Anthropic

Configuration

'anthropic' => [
    'api_key' => env('ANTHROPIC_API_KEY', ''),
    'version' => env('ANTHROPIC_API_VERSION', '2023-06-01'),
    'default_thinking_budget' => env('ANTHROPIC_DEFAULT_THINKING_BUDGET', 1024),
    // Include beta strings as a comma separated list (e.g. output-128k-2025-02-19, code-execution-2025-05-22).
    'anthropic_beta' => env('ANTHROPIC_BETA', null),
]

You can also override anthropic_beta per request with ->using('anthropic', $model, ['anthropic_beta' => '...']). On top of the configured list, Prism adds the betas a request's provider options need (fallbacks, thinking.display: updates, thinking.block_binding), de-duplicated, on every request and step.

Prompt caching

Anthropic's prompt caching feature allows you to drastically reduce latency and your API bill when repeatedly re-using blocks of content within five minutes or one hour of each other, depending on the Anthropic compatible TTL option you provide.

There are two ways to enable prompt caching:

• Automatic caching
• Explicit cache breakpoints

To enable automatic caching, simply add a single cache_control field at the top level of your request:

use Prism\Prism\Enums\Provider;
use Prism\Prism\Facades\Prism;
use Prism\Prism\Tool;
use Prism\Prism\ValueObjects\Messages\UserMessage;
use Prism\Prism\ValueObjects\Messages\SystemMessage;

Prism::text()
    ->using(Provider::Anthropic, 'claude-3-5-sonnet-20241022')
    ->withSystemPrompt(
        (new SystemMessage('I am a long re-usable system message.'))
    )
    ->withMessages([
        (new UserMessage('I am a long re-usable user message.'))
    ])
    ->withTools([
        Tool::as('cache me')
    ])
    ->withProviderOptions(['cache_control' => ['type' => 'ephemeral']])
    ->asText();

We support Anthropic explicit cache breakpoints on:

• System Messages (text only)
• User Messages (Text, Image and PDF (pdf only))
• Assistant Messages (text only)
• Tools

The API for enabling prompt caching is the same for all, enabled via the withProviderOptions() method. Where a UserMessage contains both text and an image or document, both will be cached.

use Prism\Prism\Enums\Provider;
use Prism\Prism\Facades\Prism;
use Prism\Prism\Tool;
use Prism\Prism\ValueObjects\Messages\UserMessage;
use Prism\Prism\ValueObjects\Messages\SystemMessage;

Prism::text()
    ->using(Provider::Anthropic, 'claude-3-5-sonnet-20241022')
    ->withSystemPrompt(
        (new SystemMessage('I am a long re-usable system message.'))
            ->withProviderOptions(['cacheType' => 'ephemeral', 'cacheTtl' => '1h'])
    )
    ->withMessages([
        (new UserMessage('I am a long re-usable user message.'))
            ->withProviderOptions(['cacheType' => 'ephemeral'])
    ])
    ->withTools([
        Tool::as('cache me')
            ->withProviderOptions(['cacheType' => 'ephemeral'])
    ])
    ->asText();

If you prefer, you can use the AnthropicCacheType Enum like so:

use Prism\Prism\Enums\Provider;
use Prism\Prism\Providers\Anthropic\Enums\AnthropicCacheType;
use Prism\Prism\ValueObjects\Messages\UserMessage;
use Prism\Prism\ValueObjects\Media\Document;

(new UserMessage('I am a long re-usable user message.'))->withProviderOptions(['cacheType' => AnthropicCacheType::ephemeral])

Important: To enable prompt caching:

• System messages must use withSystemPrompt() or withSystemPrompts() (Anthropic does not allow SystemMessages in the messages array)
• User and Assistant messages must use withMessages()
• Tools use withTools()
• All message types support caching via withProviderOptions(['cacheType' => 'ephemeral'])
• You cannot use withPrompt() for caching as it doesn't allow adding provider options to individual messages
• Anthropic supports two TTL options: 5m (default) or 1h. You can leave the cacheTtl unset and Anthropic will use the default TTL of 5m.

