Node.js SDK Changelog

​

0.32.1 (2026-11-12)

​

Changes

• benchmarks: auto_disallow_too_costly_reasoning_efforts also drops a model from the benchmark when, at the first reasoning effort it is tested at, it writes at least 250,000 output tokens to a call whose logged answer wrote fewer than 10,000

​

0.32.0 (2026-11-09)

​

Changes

• a model benchmark’s config takes ignore_max_tokens, to replay calls without the max_tokens they set
• a fixed order routing strategy’s allowed deployments take a reasoning_effort, the effort its calls are sent at

​

0.31.0 (2026-11-07)

​

Bug Fixes

• a model benchmark checks each reasoning effort’s cost before its first example and again after its first 3, so a too costly effort is dropped before most examples run it
• streamed calls with thinking are recorded like non-streamed ones, through the AI gateway and the SDKs

​

Changes

• the benchmark report and routing strategy pages show how much Weflayr’s caching saves
• the public API adds a Get API key info route, and its import past calls route takes named fields
• a model benchmark stops once it has spent its credit reservation, emails you what finishing it would cost and which reasoning efforts cost far more than expected, and reserves a 30% buffer instead of 20%
• a model benchmark stops testing a model at a reasoning effort costing far more than its medium effort
• model benchmarks and routing strategies can be managed through the public API and the MCP server

​

0.30.0 (2026-10-30)

​

Bug Fixes

• web searches run on Vertex are priced at Google’s price

​

Changes

• images and long prompts are now kept in your recorded calls instead of being stripped

​

0.29.1 (2026-10-19)

​

Bug Fixes

• a cache breakpoint placed on a block before the end of its message is recorded at that block’s position

​

0.29.0 (2026-10-13)

​

Changes

• propagate_metadata no longer requires a customer name, and its provider override is renamed provider_name_override_if_telemetry_instrumentation (providerNameOverrideIfTelemetryInstrumentation in JS/TS), since it only applies to calls the SDK instruments, not to calls sent through the AI gateway

​

0.28.0 (2026-10-12)

​

Bug Fixes

• Mistral tool calls are recorded correctly, and embeddings accept a bare text input
• Bedrock models served in the OpenAI format are recorded
• model names use American spelling

​

Changes

• inline images and documents are recorded without their content
• propagate_metadata takes a feature_run_uuid, to group the calls of one feature run
• the SDK instruments Mistral, including Mistral on Vertex
• new providers: xAI (Grok) and Alibaba (Qwen)
• calls to OpenAI’s /v1/responses endpoint are recorded
• the web search a provider runs is recorded and priced, for every provider that offers one, PDF inputs to xAI included
• the Google provider is named google instead of gemini
• calls are priced by their service tier (flex, priority) and by the region they ran in
• name each call with set_name_call, and send its metadata to the AI gateway in the x-weflayr-metadata header
• the SDKs no longer instrument the Vercel AI SDK, and AIProviderName.VERCEL_AI_GATEWAY is removed

​

0.27.0 (2026-08-26)

​

Bug Fixes

• prompts holding media (images, audio…) work in benchmarks and prompt caching analysis

​

0.26.1 (2026-08-13)

​

Bug Fixes

• fix the packaging of the SDK

​

0.26.0 (2026-08-13)

​

Bug Fixes

• Bedrock tool calls are recorded correctly

​

Changes

• support for Bedrock Mantle

​

0.25.0 (2026-08-07)

​

Changes

• new cost route, in the management API, the SDK and the MCP server
• new get_performance route
• get_margin returns more consistent figures
• GPT-5.6 prompt caching is recorded and priced
• the SDK is now written in TypeScript and ships its types