# Pydantic AI 2.59.0 patches web_fetch slowdown, adds ElevenLabs realtime model

published: 2026-10-21T08:05:47.321Z
topic: frameworks
source: https://github.com/pydantic/pydantic-ai/releases/tag/v2.59.0
url: https://aiagentnews.io/stories/pydantic-ai-2-59-0-patches-web-fetch-slowdown-adds-elevenlabs-realtime-model

Pydantic AI v2.59.0 is out with a high-severity fix for slow HTML conversion in web_fetch that could delay unrelated work in the same process. The release also adds a Daytona sandbox, an ElevenLabs realtime speech model, and several compatibility changes.

The main fix is security advisory GHSA-4whf-38hw-pxxh, rated high. When web_fetch_tool fetched attacker-controlled HTML with many elements, the conversion step could delay other work running in the same process. The fix ships in v2.59.0 and v1.107.9, and the changelog advises upgrading if the tool fetches such pages.

New features include DaytonaSandbox, a workspace that runs Coder, Shell, FileSystem and other harness capabilities inside a Daytona sandbox, and ElevenLabsRealtimeModel for Eleven Realtime speech-to-speech models. FunctionToolset.instructions and instruction functions can now opt into on_change='append', and TemporalDurability gains non_retryable_model_errors to stop Temporal from retrying chosen model error categories.

The compatibility notes cover several changes that affect upgrades. The Groq models playai-tts, playai-tts-arabic, meta-llama/llama-guard-4-12b and meta-llama/llama-4-maverick-17b-128e-instruct were removed. The release also adds provider-neutral model error categories such as ModelRateLimitError and ModelUnavailableError, and separates omitted provider usage from reported zero through UsageLimitUnavailableWarning.

## Why it matters if you build agents

If your agent fetches arbitrary pages with web_fetch, upgrade to v2.59.0 or v1.107.9 to get the fix. Code that calls the removed Groq models will need a different model ID.

If your code branches on model failures, match the new error categories instead of provider-specific errors. Check any logic that treats missing usage data as zero, since the library now reports the two cases differently.