LiteLLM 1.98.0 Adds Native Structured Output for GLM 5 and DeepSeek V3.2 on Bedrock and Hashes Raw API Keys Before Writing Spend Logs

LiteLLM GitHub Releases·low signalXBlueskyLinkedInCopy link

Shipped August 23, this release is mostly proxy correctness work but two items matter if you run LiteLLM as a gateway. Native structured output is now enabled for GLM 5 and DeepSeek V3.2 on Bedrock, and toolSpec.strict is dropped for Claude Sonnet 5 on Converse. On the security side, raw API keys are now hashed before being persisted to spend logs, and per-model budgets were fixed so tracking, enforcement and reporting all read the same counter rather than diverging. Provisioned-throughput flat cost can now be configured per deployment with daily per-model rollups by active hour.