Laravel AI SDK v1.0 Released: Tool Approvals, Classification Models and Upgrade Notes

On September 23, 2026, Laravel released version 1.0 of the Laravel AI SDK, its first-party package for adding AI features to PHP applications. The SDK had been in beta since earlier in the year. Version 1.0 is the point at which Laravel describes it as ready for production.

For a business running or planning a Laravel application, the release matters for two reasons. It makes AI features such as assistants, document handling and automated decisions a supported part of the framework's own toolset, and it adds controls that companies usually ask for before letting software act on their behalf. Teams already using the beta also have upgrade work to plan. If you work with a Laravel development team, this is a release worth raising with them.

What Laravel released on September 23, 2026

The Laravel announcement describes the AI SDK as a single Laravel-native interface that works across AI providers, so that application code is not tied to one vendor's API. According to the post, the team closed hundreds of bugs between the beta and v1.0 and added several new capabilities. The package is installed with Composer and is documented in the Laravel AI SDK documentation.

Version 1.0 is released and available now. Some individual features, covered below, currently work only with specific AI providers, and Laravel says so feature by feature.

Human approval before an agent acts

The feature most relevant to business risk is what Laravel calls approvable tool calls. In AI terms, a tool is an action the model can request, such as issuing a refund, sending an email or updating a customer record. The announcement explains that a developer can now mark a tool as requiring approval. When the AI agent reaches that tool, it pauses instead of running it, and the application can show a person exactly what the model wants to do and with which values.

The reviewer then has three options, according to the post: approve the action, reject it with a reason that is passed back to the model, or edit the values before the action runs. Laravel states that approvals work whether the request is handled immediately, streamed, queued in the background or broadcast.

This is the difference between an assistant that drafts and an agent that executes. If that distinction is new to you, our guide to AI agents, chatbots and workflow automation explains it in more detail.

Classification models for fast yes-or-no decisions

The second notable addition is a new category of AI work: classification. The announcement points out that much business automation is not about writing text at all. Routing a support ticket, flagging a comment for review or choosing the next step in a workflow are decisions between a small number of options.

Laravel says the SDK now treats classification as its own capability alongside text, images, audio and embeddings, and names Jev, a model from a company called TypeSafe, as the example. The post claims that Jev can answer classification questions in milliseconds at a fraction of the price of traditional large language models. That is the vendor's description, and the announcement gives no benchmark or price figures, so it should be tested against your own workload before you rely on it.

There is also a simple helper that asks one yes-or-no question and returns true or false, with a threshold setting that controls how certain the model must be. Per the announcement, classification currently runs through TypeSafe and OpenRouter, with other providers to be added behind the same interface as they release comparable models. That last part is a stated intention with no date attached.

Controlling cost and connecting chat interfaces

Several changes are aimed at keeping AI features affordable and easier to build.

Loading tools only when needed

The post explains that an agent with 30 tools normally describes all 30 on every request, which adds cost and makes the model's choice less accurate. A new tool search feature lets rarely used tools be loaded only when a request calls for them. Laravel states this is supported on OpenAI and Anthropic.

Middleware at every step

Developer-written checks, known as middleware, previously ran once per prompt. They now run around each step an agent takes. The announcement gives examples such as removing an expensive tool once it has been used, switching to a different model partway through, or returning a cached answer without calling the provider.

Code execution, chat protocols and usage reporting

A code execution tool lets the model run calculations in the AI provider's own sandbox, which the post says is supported on Anthropic, OpenAI, Azure, Gemini and xAI. The SDK also reads and writes the Vercel chat protocol and the AG-UI protocol, so common front-end chat components can connect to a Laravel back end. Finally, token usage is now reported consistently across providers, which makes it simpler to track what each AI feature costs.

Breaking changes for teams already on the beta

Laravel is direct about the fact that v1.0 includes breaking changes, mainly in conversation storage, agent middleware, token usage and stream protocols. The one that needs planning is storage. The announcement explains that the way stored conversations record tool calls has been restructured into a single per-step record, because the earlier layout could not reliably reconstruct a conversation for some providers and could not distinguish an action that never ran from one awaiting approval.

Two consequences follow, both stated in the post. Any custom SQL that reads the old tool call columns directly has to be rewritten, and the upgrade guide includes a data migration that must be run once before v1.0 is deployed. Laravel also offers a guided upgrade through Laravel Boost, its tool for AI coding assistants, which walks an assistant through the guide one change at a time. A developer should still review and test the result.

What this means for your business

The following is our interpretation rather than anything Laravel has stated. A 1.0 release from the framework's own team lowers one common objection to building AI features into a Laravel product: the worry that the tooling will change underneath you. A provider-neutral layer also reduces the cost of switching AI vendors later, which matters given how often model pricing and capabilities change.

The approval feature deserves particular attention from operations and compliance leads. Many companies are comfortable with AI that suggests and uncomfortable with AI that acts. A built-in way to pause, show the proposed action and record a human decision makes it more realistic to automate tasks that touch money, customer data or outbound communication. It does not remove the need to decide which actions require sign-off, who gives it, and how those decisions are logged.

The SDK is a building block, not a finished product. Data protection, accuracy testing, access permissions and monitoring remain the responsibility of whoever builds and operates the application.

What to do next

• If you already run the beta SDK in production, ask your developers to read the upgrade guide, estimate the effort, and schedule the data migration with a backup and a staging test before deploying v1.0.
• If you are planning an AI feature, list the actions the AI would take and mark which ones need human approval. That list shapes the design more than the choice of model.
• For sorting and routing tasks, ask whether a classification model fits better than a general-purpose one, and measure accuracy on your own data.
• Check that the features you need are supported by the AI provider you intend to use, since tool search, code execution and classification each have their own provider list.

Conclusion

Laravel AI SDK v1.0, released on September 23, 2026, gives Laravel applications a supported, provider-neutral way to build AI features, with human approval for sensitive actions, a new classification capability and better cost controls. Teams on the beta need to plan for breaking changes, including a one-time data migration for stored conversations.

If you are weighing up an AI feature for an existing Laravel application or a new product, you can request a quote from Entrant Technologies and we will help you scope what is practical.