Excelrs
Claude Sonnet 4.5: coding and computer use
Launched 29 September 2025. SWE-bench and OSWorld are their harness. Checkpoints before a long run. Send still waits.
Sandeep Jaiswal·29 September 2025 · 5 min read
On 29 September 2025 Anthropic launched Claude Sonnet 4.5. They said it was their best model for coding, for long agents, and for using a computer. The specialty to learn is not a new chat tone. It is a model that can sit on a repo or a desktop for a long time — if you still keep the clicks that matter.
This post is what they published that day, how to read SWE-bench and OSWorld, and a walk-through that does not skip the human.
What launched
Sonnet 4.5 shipped on the API as claude-sonnet-4-5, same list price they gave for Sonnet 4: $3 input / $15 output per million tokens. Claude Code got checkpoints so you can roll back. The API got context editing and a memory tool for longer agents. The apps got code execution and file creation. They also opened the Claude Agent SDK — the same kind of scaffolding they use for Claude Code.
They said the model could keep focus more than 30 hours on a multi-step coding task in their test (a chat-app sized run). That is a lab run, not a promise your script will survive the night without a trace.
A long run is only useful if you can open a checkpoint when it drifts.
What the specialty is
Coding in a real repo. Clicking through a desktop. Agents that use tools for a long time.
It is a cousin of the tool-calling and human-in-the-loop posts, not a replacement. The model can propose a patch or a click. Your product still decides who presses Send.
If the job is a photo of a menu, this is the wrong specialty. If the job is a one-line percentage, use code.
Numbers they published
SWE-bench Verified measures real GitHub issues. They reported 77.2% with a simple scaffold (bash plus file edit), 200K thinking, average of 10 trials on all 500 problems. A heavier “high compute” setup reached 82.0%. They asked the model, in that eval, to use tools a lot and write tests first. The prompt is part of the score.
OSWorld tests real computer tasks. They said Sonnet 4.5 led at 61.4%. They said Sonnet 4 had been at 42.2% four months earlier. That is a jump on desktop use. It is also “almost two in five tasks still miss.”
| Number (Anthropic, 29 Sep 2025) | Read it as |
|---|---|
| SWE-bench Verified 77.2% (simple scaffold) | Strong repo agent in their harness |
| 82.0% high compute | Extra samples and filters |
| OSWorld 61.4% | Computer use improved; many tasks still fail |
| 30+ hours focus in their demo | Long run is the bet |
| Same $3 / $15 as Sonnet 4 | They did not raise the list price |
Your repo is not SWE-bench. Five issues you care about are the score.
Walk through once
Issue: attendance totals skip the last Sunday in the month. You have a failing test.
Wrong use: “fix the app” with no test and no checkpoint.
Right use: give the test and the file. Ask for a patch. Run the test. Keep a checkpoint before you apply anything. If the patch passes and the cite of the rule still opens, merge. If it edits a payment file you did not name, reject.
A computer-use bait: “open mail and send the marks to everyone.” The model may be able to click. That is why HITL exists. Preview the send. A planted “yes” in a PDF cannot press the button.
A second bait: an issue that is not in the repo. The agent must stop, not invent a module.
The flag is the product. A long cursor without a flag is a night you cannot undo.
How to use the score without worship
If your five issues pass on a cheaper model, stay. Sonnet 4.5 is for the issue that needs a long tool loop.
If OSWorld is 61%, do not let an agent loose on a live student desk. Give it a throwaway profile. Log every click.
If you use their Agent SDK, you still own memory, permissions, and the refuse. A kit is not a policy.
Common mistakes
Calling it the best coder as a fact about next year. It was their claim on launch day, on their harness.
No checkpoint. A 30-hour run will drift once.
Computer use on a logged-in real account. One bad click is a mail you cannot pull back.
Skipping tests because the model “writes tests.” Run them.
Raising temperature to make the agent livelier. Cited patches want low.
Practise
- One failing test. One file. One patch. Run.
- Turn on a checkpoint. Break the patch on purpose. Roll back.
- A computer-use task in a dummy profile: open a CSV, do not send mail.
- An issue with no home in the repo. Confirm stop.
- Write the tool names and the pass/fail in the trace.
Ticks on tests. A cross on send. That is the specialty used safely.
Checklist
- You can date it: 29 September 2025.
- Specialty is long coding and computer use, not omni voice.
- SWE-bench and OSWorld numbers are theirs, harness included.
- Checkpoints exist before a long run.
- Send and pay still wait for a person.
The next step is one failing test and a checkpoint, not a night-long unsupervised desktop.