Tool result caching

In addition to caching prompts and tool definitions, Prism supports caching tool results. This is particularly useful when making multiple tool calls where results might be referenced repeatedly.

To enable tool result caching, use the tool_result_cache_type provider option on your request:

use Prism\Prism\Facades\Prism;

$response = Prism::text()
    ->using('anthropic', 'claude-3-5-sonnet-20241022')
    ->withMaxSteps(30)
    ->withTools([new WeatherTool()])
    ->withProviderOptions([
        'tool_result_cache_type' => 'ephemeral'
    ])
    ->withPrompt('Check the weather in New York, London, Tokyo, Paris, and Sydney')
    ->asText();

When multiple tool results are returned, Prism automatically applies caching to only the last result, which caches all preceding results as well. The breakpoint rolls forward to the newest tool result on each step of a tool loop (text, streaming and structured alike), so a long loop never exceeds Anthropic's 4-cache-breakpoint limitation.

Please ensure you read Anthropic's prompt caching documentation, which covers some important information on e.g. minimum cacheable tokens and message order consistency.

Thinking

Claude can reason before it answers. On current models that reasoning is adaptive: Claude decides when and how deeply to think, and you steer it with effort. Claude Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 think by default, even when you send no thinking option at all. Older models (Opus 4.5, Sonnet 4.5, Sonnet 3.7 and earlier) use manual thinking with a fixed token budget.

Please refer to Anthropic's thinking documentation for limitations and behaviours when using thinking with caching and/or tools.

Adaptive thinking (recommended)

use Prism\Prism\Facades\Prism;

Prism::text()
    ->using('anthropic', 'claude-opus-5-5')
    ->withPrompt('What is the meaning of life, the universe and everything in popular fiction?')
    ->withProviderOptions(['thinking' => ['type' => 'adaptive']])
    ->asText();

On the 5.x models thinking is on without this option, and the thinking text is hidden by default (display: omitted): you get thinking blocks with an empty thinking field and a signature. To read a summary of Claude's reasoning, ask for it with display:

Prism::text()
    ->using('anthropic', 'claude-opus-5-5')
    ->withPrompt('Analyze the trade-offs between microservices and monolithic architectures')
    ->withProviderOptions([
        'thinking' => ['type' => 'adaptive', 'display' => 'summarized'],
        'effort' => 'high',
    ])
    ->asText();

Use the effort parameter to guide how much work Claude puts into the whole response, thinking included:

Level	Description
max	Maximum capability with no constraints on thinking depth.
xhigh	Deeper than high for the hardest tasks.
high	Deep reasoning on complex tasks. The default on most models; Opus 5.5 and Haiku 5.5 default to medium.
medium	Balanced approach with moderate token savings.
low	Most efficient. Significant token savings with some capability reduction.

Which levels a model accepts varies; an unsupported level comes back as a PrismException from the API.

TIP

The effort parameter can also be used without a thinking option, in which case it controls overall token spend for text responses and tool calls.

Works identically with Prism::structured() and streaming.

The thinking option

When you set thinking.type, Prism sends your thinking array to Anthropic as given, using the API's field names. That means anything the API accepts works without a Prism release:

Option	Example	Notes
type	adaptive, enabled, disabled, between_tools	Required for passthrough.
display	summarized, omitted, updates	updates is in beta; Prism adds the thinking-display-updates-2026-08-18 header for you.
block_binding	['prefix_mismatch_behavior' => 'error']	Prefix enforcement for preserved thinking (error or drop_block); Prism adds the thinking-binding-controls-2026-08-01 header for you.
budget_tokens	2048	Only meaningful for type: enabled. Prism's budgetTokens alias is translated for enabled, and the config default is used when neither is set.

