OpenAI SDKs Ship Four Releases in One Day: safety_identifier and Translation (v3.33-3.34, v7.36-7.37)

The openai-python 3.32.0 and openai-node 7.35.0: prompt-created voices, Live audio fixtures, and Decisions in Live covered v3.32.0/v7.35.0 barely a day ago, and OpenAI's two-package release train ran four more stops: on 2026-10-17 openai-python shipped v3.33.0 and v3.34.0 while openai-node shipped v7.36.0 and v7.37.0. Read together, the four release notes describe one expansion of the API surface around safety identifiers and realtime translation. Every parameter-level fact in this article comes from the day's SDK source diffs and the release notes' own wording.

safety_identifier: the image APIs get a unified safety parameter

v3.33.0 (python) and v7.36.0 (node) share one feature: add safety identifiers to image requests and responses (python #4093, node #2957). The parameter itself can be read verbatim from the v3.33.0 source diff: safety_identifier is a string or null with maxLength 128 and the example value user-1234. The official description reads "A stable identifier for an end user in your organization, used for safety monitoring and abuse detection" — and OpenAI advises hashing usernames or email addresses before sending them, so no identifying information travels in plain text.

The precedence rule is in the source too: a non-empty safety_identifier takes precedence over user; user is marked deprecated and remains accepted only when safety_identifier is omitted, null, or empty. The migration path matches the docs — OpenAI's safety best practices page (verified reachable, direct 200, on 2026-10-18) carries an "implement safety identifiers" section anchor.

Scope matters: the notes explicitly say image requests and responses — the same identifier rides the request in and appears on the response object, symmetrically. Whether other API surfaces already carry the parameter is not stated in either release, and this article does not infer beyond the wording.

Realtime Translation: a new session in python, generated bindings in node

v3.34.0's second feature is add Realtime Translation WebSocket sessions (#4095) — another WebSocket session type in the Realtime family. node's counterpart move appears in v7.37.0's Refactors section: realtime now uses generated Translation bindings (#2961). Neither note documents fields or endpoints beyond those lines; taken together, the two moves are the usual cadence — handwritten in python, code-generated in node — for aligning one API surface, and this article asserts nothing further.

The synchronized item and two secondary changes

A degraded external storage status appeared as a feature in both packages the same day (python #4097, node #2963). The release notes' entire description is that one line — by naming it looks like a new value on some status enum, but where it attaches and when it appears is undocumented, and this article does not vouch for inferred semantics.

Two secondary changes are worth recording. First, node v7.37.0's Chores section added shutdown notices for GPT-4 model aliases (#2959) — the SDK now flags GPT-4 aliases as shutting down, in step with the official 2026-10-14 deprecation wave (OpenAI's October 2026 deprecation wave: GPT-4o and GPT-4.1 on 2027-01-14, TTS to gpt-realtime-2.1-mini); the SDK notice carries no new dates, so timing remains governed by the official announcements. Second, python v3.33.0 shared the Decisions cache breakpoint parameter type (#4092) — a type-level refactor making the type reusable across surfaces, with no behavior change stated.

Version comparison

Publish times for all four versions follow GitHub releases' published_at (2026-10-17). python v3.33.0's semver comparison header says 2026-10-16, one day off the API field; this article standardizes on the API value.

Next steps

• Callers already using user for safety attribution: migrate to safety_identifier, pass a hashed identifier, and mind the 128-character cap.
• Teams on Realtime: watch how the Translation session surface aligns across the two packages before switching node over to the generated bindings.
• Tracking deprecation timelines: anchor on the OpenAI's October 2026 deprecation wave: GPT-4o and GPT-4.1 on 2027-01-14, TTS to gpt-realtime-2.1-mini; the SDK notices are a hinting layer only.