Two days, two launches, and two very different ideas of where AI is heading.

On Monday, Anthropic released a model built to sit in your terminal and write code for hours. On Tuesday, OpenAI released a model that makes videos, plus an app to share them. If you only skimmed the headlines, it looked like "AI news, again". Look a bit closer and the week says a lot about what these companies think their tools are for.

Here's what shipped, and what I think is worth your attention as a developer.

What was released

• 29 September: Claude Sonnet 4.5. Anthropic's new Sonnet model, pitched squarely at coding and at agents that keep working on long, multi-step tasks without losing the thread.
• Same day: Claude Code 2.0. A batch of updates to Claude Code, including a native VS Code extension and checkpoints, which let you roll back to an earlier state when a session goes somewhere you didn't want.
• Same day: the Claude Agent SDK. The machinery behind Claude Code, opened up so developers can build their own agents on top of it.
• 30 September: Sora 2. OpenAI's next video generation model, released together with a new Sora app for creating and sharing videos.

Sonnet 4.5: the boring word that matters is "long"

Every model launch says "better at coding". That alone isn't news anymore. The more interesting part of this release is the emphasis on long-running agents.

If you've used any coding agent seriously, you know where they usually fall over. Not on the first file. On the twelfth step, when the agent has forgotten a decision it made on step three, repeats a fix it already tried, or wanders off into an unrelated refactor. A model that stays coherent over a long task is worth more in practice than one that writes a slightly nicer single function.

My plan is the same as with every new model: don't trust the launch post, trust your own repo. Give it the kind of task that tripped up the previous model for you, ideally one with tests, and see whether it gets further before you have to step in.

Claude Code 2.0: checkpoints are the quiet headline

I use Claude Code daily, so this is the part of the week that changes my workflow most directly.

The VS Code extension is the visible change. If you prefer seeing diffs in an editor panel instead of a terminal, the agent now lives where you already work. Plenty of developers bounced off terminal-only agents for exactly that reason.

But checkpoints are the feature I'd tell a colleague about first. Agents are most useful when you let them try things, and you only feel comfortable letting them try things when undoing is cheap. Being able to rewind a session to a known good state lowers the cost of a bad attempt.

The Agent SDK: Claude Code as a building block

This one is easy to overlook, and I think it matters more over time. A coding agent is a lot more than a model call. It's a loop: read files, run commands, look at the output, decide what to do next, manage context, ask for permission. Getting that loop right is most of the work.

Opening it up as an SDK means you can build agents for things that aren't "write code in my repo": a support triage agent, an internal tool that investigates failing jobs, a research helper for your own documentation. The same patterns that make a coding agent useful (tools, permissions, context management) apply to all of them.

The usual caution applies: an agent you build is an agent you're responsible for. Scope its tools tightly and log what it does.

Sora 2: a different kind of bet

Sora 2 is a video model, so most developers won't touch it in their day job. What's interesting is the packaging. OpenAI shipped it with a Sora app, which makes it feel less like a research demo and more like a consumer product people create and share with.

For developers, the practical points are:

• Generated video is getting harder to tell apart from real footage. If you build anything that accepts user-uploaded media (marketplaces, verification, reviews), "is this real?" becomes a product question, not just a philosophical one.
• Product and marketing teams will ask about it. Expect "can we make our explainer videos with this?" in your next planning meeting. The honest answer involves rights, brand review and a human in the loop.

Two directions, one lesson

Put the two launches side by side and you get a neat summary of 2025 in AI. One direction is AI as a worker: agents that do long, structured tasks inside your tools. The other is AI as a creator: models that produce media for people to watch and share.

As developers, we mostly live in the first world. And the lesson there hasn't changed with this release: the better agents get, the more your engineering habits matter. Clear tasks, tests, small branches and real code review are what turn a capable model into a safe one.

What I'd do this week

• Try Sonnet 4.5 on one real task from your backlog that has tests, not a toy demo.
• If you use Claude Code, update it and get used to checkpoints before you need them in a hurry.
• If you've been meaning to build an internal agent, read the Agent SDK docs before writing your own loop from scratch.
• Keep committing before big agent runs. No release changes that.

Which of these launches will actually change your week: the coding model, the editor extension, the SDK, or none of them?