Prism doesn't validate these against the model. The API's answer is the source of truth, and a rejected combination surfaces as a PrismException. As of this writing:

• enabled (with budget_tokens) returns a 400 on all of the 5.x models above.
• disabled returns a 400 on Opus 5.5, Sonnet 5.5 and Fable 5.1. Haiku 5.5 accepts it at high effort or below.
• Sonnet 5.5's lowest setting is between_tools, which turns off up-front thinking but still returns the short progress updates it writes between tool calls as thinking blocks. It's accepted at high effort or below and takes no other field. Opus 5.5 rejects it.

// Turn up-front thinking off on Sonnet 5.5
Prism::text()
    ->using('anthropic', 'claude-sonnet-5-5')
    ->withPrompt('Summarize this article in one sentence.')
    ->withProviderOptions(['thinking' => ['type' => 'between_tools']])
    ->asText();

Manual thinking (legacy)

WARNING

Manual thinking with enabled and a token budget is deprecated on Claude 4.6 and rejected by the 5.x models. Use adaptive thinking instead. Manual thinking is still required for older models (Sonnet 4.5, Opus 4.5, Sonnet 3.7, etc.).

Prism::text()
    ->using('anthropic', 'claude-3-7-sonnet-latest')
    ->withPrompt('What is the meaning of life, the universe and everything in popular fiction?')
    ->withProviderOptions(['thinking' => ['enabled' => true]])
    ->asText();

By default Prism will set the thinking budget to the value set in config, or where that isn't set, the minimum allowed (1024).

You can override the config (or its default) using withProviderOptions:

Prism::text()
    ->using('anthropic', 'claude-3-7-sonnet-latest')
    ->withPrompt('What is the meaning of life, the universe and everything in popular fiction?')
    ->withProviderOptions([
        'thinking' => [
            'enabled' => true,
            'budgetTokens' => 2048,
        ],
    ]);

Note that thinking tokens count towards output tokens, so you will be billed for them and your token budget must be less than the max tokens you have set for the request. If you expect a long response, ensure there's enough tokens left for the response — i.e. does (maxTokens - thinkingBudget) leave a sufficient remainder.

Inspecting the thinking block

Prism reads thinking from the response whenever the model returns it, whether or not you sent a thinking option. You can access it via the additionalContent property on either the Response or the relevant step:

Key	Contents
thinking_blocks	Every thinking and redacted_thinking block, in order, exactly as returned.
thinking	The text of the first thinking block (an empty string when display is omitted).
thinking_signature	The signature of that same block.

$response = Prism::text()
    ->using('anthropic', 'claude-opus-5-5')
    ->withPrompt('What is the meaning of life, the universe and everything in popular fiction?')
    ->withProviderOptions(['thinking' => ['type' => 'adaptive', 'display' => 'summarized']])
    ->asText();

$response->additionalContent['thinking'];

When using tools, Anthropic returns thinking on the tool call steps, so look at the steps:

$response->steps->first()->additionalContent['thinking_blocks'];

NOTE

Claude may skip thinking for simple queries, in which case no thinking keys are present.

Thinking tokens

When Anthropic reports usage.output_tokens_details.thinking_tokens, Prism puts it on $response->usage->thoughtTokens, for text, structured and streamed responses, per step and summed across a tool loop. completionTokens is still Anthropic's output_tokens, which already includes those thinking tokens, so don't add the two together. Older models don't send the field, so thoughtTokens stays null (a streamed tool loop's total reports 0).

$response->usage->completionTokens; // everything Claude generated, thinking included
$response->usage->thoughtTokens;    // the thinking part of it

Replaying conversations with thinking

Thinking blocks have to go back to Anthropic complete and unmodified, in their original position, when you continue a tool loop. On the 5.x models the API can also check that everything before a thinking block (system prompt, tools and earlier messages) is unchanged. That check is on by default for accounts created on or after 2026-08-31, and you can turn it on with thinking.block_binding. Prism handles this for you inside its own tool loops, and it keeps what you need to do the same with conversations you store:

