Changelog

This page summarizes recent versions. The full notes are in CHANGELOG.md. The package follows Semantic Versioning.

1.3.0 (2026-11-05)

The package now covers Cloudflare’s other model types, through the Laravel AI code you already write.

• Reranking. Reranking::of() and Collection::rerank(). See Reranking.
• Image generation. Image::of(), with sizes, quality and FLUX.2 reference images. See Image generation.
• Speech and transcription. Audio::of() and Transcription::of(), with speaker labels on nova-3. See Speech and transcription.
• Documents. Markdown::of() converts PDF, Word, Excel and other files. convert_documents does it for attachments on chat calls. See Documents.
• Web search. CloudflareWebSearch is a normal tool. See Web search.
• Neurons on every model call. WorkersAiUsageReported now fires for Clef, Clairic and the new calls.
• Tested on an authenticated gateway. Text and embeddings need gateway_token. Other calls need only key, except FLUX.2. nova-3 cannot run through a gateway, and throws. See Non-JSON models.
• Docs corrected. Laravel AI provider tools do not throw on laravel/ai 1.1 and later. They are removed with no error. A document or audio attachment has a clearer error.

1.2.0 (2026-11-05)

Upgrade from 1.1.0. Clef and Clairic fail on every live call in that version. This release fixes them, and they now behave like the rest of Laravel AI. The docs have a separate landing page, a new overview and a guide to the routes.

• Clairic has Laravel AI’s testing and failover shape. Adds Extraction::fake(), assertExtracted(), assertNotExtracted(), assertNothingExtracted(), the Extracting, Extracted and ExtractionFailed events, failover lists, when(), withHeaders(), and closures in withProviderOptions(). See Clairic.
• Clef and Clairic read Cloudflare’s result envelope, and accept errors: [] as success. In 1.1.0, Clef returned no answers and Clairic threw Workers AI Error: [unknown] Unknown error.
• Clef throws on a bad reply. A 200 reply with an error or no answers throws AiException, and a 408 throws GatewayTimeoutException. Before, Clef returned an empty response.
• Third-party models are accepted on the direct route. Cloudflare’s REST API serves openai/gpt-5.5 on /ai/v1. See third-party models.
• Clairic answers gain guess(), spans(), steps(), pBelow(), pAbove(), and the response gains stateText.
• Clef and Clairic validate config as chat does, so url with account_id throws.
• The classifying provider is a normal class, not eval().
• Corrects the third-party docs. A Cloudflare token in key returns 401 on /compat, so use the direct route.
• The docs now say where each value goes: .env holds the values, and config/ai.php reads them.

1.1.0 (2026-11-05)

Known issue: Clef and Clairic fail on every live call. Use 1.2.0.

• Clef decisions through Classification::of(), on laravel/ai 1.0 and later. See Clef and Clairic.
• Clairic extraction through Extraction::of(). Not a laravel/ai API. Works on 0.9 to 1.3.
• A shorter README. The full reference is on the docs site.

1.0.1 (2026-11-02)

• Streamed calls now dispatch WorkersAiUsageReported with the call’s neuron total. Before, only non-streamed calls reported neurons. See Cost tracking.

1.0.0 (2026-11-02)

First stable release. One codebase supports laravel/ai 0.9 to 1.3.

• Adds laravel/ai 1.x support (^1.0, tested on v1.0.0 and v1.3.0).
• Sends the headers config key and, on 1.x, Provider::withHeaders() headers. Before 1.0.0 they were dropped.
• Fills StepResponse::$reasoning on 1.x.
• On 0.9 to 0.11, Usage::$promptTokens excludes cached tokens. On 1.x, TextUsage::$inputTokens includes them. WorkersAiUsageReported::$promptTokens is uncached on both.
• Fixes 6 attempts on a persistent 429 with laravel/ai 1.2.1 and later. The total is 3 again.
• Reads reasoning_tokens from completion_tokens_details too.

Upgrading from 0.8.x? See Upgrade to 1.0.