Anthropic drops Claude Sonnet 4.5 to supercharge coding and agents

AI coding assistants are great at fixing single functions, but they usually fall apart the moment you ask them to handle a multi-hour project. Context windows clutter up, logic drifts, and you end up spending more time babysitting the agent than writing code yourself.

Anthropic wants to end that babysitting phase. On Monday, the company officially launched Claude Sonnet 4.5, declaring it the best coding and computer-use model in the world. Pricing stays completely untouched at $3 per million input tokens and $15 per million output tokens.

Thirty hours without losing the plot

The most eye-catching claim here is Sonnet 4.5's sheer stamina. Anthropic says the model can maintain autonomous focus for over 30 hours on complex, multi-step engineering tasks. In one showcase, it spent 30 hours building a functional Slack-style chat application from scratch, churning out roughly 11,000 lines of code before stopping.

Anthropic researchers also noted enterprise test runs where the model didn't just write scripts, but provisioned database instances, bought domain names, and ran a SOC 2 security audit. On X, developers quickly pointed out that previous benchmarks like Opus 4 and GPT-5-Codex tapped out around the 7-hour mark.

Benchmark scores back up the hype. Sonnet 4.5 scored 77.2% on SWE-bench Verified out of the box, jumping to 82.0% with parallel test-time compute. On OSWorld, which evaluates how well models can drive a real desktop operating system, Sonnet 4.5 jumped to 61.4%, up from Sonnet 4's 42.2% just four months ago.

The tools behind the model

Anthropic didn't just ship raw weights. It also overhauled its developer tooling to make sure people actually trust this thing with real codebases.

Claude Code gained an instant checkpointing system. If an agent goes down an unhelpful rabbit hole, tapping Escape twice or typing a quick command rewinds both code edits and conversation states. Alongside a new VS Code extension, Anthropic released the Claude Agent SDK, exposing the exact same orchestration, memory, and permission frameworks it uses to build Claude Code internally.

On the safety side, Sonnet 4.5 ships under ASL-3 safeguards, using mechanistic interpretability to curb deceptive behavior, sycophancy, and prompt injections. If the model's CBRN classifier catches false positives, users can instantly continue the interrupted thread on Sonnet 4.

Why it matters

Anthropic was facing real heat from OpenAI's GPT-5 on technical leaderboards. With Sonnet 4.5, it reclaims clear bragging rights across software engineering and agentic workflows without hiking API costs.

For builders and dev shops, this shifts AI from an autocomplete suggestion into an asynchronous contractor. You can genuinely delegate an entire test suite overhaul or repository refactor before heading to bed, knowing the checkpoint system provides an instant parachute if things go sideways.

We think packaging the internal Agent SDK is the shrewdest move of this release. Writing reliable subagent orchestration and state management from scratch has been a massive headache for AI startups; by handing over the blueprints, Anthropic effectively cements Claude as the default engine for modern agent builders.

Still, giving an autonomous agent a 30-hour leash comes with genuine operational risks. An 11,000-line PR that technically runs can still introduce insidious architectural debt. Teams diving into Sonnet 4.5 should start with isolated microservices and internal tooling before letting it manage production infrastructure.