• Every step's additionalContent['content_blocks'] holds the response's content array as returned: thinking, text, tool calls, server tool blocks and fallback markers, in order. The AssistantMessage Prism builds for each step carries the same keys, so the stored messages replay the conversation exactly:
  • For asText(), persist $response->messages. It holds the messages you sent, every step of the tool loop and the final assistant message. $response->steps->last()->messages is only what Prism sent for the last step, so it stops before the final answer.
  • For asStructured(), which has no messages, persist $response->steps->last()->messages. Structured steps record the assistant message they produced, so it's already there.
  • For streams, persist the messages StreamedMessages::fromEvents($events) builds in your stream callback, after the messages you sent.
• A turn that ends without tool_use (a refusal, max_tokens) never runs its tool calls, so Prism stores it without them: content_blocks leave out its tool_use blocks and any server tool use left without a result (except on pause_turn, where it goes back unchanged so the turn can resume), and the step has no toolCalls. A refused turn replays as whatever text it produced, so don't expect tool calls from it when you continue the conversation. The untouched response is still in the step's raw.
• When an AssistantMessage is mapped back to Anthropic, Prism uses the first of these it finds: content_blocks, then thinking_blocks (followed by the message's text and tool calls), then the legacy thinking/thinking_signature pair. Messages stored before content_blocks existed keep replaying as they always have.
• To deliberately drop replay state (for example, when you edit a message), remove those keys from additionalContent.

A cacheType on an AssistantMessage that carries content_blocks goes on its last cacheable block; thinking and fallback blocks never get cache_control.

Extended output mode

Claude Sonnet 3.7 also brings extended output mode which increase the output limit to 128k tokens.

This feature is currently in beta, so you will need to enable to by adding output-128k-2025-02-19 to your Anthropic anthropic_beta config (see Configuration above).

Streaming

Claude supports streaming responses in real-time. All the standard streaming methods work with Anthropic models:

// Stream events
$stream = Prism::text()
    ->using('anthropic', 'claude-3-7-sonnet-latest')
    ->withPrompt('Write a story')
    ->asStream();

// Server-Sent Events
return Prism::text()
    ->using('anthropic', 'claude-3-7-sonnet-latest')
    ->withPrompt(request('message'))
    ->asEventStreamResponse();

Streaming with Thinking

When using thinking (adaptive or manual), the reasoning process streams separately from the final answer:

use Prism\Prism\Enums\StreamEventType;

foreach ($stream as $event) {
    match ($event->type()) {
        StreamEventType::ThinkingDelta => echo "[Thinking] " . $event->delta,
        StreamEventType::TextDelta => echo $event->delta,
        default => null,
    };
}

For complete streaming documentation including Vercel Data Protocol and WebSocket broadcasting, see Streaming Output.

Step metadata in streams

Streams carry the same metadata as asText(), split across two events:

• Each StepFinishEvent has the step's additionalContent: content_blocks, thinking_blocks, thinking, thinking_signature, stop_reason, stop_details, fallbacks, usage_iterations and the citation/provider tool keys, plus that step's usage. These replay blobs stay out of toArray(), so they never reach SSE, Data Protocol or broadcast output.
• The StreamEndEvent carries the final step's thinking, thinking_signature, stop_reason, stop_details, fallbacks and usage_iterations, but not content_blocks or thinking_blocks.

If you persist streamed conversations, pass the events your stream callback receives to Prism\Prism\Streaming\StreamedMessages::fromEvents() (see Persisting Streamed Conversations). It builds one AssistantMessage per step and attaches that step's additionalContent, so the stored messages replay exactly like the ones Prism sends inside its own tool loop.

The stream's finishReason reflects Anthropic's stop_reason, and Prism only runs the tools Claude asked for when the turn actually ended with tool_use. A turn that ends in a refusal, max_tokens or an error doesn't execute the tool calls it streamed. You still get their ToolCallEvents, but the step's content_blocks leave them out (along with any block the stream never finished), the StepFinishEvent reports the step's finishReason, and the AssistantMessage StreamedMessages builds for that step has no tool calls.

Paused turns

Server tools such as web search, web fetch and code execution run in a loop on Anthropic's side. When that loop reaches its iteration limit, the response stops with pause_turn (FinishReason::Pause), and the turn continues once you send the paused assistant content back as it is.

