Anthropic launched Claude Sonnet 4.5 on September 29, 2025, calling it the best coding model in the world. It is the strongest Claude model for building complex agents and for using computers, with large gains in reasoning and math over earlier Claude models. It shipped the same day in the Claude apps and the Claude API, together with Claude Code 2.0 and a new Agent SDK. This guide explains what changed, what the benchmark numbers mean, and what it costs.
Claude Sonnet 4.5 at a glance: what changed
Claude Sonnet 4.5 launched on September 29, 2025 as Anthropic’s new middle model, between the smaller Haiku and the larger Opus. It is the company’s strongest model for coding, agents and computer use, with a 77.2 percent SWE-bench Verified score and a 61.4 percent OSWorld score, both reported by Anthropic.
Sonnet is the workhorse of the Claude family. Haiku is the small fast model, Opus is the largest and most capable, and Sonnet sits in the middle with the best balance of speed, cost and intelligence for everyday work. Sonnet 4.5 replaces Sonnet 4 as the recommended model, and Anthropic says developers should upgrade for all uses because it is a drop-in replacement at the same price. The model id for the API is claude-sonnet-4-5, and it works in the Claude apps on the web, desktop and mobile, including on an iPhone 17 or an Android 16 phone.
The launch was more than a model swap. Anthropic shipped Claude Code 2.0 with checkpoints and a native VS Code extension, a context editing feature and a memory tool for the API, code execution and file creation inside Claude chats, the Claude for Chrome extension for Max users, and the Claude Agent SDK for developers. The pace is fast: Sonnet 4.5 arrived less than two months after Anthropic’s previous model, Claude Opus 4.1, which shows how quickly the company is shipping new models.
Coding and the SWE-bench Verified result
Claude Sonnet 4.5 scored 77.2 percent on SWE-bench Verified, a test built from real software engineering tasks, which Anthropic reports as a state of the art result. The score was measured with a simple two tool setup, bash and file editing, averaged over 10 trials with a 200,000 token thinking budget and no test-time compute.
SWE-bench Verified is the coding test that matters most to working developers, because it uses real issues from real open source projects rather than textbook problems. A model has to read the issue, find the right files, write a fix and pass the project tests. Sonnet 4.5 at 77.2 percent sits ahead of GPT-5 at 74.9 percent on the same test, as reported by OpenAI, and ahead of Gemini 2.5 Pro at 59.6 percent, as reported by TechCrunch. Anthropic also reports 78.2 percent with a 1 million token context setup and 82.0 percent with extra parallel compute and candidate selection, but the 77.2 percent figure is the headline result under standard conditions.
The test conditions matter. Anthropic used a simple setup with only bash and file editing, so the score reflects the model rather than a complex harness built around it. The score fits what early users reported. Cursor chief executive Michael Truell told TechCrunch that Sonnet 4.5 shows state of the art coding performance, above all on longer tasks. Windsurf chief executive Jeff Wang called it a new generation of coding models. For a developer, the practical test is simple: give Sonnet 4.5 a real task in your own codebase, such as fixing a bug or adding a feature with tests, and compare the result with what Sonnet 4 produced. Our guide to OpenAI GPT-5 covers the rival model in the same detail.
Long tasks: agents that run for 30 hours
Anthropic says it watched Claude Sonnet 4.5 stay focused for more than 30 hours on complex multi-step tasks. That is the headline change for agent work: the model can keep going on a long job without drifting off task or losing track of the goal.
A 30 hour run is not a normal work session, and Anthropic is clear that the figure comes from its own tests with enterprise customers. Anthropic researcher David Hershey told TechCrunch that he watched the model code on its own for up to 30 hours in early trials. In that time it built an application, set up database services, bought domain names and ran a SOC 2 security audit to check the result. Those are the kinds of jobs that used to need a person checking in every few minutes.
Long runs work because the model holds context better and recovers from its own errors. Two new API features help: a context editing feature that manages the conversation as it grows, and a memory tool that lets an agent store facts across a long task. Anthropic added both so agents can run longer and handle greater complexity, which is what makes multi-hour work possible at all. Together they let an agent work through a large project in stages instead of starting over when the context fills up. The limit is still real. A long agent run needs a person to review the result, because a model that works for 30 hours can also repeat a wrong assumption for 30 hours.
Computer use and the OSWorld result
Claude Sonnet 4.5 scored 61.4 percent on OSWorld, a benchmark that tests whether an AI can carry out real tasks on a computer, such as using apps, files and a browser. Sonnet 4 scored 42.2 percent on the same test four months earlier, so the jump is large.
