What LangChain4j 1.22.0 adds on top of 1.21: OpenAI decision models, images on Decision Services, Anthropic automatic prompt caching, Bedrock batch chat, and async MCP cancellation.
OpenAI decisions
OpenAiDecisionModel and OpenAiOfficialDecisionModel call gpt-6-luna. The model name is required.
Images and scales
A Decision Service can take an ImageContent. @Description on a scale level is the description the model sees.
Anthropic cache
cacheAutomatically is off by default. It pays when the next call keeps the history. Independent calls can cost more.
Bedrock batch
BedrockBatchChatModel writes JSONL to S3 and starts a Converse batch job. The job name is the idempotency token.
MCP async
Cancelling executeToolAsync now cancels the request. Over stdio it sends notifications/cancelled. The server may still ignore it.
Quiet fixes
An SPI warning fires only on a real priority tie. A @UserMessage TextContent is no longer sent twice.