Prism resumes paused turns for you in asText(), asStructured() and asStream(), within maxSteps. The paused response becomes a step with FinishReason::Pause, its content_blocks go back unchanged (server tool use still waiting on its result included), and the next request picks the turn up. Structured output comes from the final step. A stream stays a single stream: each paused step ends with its own StepFinishEvent, and there is one StreamEndEvent at the very end.

Every resumed request is a step, so it counts toward maxSteps, and the default of 1 leaves no room to resume. You only see FinishReason::Pause on the response (or the StreamEndEvent) when maxSteps runs out before the turn finishes. Raise maxSteps when you use server tools, or continue from the messages yourself: send the stored messages back as they are, and the paused assistant message at the end resumes the turn.

use Prism\Prism\Enums\FinishReason;
use Prism\Prism\ValueObjects\ProviderTool;

$response = Prism::text()
    ->using('anthropic', 'claude-opus-5-5')
    ->withProviderTools([new ProviderTool(type: 'web_search_20250305', name: 'web_search')])
    ->withMaxSteps(5)
    ->withPrompt('Summarize this week\'s Laravel releases.')
    ->asText();

if ($response->finishReason === FinishReason::Pause) {
    // Still paused after 5 steps: raise maxSteps or continue from $response->messages.
}

A paused turn never asks for your tools, so Prism doesn't run client tool calls from one.

Refusals

Claude's safety classifiers can decline a request. That comes back as a normal response, not an error, with FinishReason::Refusal. Check the finish reason rather than the content, since a refusal can arrive before any output or part-way through it:

use Prism\Prism\Enums\FinishReason;

$response = Prism::text()
    ->using('anthropic', 'claude-opus-5-5')
    ->withPrompt($prompt)
    ->asText();

if ($response->finishReason === FinishReason::Refusal) {
    $response->additionalContent['stop_details'];
    // ['type' => 'refusal', 'category' => 'cyber', 'explanation' => '...']
}

Every step and response carries the raw stop_reason in additionalContent, and stop_details whenever Anthropic sends one. Both category and explanation can be null. The same keys are on the StepFinishEvent and StreamEndEvent when streaming. model_context_window_exceeded maps to FinishReason::Length.

A refusal before any output still adds an assistant message to $response->messages, but that message has nothing to replay. Prism leaves it out of the next request, since an empty assistant turn reads to the model as a reply it skipped. The refused user prompt stays in the history, though, so if you continue the conversation as is, the model will usually decline again. Drop the refused exchange from the messages you send back, or use fallbacks to get an answer in the same call.

Fallbacks

Anthropic can retry a declined request on another model inside the same API call. Pass fallbacks as a provider option, either 'default' (Anthropic picks the recommended fallback for the refusal category) or a list of up to three models with optional overrides. Prism sends it as the top-level fallbacks field and adds the server-side-fallback-2026-07-01 beta header for you, on text, structured and streaming requests and on every step of a tool loop.

$response = Prism::text()
    ->using('anthropic', 'claude-opus-5-5')
    ->withPrompt($prompt)
    ->withProviderOptions([
        'fallbacks' => 'default',
        // or: 'fallbacks' => [['model' => 'claude-opus-4-8']],
    ])
    ->asText();

$response->meta->model;                            // the model that served the response
$response->additionalContent['fallbacks'];         // [['from' => 'claude-opus-5-5', 'to' => 'claude-opus-4-8']]
$response->additionalContent['usage_iterations'];  // one entry per attempt, with its own model

A few things worth knowing:

• fallbacks is only present when the response contains a fallback block. With sticky routing, a later turn can go straight to the fallback model; that turn has no fallback block, so look for a fallback_message entry in usage_iterations and at meta->model.
• usage_iterations is Anthropic's raw usage.iterations for that step. Prism's Usage keeps the API's meaning: the attempt that produced the returned message, never tokens from different models added together. Bill each attempt at its own model's rates, and for a tool loop sum usage_iterations across $response->steps, since each step is a separate API call.
• Replayed turns follow Anthropic's echo rules for you: the fallback blocks stay where they appeared, text and everything after the last fallback is kept, and declined thinking, connector text, tool calls and unpaired server tool calls before it are dropped. Tool calls made before the last fallback are never executed.
• Claude Haiku 5.5 has no server-side fallback, and fallbacks isn't available on the Message Batches API or on Bedrock, Google Cloud or Foundry.