OSWorld matters because it tests the whole job, not one skill. The model has to look at a screen, decide what to click, type into the right field, open the right file and check the result. A model can be good at writing code and still fail here, because using a computer means dealing with menus, dialogs and apps that change state. Sonnet 4.5 leads the test at 61.4 percent, which means it still fails about four in ten tasks, so it is a working tool rather than a finished replacement for a person at a keyboard. The gain over Sonnet 4 is large: the score rose from 42.2 percent to 61.4 percent in four months, so the model now completes roughly three in five computer tasks instead of two in five.
Claude for Chrome puts this ability in the browser. The extension lets Claude work beside you in Chrome, clicking buttons, filling in forms and moving through sites. At launch it was available to Max users who had joined the waitlist. Anthropic also showed Claude navigating sites and filling spreadsheets in a demo. The risk with a browser agent is the same as with any agent that can act for you: it can click the wrong button or submit the wrong form, so watch the first runs closely and keep it away from payments until you trust it.
Claude Code 2.0: checkpoints, terminal and VS Code
Claude Code 2.0 arrived with Sonnet 4.5, and its most requested feature is checkpoints. A checkpoint saves your progress, so you can roll back instantly to an earlier state if the model takes the code in the wrong direction. The terminal interface was refreshed, and a native VS Code extension shipped.
Checkpoints change how it feels to work with an AI coder. Before them, a bad suggestion meant undoing edits by hand or restoring files from version control. Now you can let the model try an approach, see the result, and step back to the checkpoint if it fails. That makes it safer to give the model bigger jobs, because the cost of a wrong turn is a click instead of an afternoon of repair work. Anthropic says checkpoints were one of the most requested features from Claude Code users.
The native VS Code extension brings Claude Code into the editor most developers already use, instead of keeping it in a separate terminal window. The refreshed terminal interface keeps the command line workflow for people who prefer it. Both shipped in the same release, so developers can use either surface with the same model behind it. Anthropic also notes that Claude Code updates are available to all users, not only to paid plans. If you write code on a MacBook Pro M5, the extension runs in VS Code on macOS the same way it does on Windows and Linux, so the workflow is the same across machines.
The Claude Agent SDK
The Claude Agent SDK is the set of building blocks Anthropic uses to build Claude Code, now open to all developers. It is the same infrastructure that powers Anthropic’s own products, packaged so developers can build agents for their own work instead of only for coding.
Anthropic spent more than six months shipping Claude Code updates before it released the SDK, and the company says that work taught it what agents need in practice. The hard problems are memory across long tasks, permission systems that balance autonomy with user control, and coordination between subagents that share one goal. The SDK gives developers those pieces instead of asking each team to build them from scratch.
The SDK is not only for coding tools. Anthropic says it works for a wide range of agents, from research assistants that gather and check facts to workflow tools that move data between apps. It is available to all developers from launch day, and a developer can start with the SDK and the Sonnet 4.5 model, then add tools, memory and permissions as the agent grows. The practical advice is to start with one narrow job, such as triaging support tickets or drafting a weekly report, and add scope only after the agent handles that job well.
Reasoning, math and everyday knowledge
Claude Sonnet 4.5 shows substantial gains in reasoning and math over earlier Claude models, including Opus 4.1, according to Anthropic. Experts in finance, law, medicine and science, technology, engineering and math found its domain knowledge and reasoning dramatically better than older models.
The gains matter outside coding because the same reasoning applies to analysis, planning and writing. A model that reasons better can work through a contract clause, check a calculation step by step, or compare two options with their trade offs stated plainly. Anthropic says the improvement comes from the model itself, not from a new prompt trick, so it shows up across tasks rather than in one narrow test. The expert review covered finance, law, medicine and science, technology, engineering and math, and the experts found the gains over older models dramatic rather than small.
There is a limit to what the announcement proves. Anthropic did not publish a full table of reasoning scores in the launch post, and the expert findings are described in general terms rather than with numbers. The coding and computer use scores are the hard evidence, and they are strong. For other work, the honest test is your own: give Sonnet 4.5 a task you already did with Sonnet 4, such as summarizing a long document or checking a spreadsheet, and see whether the result needs less fixing.
Safety, alignment and ASL-3 protections
Anthropic calls Claude Sonnet 4.5 its most aligned frontier model yet, with large improvements in behavior over earlier Claude models. It shows less sycophancy, deception, power seeking and encouragement of delusional thinking, and it defends better against prompt injection attacks.
Alignment is about how the model behaves when no one is watching, and Anthropic treats it as a core part of this release rather than an add-on. Sycophancy means agreeing with the user even when the user is wrong, and deception means stating things the model has no basis to believe. Both are worse in an agent that works for hours without supervision, because a small error early on can grow into a large one. Better defenses against prompt injection matter for the same reason: an agent that reads web pages and files can be tricked by text hidden in those pages.
