OpenAI released GPT-5.6 yesterday, and it’s shipping exclusively in Kiro, the company’s AI coding assistant. The pitch is straightforward: better price-performance than the full o3 model for developers who want help planning, building, reviewing, and testing software.

The timing is notable. OpenAI has been talking about AI agents for months, pushing the idea that these systems should handle everything from customer support to software engineering. But Kiro is the first product where they’re actually shipping a model optimized for a specific workflow at a price point that isn’t o3.

No pricing details yet, and OpenAI hasn’t published benchmarks comparing 5.6 to o3 or GPT-4.5. The announcement focuses on “price-performance,” which usually means “cheaper and almost as good.” If you’re already using o3 for code and it’s working, there’s no obvious reason to switch. If you’ve been priced out of o3, this might be worth checking out.

The bigger question is whether anyone outside OpenAI’s existing customer base will actually use Kiro. Cursor and GitHub Copilot already own the AI coding assistant market, and both integrate multiple models. Developers tend to stick with tools that work, and “OpenAI’s coding assistant” isn’t an automatic win anymore.

llm-anthropic Gets a Compatibility Update

Simon Willison released llm-anthropic 0.27, updating the plugin to work with Anthropic’s new Python SDK v1.0.0. The main change: Anthropic switched from httpx to httpx2, the same move OpenAI made two weeks ago in their v3.0.0 release.

If you’re using llm-anthropic in production, you’ll need to upgrade. Anthropic published a migration guide that covers the breaking changes. Willison used Claude Code to handle the upgrade, which is the kind of recursive AI use case that either feels incredibly futuristic or deeply silly depending on your mood.

This matters if you’re building tools on top of Anthropic’s API. If you’re just using Claude through the web interface, you can ignore it.

GitHub’s Alt Text Checker That Actually Checks

GitHub shipped a new plugin for the GitHub Accessibility Scanner that evaluates whether your alt text is actually useful. The problem it’s solving is real: most automated accessibility checks just verify that alt text exists, not that it describes the image in a way that helps screen reader users.

The plugin uses an AI model to analyze images and compare them against the alt text you wrote. If your alt text is “image.png” or “screenshot,” it’ll flag it. If you wrote “dashboard showing user metrics” for a complex chart, it’ll check whether that description actually covers what’s in the image.

This is a good use of AI. The tool isn’t writing alt text for you, it’s checking the alt text you wrote against what’s actually in the image. It catches lazy descriptions and prompts you to be more specific. If you maintain documentation or a public-facing site, it’s worth adding to your workflow.

The plugin is open source and works with the existing GitHub Accessibility Scanner, which means it fits into CI/CD pipelines without reinventing your whole testing setup.

Everything Else

Headlong is a microharness for persistent agents that keeps them running across sessions. If you’re building agents that need to maintain state or run tasks over days instead of minutes, it’s worth reading through. It’s not a framework, just a lightweight way to handle persistence.

Bookshelf is a self-hosted ebook library that runs on object storage. If you’ve been looking for a Calibre alternative that doesn’t require a traditional database, this might work. It’s early, but the architecture is clean.

Instinct, the AI assistant that’s been getting attention for how much it can do, is also raising privacy concerns. The trade-off is predictable: the more access you give an agent, the more it can do, and the more uncomfortable the security implications become. Early testers say it’s powerful. They also say the terms are broad and the permissions are sweeping.

And OpenAI got caught disrupting a Russian influence campaign that used AI to promote a fake Israel-based think tank. The accounts were pushing a “sovereignty index” that praised Russia and criticized the West. OpenAI banned them. The fact that this is now a regular disclosure category tells you everything about where we are.