes (even risky refactors) and still recover if they fail. The VS Code extension and better CLI let developers work in the environment they prefer, with AI suggestions appearing seamlessly. And subagents/hooks turn Claude Code from a single AI helper into an entire team of AI agents, each following your project’s rules. As Anthropic notes, these features make Claude Code “more capable of handling sophisticated tasks” (^[28]).

Impact and Availability

Sonnet 4.5 is live in Anthropic's API and on the Claude chatbot across all plan tiers. You can invoke it via the API model name claude-sonnet-4-5-20250929 (or the alias claude-sonnet-4-5) at $3/$15 per million tokens (^[3]). For those needing maximum intelligence, Opus 4.6 is available at $5/$25 per million tokens with a 1M-token context window in beta and 128K max output (^[6]). Claude Code continues to receive rapid updates: the VS Code extension is out of beta, and the CLI is available on npm, Homebrew, and now Windows Package Manager (winget) (^[29]). Claude models are widely integrated into GitHub Copilot, Cursor, Amazon Bedrock, Google Vertex AI, and other developer platforms (^[30]).

The combined effect has been dramatic. Anthropic cites developer reports of 10× faster complex refactors and 77% higher success on real GitHub issues using Sonnet 4.5. With Opus 4.6, Claude has also uncovered more than 500 previously unknown zero-day vulnerabilities in open-source code, each validated by Anthropic's team or outside security researchers (^[31]). On the GDPval-AA benchmark of economically valuable knowledge work, Opus 4.6 outperforms OpenAI's GPT-5.2 by approximately 144 Elo points (^[2]). In short, the Claude model family now covers the full spectrum — from Haiku 4.5 ($1/$5) for speed-sensitive tasks, through Sonnet 4.5 for balanced coding, to Opus 4.6 for the most demanding agentic workflows — while Claude Code's ever-expanding tools (checkpoints, hooks, agent teams, IDE integrations) provide the controls needed to wield this intelligence safely. Together, they represent a defining step toward AI that truly partners on software engineering.

Sources: Anthropic's official announcements and docs (^[32]) (^[2]) (^[33]) (^[6]) (^[34]) (^[35]) (^[23]) (^[27]); press coverage (^[36]) (^[37]) (^[31]) (^[30]).