Sonnet 4.5 ships under AI Safety Level 3, or ASL-3, protections. That is the tier in Anthropic’s framework for models with stronger capabilities and stronger safeguards. The safeguards include classifiers that look for dangerous inputs and outputs, above all those linked to chemical, biological, radiological and nuclear weapons. Anthropic says it cut false positives from those classifiers by a factor of ten since it first described them, and by a factor of two since Claude Opus 4 shipped in May. The classifiers are not perfect, and Anthropic says they can still flag normal content by mistake, so it has made it easy to continue an interrupted conversation with Sonnet 4 instead.
Plans and API pricing
Claude Sonnet 4.5 costs the same as Claude Sonnet 4 through the API: 3 dollars per million input tokens and 15 dollars per million output tokens. It is available in the Claude apps and the API, and Anthropic recommends upgrading for all uses because it is a drop-in replacement.
The API price is the figure developers should plan around. One million input tokens is about 750,000 words, which is longer than the full Lord of the Rings series, so a typical coding session costs a small amount in tokens even when the context is large. Output tokens cost five times as much as input tokens, which is normal for this class of model, because generating text takes more work than reading it. Anthropic also offers prompt caching and batching, which cut the cost for repeated or non urgent work.
Consumer plans are listed on Claude’s pricing page. The free plan costs nothing and covers everyday chats with limits, and it is the place to start because it includes Sonnet 4.5. The paid plans raise the limits rather than changing the model itself. Pro costs 20 dollars a month when billed monthly, or 17 dollars a month on the annual plan, and includes Claude Code. Max costs 100 dollars a month with five times the usage of Pro, and Ultra costs 200 dollars a month with ten times the weekly usage of Pro and no five hour limit for verified accounts. Code execution and file creation in the Claude apps are available on all paid plans. Check the current pricing page before you subscribe, because plans and limits change.
Frequently asked questions
These are the questions readers ask most about Claude Sonnet 4.5. Each answer below comes from the Anthropic announcement, the TechCrunch report or Claude’s pricing page, and the benchmark numbers are labeled as Anthropic reported because they come from the company.
When did Anthropic launch Claude Sonnet 4.5?
Anthropic launched Claude Sonnet 4.5 on September 29, 2025. It was available the same day in the Claude apps and through the Claude API under the model id claude-sonnet-4-5, at the same price as Claude Sonnet 4.
What is Claude Sonnet 4.5 best at?
Coding, building agents and using computers, according to Anthropic. It scored 77.2 percent on SWE-bench Verified, 61.4 percent on OSWorld, and Anthropic says it can stay focused for more than 30 hours on complex multi-step tasks.
What is Claude Code 2.0?
Claude Code 2.0 is the updated version of Anthropic’s coding tool, released with Sonnet 4.5. It adds checkpoints that save your progress so you can roll back to an earlier state, a refreshed terminal interface and a native VS Code extension.
What is the Claude Agent SDK?
The Claude Agent SDK is the set of building blocks Anthropic uses to build Claude Code, now open to all developers. It gives developers the infrastructure for agents, including memory, permissions and subagent coordination, so they can build agents for work beyond coding.
How much does Claude Sonnet 4.5 cost?
Through the API it costs 3 dollars per million input tokens and 15 dollars per million output tokens, the same as Claude Sonnet 4. In the Claude apps it is included in the free plan with limits, and in paid plans: Pro at 20 dollars a month, Max at 100 dollars a month and Ultra at 200 dollars a month.
What is ASL-3?
ASL-3, or AI Safety Level 3, is the tier of Anthropic’s safety framework under which Sonnet 4.5 ships. It adds safeguards for a more capable model, including classifiers that look for inputs and outputs linked to chemical, biological, radiological and nuclear weapons.
Is Claude Sonnet 4.5 better than GPT-5?
On the coding test Anthropic highlights, yes: Sonnet 4.5 scored 77.2 percent on SWE-bench Verified, against 74.9 percent for GPT-5 as reported by OpenAI. On other tasks the answer depends on the work, so test both models on your own code before you choose. Our guide to OpenAI GPT-5 covers the GPT-5 launch in the same detail.
Sources
- Anthropic, Introducing Claude Sonnet 4.5: https://www.anthropic.com/news/claude-sonnet-4-5
- TechCrunch, Anthropic launches Claude Sonnet 4.5, its best AI model for coding: https://techcrunch.com/2025/09/29/anthropic-launches-claude-sonnet-4-5-its-best-ai-model-for-coding/
- MacRumors, Anthropic Debuts Claude Sonnet 4.5 With Improved Coding: https://www.macrumors.com/2025/09/29/claude-sonnet-4-5/
- Claude, Plans and pricing: https://claude.com/pricing
