Spring AI 1.1 Released: Explosive New Features for Java AI Development

Spring AI 1.1.0 GA Release

Released on 2025‑05‑20, Spring AI 1.1.0 contains 854 changes: 354 feature enhancements, 241 bug fixes, 100 documentation updates, and 23 security patches.

1. Model Context Protocol (MCP) – out‑of‑the‑box support

MCP defines a public standard for AI models (e.g., Claude, Gemini, Qwen) to communicate with development tools and external services. Spring AI 1.1.0 provides auto‑configuration and annotation‑driven setup for MCP servers or clients. @McpTool – defines tool functions such as calling external APIs for real‑time data. @McpResource – accesses resources like database schemas. @McpPrompt – creates prompt templates to guide model output (e.g., generating SQL queries).

Example:

@McpTool
public String getCurrentWeather(String location) {
    // Call weather API, return result
    return weatherService.fetch(location);
}

@McpResource
public String getDatabaseSchema() {
    // Return database structure information
    return dbConnector.getSchema();
}

@McpPrompt
public String generateSqlQuery(String userIntent) {
    // Generate SQL based on user intent
    return sqlGenerator.create(userIntent);
}

Supported transport mechanisms include STDIO, HTTP SSE, and Streamable HTTP. Dedicated starters are available for WebFlux, WebMVC, and Servlet environments, with integrations for Docker Compose, Testcontainers, and OAuth2.

2. Prompt Caching

Prompt caching reduces the cost of large‑model API calls by up to 90% and speeds up responses. Supported providers:

Anthropic Claude – five cache modes: NONE , SYSTEM_ONLY , TOOLS_ONLY , SYSTEM_AND_TOOLS , CONVERSATION_HISTORY .

AWS Bedrock – uses the Converse API to cache prompts for Claude and Nova models.

3. Reasoning Mode & Self‑Evolving Agents

The new ReasoningContent API exposes the model’s reasoning process before an answer is returned, improving reliability and explainability. Supported models include Ollama, Zhipu AI, Anthropic, and OpenAI.

Recursive Advisor enables an advisor to invoke another advisor, forming multi‑step workflows (e.g., draft → self‑review → improvement). Built‑in implementations support monitoring, debugging, and “LLM‑as‑a‑Judge” patterns.

4. Model Ecosystem Expansion

New providers:

Google GenAI SDK – integrates Gemini Pro, 1.5 Pro, and 2.0 Flash with chat, embedding, and caching support.

ElevenLabs – streaming text‑to‑speech with multiple voices via a unified interface.

Enhanced existing providers:

OpenAI – adds GPT‑5 series, file API, enhanced TTS and transcription.

Anthropic Claude – supports Sonnet 4.5 and Opus 4.1, adds reference API, tool selection, and cache management.

Zhipu AI – adds GLM‑4.6, GLM‑4.5, and Z1 models.

Mistral AI – adds OCR API and Codestral embedding model.

5. Vector Store, Chat Memory & Observability

Vector‑store enhancements for Retrieval‑Augmented Generation (RAG):

MariaDB – similarity scoring.

OpenSearch – k‑NN search.

GemFire – metadata filtering.

Weaviate – custom configuration.

Introduces VectorStoreRetriever interface for secure read‑only access.

Chat memory now supports MongoDB, Oracle JDBC, and Azure Cosmos DB storage options.

Observability improvements via Micrometer add context propagation, logging, and metric mapping with native Prometheus and OpenTelemetry exporters.

Official documentation: https://spring.io/blog/2025/11/12/spring-ai-1-1-GA-released