Anthropic launches Claude Sonnet 4.5 for coding and agents

On September 29, 2025, Anthropic introduced Claude Sonnet 4.5 for coding, agents and computer use. The original source supports the documentary core of the event; performance and autonomy comparisons are vendor measurements under particular configurations.

A leap in coding that needs context

Sonnet 4.5 scores 77.2% on SWE-bench Verified, according to evaluations published by Anthropic. The benchmark uses 500 real, vetted issues from projects hosted on GitHub: the model must understand the repository, locate the problem and produce a code change that passes the relevant checks. Document supporting the figure.

It is one of the most widely used indicators for comparing coding models, although it is not equivalent to a developer’s day-to-day work. A correct benchmark result alone does not measure design quality, code security, subsequent maintenance or the ability to clarify ambiguous requirements with a client.

Anthropic also says it has observed Sonnet 4.5 maintaining focus for more than 30 hours on complex, multi-step tasks. The wording matters: it does not mean that any assignment can be left running for that long without supervision. Rather, the model sustained certain long-running processes within the company’s infrastructure and testing environment. Document supporting the figure.

The progress fits the broader shift in AI-assisted programming. Early copilots completed lines or generated isolated functions. Today’s tools can explore a repository, modify multiple files, run tests, consult documentation and fix their own errors. That is closer to an agent’s work than to autocomplete.

Claude gets better at using a computer

The other key figure comes from OSWorld, a benchmark that evaluates tasks performed on real computer interfaces. Sonnet 4.5 reaches 61.4%, compared with 42.2% for Sonnet 4 four months earlier. That is a 19.2-percentage-point improvement. Document supporting the figure.

This capability lets Claude navigate web pages, fill out spreadsheets and interact with applications through their interfaces. Anthropic is applying it in its Claude for Chrome extension, available today to Max subscribers who joined the waitlist.

Autonomous computer use greatly expands the range of possible applications, but it also increases the risk. An agent with browser access may encounter malicious instructions hidden in a website, email or document. This is known as prompt injection: external content designed to steer the model away from its original instruction and make it disclose information or take improper actions.

Anthropic says it has strengthened its defenses against these attacks. Even so, companies will need to limit permissions, log actions and require human confirmation before sensitive operations such as sending messages, deleting files, publishing changes or making payments.

Claude Code adds restore points

The launch comes with changes to Claude Code, Anthropic’s coding tool. The main new feature is checkpoints, which save progress and let users immediately return to an earlier state if the agent introduces a problematic change.

The update also brings a redesigned terminal interface and a native Visual Studio Code extension. In the API, Anthropic is adding context editing and a memory tool so agents can manage information more effectively during long-running tasks. In the Claude apps, the model can execute code and create spreadsheets, presentations and documents within the conversation.

Checkpoints address a practical problem with coding agents: the longer they work and the more files they modify, the higher the cost of a bad decision. The ability to roll back reduces that risk, although it does not replace version control, automated testing or human review.

The Claude Code engine opens up to other developers

Anthropic has also introduced the Claude Agent SDK, a package for building agents on the same technology used in Claude Code. It includes mechanisms for managing memory, permissions and subagents that collaborate on a shared task.

The bet goes beyond selling access to a model. Anthropic wants to provide the infrastructure needed to turn it into a system capable of taking action: deciding which tool to use, retaining relevant information and dividing work among multiple processes. The SDK is not limited to coding and could be applied, for example, to document analysis, research or internal business automation.

For developers, Sonnet 4.5 is available today through the API under the identifier claude-sonnet-4-5. Keeping Sonnet 4 pricing makes it easier to replace the previous model without raising the per-token cost, although the total expense will depend on how many steps, tool calls and retries each agent requires.

More capability, but also stronger safety measures

Sonnet 4.5 is being deployed under Anthropic’s AI Safety Level 3 protections. This internal level requires additional safeguards against capabilities that could increase risks related to chemical, biological, radiological or nuclear weapons.

Those measures include classifiers that examine potentially dangerous inputs and outputs. Anthropic acknowledges that these filters may block legitimate content and allows certain conversations to continue with Sonnet 4, which it considers lower-risk in this area. The company says it has reduced false positives tenfold since it first described the system and by half since the launch of Claude Opus 4 in May.

The company also reports improvements against behaviors such as excessive user flattery, deception, power-seeking and reinforcement of delusional ideas. These are results from the company’s own evaluations, documented in the model’s safety card; experience with real-world deployments will show how they translate to conversations and agents with tool access.

Sonnet 4.5 thus strengthens its position in generative AI’s most hotly contested commercial arena: systems that do more than respond, modifying code and operating applications. For businesses, the decision will no longer depend solely on which model scores highest on a benchmark, but on how much work it completes per euro, what errors it makes and what controls it provides when acting on real systems.

Turning the headline into a check

The first step is to freeze the system's identity. Anthropic introduced Claude Sonnet 4.5 for coding, agents and computer use. A commercial name may cover different revisions, automatic routes and tools. A test record should preserve date, access mode, configuration, permissions and the full output. Without that snapshot, an improvement or failure observed today cannot rigorously be attributed to the version another person will use tomorrow.

Next, turn whether code changes or computer actions should be delegated to it into cases with acceptance criteria. Build a local sample containing easy, ambiguous, long and deliberately impossible tasks. Record the input, the information the model may consult and what outcome would count as sufficient. A vendor-selected demonstration shows possibility; a test set preserved by the user measures reliability.

Autonomy needs a permission ladder. Reading and proposing are not the same as editing, sending or buying. A safer setup begins with read-only access, requires a preview and reserves execution for explicit approval. It also keeps a log and a rollback path. Judge the model by the errors the surrounding system contains, not by the confidence of its plan.

What the record must preserve

Cost and quality must be measured together. A cheap answer that must be reviewed from scratch can cost more than a slower but verifiable one. Measurement includes waiting, retries, consumption, human oversight and the consequences of failure. performance and autonomy comparisons are vendor measurements under particular configurations. That boundary turns the announcement into a testable hypothesis rather than a promise to be believed.

An evidence sheet separates four columns: what the source claims, what it shows, what it did not measure and what would change the conclusion. That discipline prevents an absence from becoming a promise and a condition from vanishing in summary. It also lets the story be updated without rewriting history from a later outcome.

Include a negative case before deciding. Find a situation where the system, rule, transaction or study does not meet the need and record the signal that would require stopping. Selected successes show that something can happen; the negative case reveals the boundary and lowers the cost of discovering it after deployment.