1.8.0

What's New

This release brings the response layer up to date with the Anthropic API changes shipped between April and July 2026: new usage fields, server-side fallback, compaction, and cache diagnostics.

New usage fields

Three additions to usage on Messages responses:

$response->usage->outputTokensDetails?->thinkingTokens; // how many billed output tokens were reasoning
$response->usage->speed;                                // 'fast' or 'standard' on fast mode requests
$response->usage->iterations;                           // per-attempt records under the fallback and compaction betas

When streaming, outputTokensDetails arrives on the final message_delta event.

Server-side fallback (beta)

When a refused request is retried on another model (the fallbacks parameter with the server-side-fallback-2026-07-01 beta header), the response now parses fully:

• fallback content blocks mark the handoff: $block->from->model and $block->to->model
• usage->iterations records every attempt: a message entry for a declined model, a fallback_message entry for the model that served the turn
• stop_details->recommended_model names a model to retry directly when the fallback itself was rate limited
• stop_details->category now covers frontier_llm, reasoning_extraction, and general_harms alongside cyber and bio

Compaction (beta)

Under the compact-2026-01-12 beta, responses can include a compaction content block whose content is a string summary. The SDK now handles that block, the compaction_delta stream event, the top-level context_management response field, and the stop_reason: 'compaction' value that appears with pause_after_compaction. All of it was verified against live API responses, including the quirk that iteration entries under this beta carry no model key.

foreach ($response->content as $block) {
    if ($block->type === 'compaction') {
        $block->content; // 'Summary of the conversation: ...'
    }
}

$response->context_management; // ['applied_edits' => [...]] or null

Cache diagnostics (beta)

With the cache-diagnosis-2026-04-07 header and a diagnostics request parameter, the API reports where two consecutive requests diverged:

$response->diagnostics?->cache_miss_reason?->type;                      // 'system_changed'
$response->diagnostics?->cache_miss_reason?->cache_missed_input_tokens; // 41850

On streams, diagnostics arrives on the message_start event.

Documentation

The Messages, Thinking, Streaming, and Server Tools guides now cover refusal handling with fallback, compaction, cache diagnostics, thinking token usage, and the latest server tool versions (web_search_20260318, web_fetch_20260318, code_execution_20260521).

Full Changelog: 1.7.0...1.8.0