Documents

Anthropic supports PDF, text and markdown documents. Note that Anthropic uses vision to process PDFs under the hood, and consequently there are some limitations detailed in their feature documentation.

See the Documents on how to get started using them.

Anthropic also supports "custom content documents", separately documented below, which are primarily for use with citations.

Custom content documents

Custom content documents are primarily for use with citations (see below), if you need citations to reference your own chunking strategy.

use Prism\Prism\Enums\Provider;
use Prism\Prism\Facades\Prism;
use Prism\Prism\ValueObjects\Messages\UserMessage;
use Prism\Prism\ValueObjects\Media\Document;

Prism::text()
    ->using(Provider::Anthropic, 'claude-3-5-sonnet-20241022')
    ->withMessages([
        new UserMessage(
            content: "Is the grass green and the sky blue?",
            additionalContent: [
                Document::fromChunks(["The grass is green.", "Flamingos are pink.", "The sky is blue."])
            ]
        )
    ])
    ->asText();

Citations

Prism supports Anthropic's citations feature for both text and structured.

Please note that citations cannot be used with native structured output mode (the default). If you need citations with structured output, use tool calling mode via withProviderOptions(['use_tool_calling' => true, 'citations' => true]). Note however that citations with tool calling mode can produce unreliable output, so you should implement proper error handling.

Code execution

Anthropic offers built-in code execution capabilities that allow your AI to run code in a secure environment. This is a provider tool that executes code using Anthropic's infrastructure. For more information about the difference between custom tools and provider tools, see Tools & Function Calling.

To enable code execution, you will first need to enable the beta feature.

Either in prism/config.php:

        'anthropic' => [
            ...
            'anthropic_beta' => 'code-execution-2025-05-22',
        ],

Or in your env file (assuming config/prism.php reflects the default prism setup):

ANTHROPIC_BETA="code-execution-2025-05-22"

You may then use code execution as follows:

use Prism\Prism\Facades\Prism;
use Prism\Prism\ValueObjects\ProviderTool;

Prism::text()
    ->using('anthropic', 'claude-3-5-haiku-latest')
    ->withPrompt('Solve the equation 3x + 10 = 14.')
    ->withProviderTools([new ProviderTool(type: 'code_execution_20250522', name: 'code_execution')])
    ->asText();

Enabling citations

Anthropic require citations to be enabled on all documents in a request. To enable them, using the withProviderOptions() method when building your request:

use Prism\Prism\Enums\Provider;
use Prism\Prism\Facades\Prism;
use Prism\Prism\ValueObjects\Messages\UserMessage;
use Prism\Prism\ValueObjects\Media\Document;

$response = Prism::text()
    ->using(Provider::Anthropic, 'claude-3-5-sonnet-20241022')
    ->withMessages([
        new UserMessage(
            content: "Is the grass green and the sky blue?",
            additionalContent: [
                Document::fromChunks(
                    chunks: ["The grass is green.", "Flamingos are pink.", "The sky is blue."],
                    title: 'The colours of nature',
                    context: 'The go-to textbook on the colours found in nature!'
                )
            ]
        )
    ])
    ->withProviderOptions(['citations' => true])
    ->asText();

Accessing citations

You can access the chunked output with its citations via the additionalContent property on a response, which returns an array of MessagePartWithCitationss.

