ApiDelta — May 13, 2026

Today's updates focus on new tool‑search capabilities, agent executor changes, a fast‑mode pricing tier for Claude Opus, and several library enhancements.

Native tool‑search added to Anthropic and OpenAI clients

Pydantic‑AI 1.95.0 introduces built‑in “Tool Search” that lets you query external search services (e.g., vector stores, web APIs) directly from Anthropic and OpenAI tool calls, with pluggable strategies for any provider. Update your pydanticai integration to use the toolsearch parameter instead of manual pre‑search logic. Source

CrewAI deprecates `CrewAgentExecutor` and switches default agents to `AgentExecutor`

Version 1.14.5a5 removes the legacy CrewAgentExecutor class; new projects should instantiate AgentExecutor directly. The release also adds improved sandbox tools for the Daytona environment. Adjust imports and any custom subclassing accordingly. Source

Claude Opus “Fast” mode launches at 6× premium price

OpenRouter now lists a fast‑mode variant of Claude Opus 4.7 that delivers the same output quality with roughly six times higher throughput, billed at a premium rate. Select the claude-opus-4.7-fast model identifier to enable it; expect higher per‑token costs. Source

LangChain 1.3.0 adds `version="v3"` support for event streaming

The streamevents and astreamevents helpers now accept version="v3" to align with the latest LangChain agent event schema. Update your streaming callbacks if you rely on version‑specific fields. Source

llama.cpp  b9127 adds optional Adreno X‑Mem F16×F32 GEMM for OpenCL prefill

The OpenCL backend now includes an opt‑in kernel for Adreno GPUs that uses mixed‑precision GEMM during the prefill stage, improving latency on supported Android devices. Enable it with the --adreno-xmem flag when building. Source

Claude service incidents resolved across Opus, Sonnet, and Haiku models

Claude’s status page reports that elevated error rates affecting Opus 4.7, Sonnet 4.6, and Haiku 4.5 have been fixed and monitoring is complete. No code changes are required, but consider adding retry logic if your application experienced transient failures yesterday. Source Source Source

---

Tooling note — Cap LLM spend, auto-downgrade to cheaper models on deprecation. → BudgetForge