LiteLLM v1.98.0 Ships Provisioned Throughput Pricing

What shipped? LiteLLM v1.98.0, published on PyPI and Docker Hub on August 22.

What changed? Three things worth flagging:

Provisioned Throughput (PTU) pricing. You can now configure a flat PTU cost on individual model deployments via the provisioned_throughput field. This lets operators track reserved capacity spend per-model instead of lumping it into generic costs. For teams running dedicated endpoints on Bedrock or Vertex, this is the difference between knowing your numbers and guessing them.

Bedrock fix for Claude Sonnet 5. Drops toolSpec.strict on the Converse API — Anthropic's latest Sonnet doesn't support it, and the proxy was rejecting valid calls. If you upgraded to Sonnet 5 and saw silent tool call failures, this is your fix.

Cosign-signed Docker images. Every image now ships with a Sigstore cosign signature. Verifiable via pinned commit hash or release tag. For production deployments that need a supply-chain attestation trail, this removes the last excuse not to verify.

Plus a batch of proxy fixes: file-list pagination is now scoped to the caller, Prisma-dependent hooks are skipped when no database is attached, and 914 type errors were cleaned out of 16 core files.

Why does a builder care? PTU cost attribution is the feature that pays for itself on the first dedicated deployment — if you're budgeting GPU reservations at scale, this saves the spreadsheet reconciliation work.