As a rough worked example, let's assume you want to implement footnotes. You'll need to loop through those chunks and (1) re-construct the message with links to the footnotes; and (2) build an array of footnotes to loop through in your frontend.

use Prism\Prism\ValueObjects\MessagePartWithCitations;
use Prism\Prism\ValueObjects\Citation;

$messageChunks = $response->additionalContent['citations'];

$text = '';
$footnotes = [];

$footnoteId = 1;

/** @var MessagePartWithCitations $messageChunk  */
foreach ($messageChunks as $messageChunk) {
    $text .= $messageChunk->outputText;
    
    /** @var Citation $citation */
    foreach ($messageChunk->citations as $citation) {
        $footnotes[] = [
            'id' => $footnoteId,
            'document_title' => $citation->sourceTitle,
            'reference_start' => $citation->sourceStartIndex,
            'reference_end' => $citation->sourceEndIndex
        ];
    
        $text .= '<sup><a href="#footnote-'.$footnoteId.'">'.$footnoteId.'</a></sup>';
    
        $footnoteId++;
    }
}

Note that when using streaming, Anthropic does not stream citations in the same way. Instead, of building the context as above, yield text to the browser in the usual way and pair text up with the relevant footnote using the citationIndex on the text chunk's additionalContent parameter.

Considerations

Message Order

• Message order matters. Anthropic is strict about the message order being:

1. UserMessage
2. AssistantMessage
3. UserMessage

Structured Output

Prism supports two approaches for structured output with Anthropic models:

Native Structured Outputs (Default)

Prism uses Anthropic's native structured outputs by default. This provides guaranteed schema compliance through constrained decoding — no beta header required.

use Prism\Prism\Enums\Provider;
use Prism\Prism\Facades\Prism;
use Prism\Prism\Schema\ObjectSchema;
use Prism\Prism\Schema\StringSchema;

$response = Prism::structured()
    ->withSchema(new ObjectSchema(
        'weather_report',
        'Weather forecast with recommendations',
        [
            new StringSchema('forecast', 'The weather forecast'),
            new StringSchema('recommendation', 'Clothing recommendation')
        ],
        ['forecast', 'recommendation']
    ))
    ->using(Provider::Anthropic, 'claude-sonnet-4-5-20250929')
    ->withPrompt('What\'s the weather like and what should I wear?')
    ->asStructured();

Benefits of native structured outputs:

• Always valid JSON: No more parsing errors or malformed responses
• Type safe: Guaranteed field types and required fields

Limitations:

• Only available on Claude Sonnet 4.5+ and Claude Opus 4.1+
• Cannot be used with citations
• Some JSON Schema features are not supported (see Schema Limitations)

Tool Calling Mode

For older models that don't support native structured outputs, or when dealing with complex content or non-English text that may contain quotes, you can use tool calling mode:

use Prism\Prism\Enums\Provider;
use Prism\Prism\Facades\Prism;
use Prism\Prism\Schema\ObjectSchema;
use Prism\Prism\Schema\StringSchema;

$response = Prism::structured()
    ->withSchema(new ObjectSchema(
        'weather_report',
        'Weather forecast with recommendations',
        [
            new StringSchema('forecast', 'The weather forecast'),
            new StringSchema('recommendation', 'Clothing recommendation')
        ],
        ['forecast', 'recommendation']
    ))
    ->using(Provider::Anthropic, 'claude-3-5-sonnet-latest')
    ->withPrompt('What\'s the weather like and what should I wear?')
    ->withProviderOptions(['use_tool_calling' => true])
    ->asStructured();

Benefits of tool calling mode:

• More reliable JSON parsing, especially with quotes and special characters
• Better handling of non-English content (Chinese, Japanese, etc.)
• Reduced risk of malformed JSON responses
• Compatible with thinking mode

Limitations:

• Cannot be used with citations (citations are not supported in tool calling mode)
• Slightly more complex under the hood but identical API usage

IMPORTANT

Without custom tools or thinking, tool calling mode forces Claude to call the structured output tool (tool_choice of type tool). Claude Opus 5.5, Sonnet 5.5 and Fable 5.1 reject forced tool use with a 400. On those models use native structured output (the default), or keep tool calling mode and pass ->withToolChoice(ToolChoice::Auto), which Prism honours. With Auto the structured output tool is best effort rather than schema-enforced, so native mode remains the recommendation. ToolChoice::None is still rejected in tool calling mode.

