Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/phpaisdk/anthropic/llms.txt

Use this file to discover all available pages before exploring further.

Claude supports extended thinking via thinking content blocks. When reasoning is enabled, Claude performs additional internal reasoning before producing its final response. The PHP AI SDK exposes this through Reasoning::effort(), which the Anthropic provider converts to a budget_tokens value and sends in the thinking field of the request body.

​
Basic Usage

Attach Reasoning::effort() to a generation request. The effort level controls how many tokens Claude is allowed to spend on its internal reasoning chain. 
use AiSdk\Anthropic;
use AiSdk\Generate;
use AiSdk\Reasoning;

$result = Generate::text('Solve: what is 2+2?')
    ->model(Anthropic::model('claude-sonnet-4'))
    ->reasoning(Reasoning::effort('high'))
    ->run();
Accepted effort values are 'low', 'medium', and 'high'. Each maps to a percentage band of the model’s configured max_tokens.

​
Budget Calculation

AnthropicReasoningBudget calculates the token budget from the effort level and the request’s max_tokens value:
​
Minimum budget
int
default:"1024"
The smallest token budget the Anthropic API accepts. Calculated budgets below this floor are clamped to 1024.
​
Default percentage
float
default:"0.30"
When no effort-specific percentage applies, the budget defaults to 30 % of max_tokens.
​
Upper bound
constraint
The calculated budget must be strictly less than max_tokens. The Anthropic API rejects requests where budget_tokens >= max_tokens.
The thinking field sent to the API looks like this: 
{
  "thinking": {
    "type": "enabled",
    "budget_tokens": 8000
  }
}

​
Response

When thinking is enabled, the API may return one or more thinking content blocks alongside the usual text blocks. The response parser maps them as follows:
Anthropic block typeSDK response part
textTextPart
thinkingReasoningPart
Both part types are available in $result->parts. You can inspect reasoning content separately from the main text: 
use AiSdk\Responses\Parts\ReasoningPart;
use AiSdk\Responses\Parts\TextPart;

foreach ($result->parts as $part) {
    if ($part instanceof ReasoningPart) {
        echo "Thinking: " . $part->text . PHP_EOL;
    }
    if ($part instanceof TextPart) {
        echo "Answer: " . $part->text . PHP_EOL;
    }
}

​
Beta Header

Some Claude models require the anthropic-beta: extended-thinking-2025-05-14 header to enable thinking blocks. Pass it via the headers option when creating the provider: 
$provider = Anthropic::create([
    'apiKey'  => 'sk-ant-...',
    'headers' => ['anthropic-beta' => 'extended-thinking-2025-05-14'],
]);
Reasoning requires a Claude model that supports it (e.g., claude-sonnet-4, claude-opus-4, claude-3-7-sonnet). Check $model->supports(Capability::Reasoning) before enabling it to avoid sending unsupported requests.

Build docs developers (and LLMs) love

Get started for freeTalk to us