Prism adds the "use the structured output tool" instruction once, before the first request, so multi-step tool loops stay append-only (required for preserved thinking on newer models).

Combining Custom Tools with Structured Output

You can combine custom tools with structured output to gather data before returning a structured response. This requires tool calling mode to be enabled:

use Prism\Prism\Facades\Prism;
use Prism\Prism\Schema\ObjectSchema;
use Prism\Prism\Schema\StringSchema;
use Prism\Prism\Tool;

$schema = new ObjectSchema(
    'weather_analysis',
    'Analysis of weather conditions',
    [
        new StringSchema('summary', 'Summary of the weather'),
        new StringSchema('recommendation', 'Recommendation based on weather'),
    ],
    ['summary', 'recommendation']
);

$weatherTool = Tool::as('get_weather')
    ->for('Get current weather for a location')
    ->withStringParameter('location', 'The city and state')
    ->using(fn (string $location): string => "Weather in {$location}: 72°F, sunny");

$response = Prism::structured()
    ->using('anthropic', 'claude-3-5-sonnet-latest')
    ->withSchema($schema)
    ->withTools([$weatherTool])
    ->withMaxSteps(3)
    ->withProviderOptions(['use_tool_calling' => true]) // Required for Anthropic
    ->withPrompt('What is the weather in San Francisco and should I wear a coat?')
    ->asStructured();

// Access structured output
dump($response->structured);

// Access tool execution details
foreach ($response->toolCalls as $toolCall) {
    echo "Called: {$toolCall->name}\n";
}

IMPORTANT

When using custom tools with structured output on Anthropic, you must:

• Set use_tool_calling: true in provider options
• Set maxSteps to at least 2

For complete documentation on combining tools with structured output, see Structured Output - Combining with Tools.

Strict Tool Use

You can enable strict validation for tool inputs, which guarantees that tool parameters exactly match your schema through constrained decoding.

To enable strict mode for a tool, use the strict provider option:

use Prism\Prism\Facades\Prism;
use Prism\Prism\Tool;

$weatherTool = Tool::as('get_weather')
    ->for('Get current weather for a location')
    ->withStringParameter('location', 'The city and state')
    ->withProviderOptions(['strict' => true])
    ->using(fn (string $location): string => "Weather in {$location}: 72°F, sunny");

$response = Prism::text()
    ->using('anthropic', 'claude-sonnet-4-5-20250929')
    ->withTools([$weatherTool])
    ->withPrompt('What is the weather in San Francisco?')
    ->asText();

Benefits of strict tool use:

• Functions receive correctly-typed arguments every time
• No need to validate tool inputs
• Eliminates runtime errors from type mismatches
• Production-ready agents that work consistently

Schema Limitations

When using native structured outputs, certain JSON Schema features are not supported by Anthropic's constrained decoding:

Not Supported:

• Recursive schemas
• Numerical constraints (minimum, maximum, multipleOf)
• String constraints (minLength, maxLength)
• Complex regex patterns (lookahead/lookbehind, backreferences)
• External $ref definitions

Supported:

• All basic types (object, array, string, integer, number, boolean, null)
• enum for simple types (strings, numbers, booleans)
• anyOf and allOf (with limitations)
• required and additionalProperties: false
• String formats (date-time, email, uri, uuid, etc.)

If you use an unsupported feature, Anthropic will return a 400 error with details. For more information, see Anthropic's structured outputs documentation.

Limitations

Messages

Most providers' API include system messages in the messages array with a "system" role. Anthropic does not support the system role, and instead has a "system" property, separate from messages.

Therefore, for Anthropic we:

• Filter all SystemMessages out, omitting them from messages.
• Always submit the prompt defined with ->withSystemPrompt() at the top of the system prompts array.
• Move all SystemMessages to the system prompts array in the order they were declared.

Images

Does not support Image::fromURL