What changed
On 17 October 2026 the OpenAI Python SDK shipped version 3.35.0. The release notes add an ultrafast service tier for Agents and typed Live sideband events. The change is API surface, not a new public model name. It gives teams a faster lane when they run agent loops that call tools many times in one session.
Agent builders already split work between chat, tools, and realtime channels. A dedicated tier signals that latency and throughput for those loops are now a first class product choice. The same changelog also fixes a Realtime websocket byte receive issue with websockets 13.
Why it matters
Product teams that ship agents care about tail latency on tool rounds. A slow round blocks the whole chain. Security and platform teams care because more agent traffic can mean more credential use and more audit logs. Finance teams should watch list price and quota for a new tier before they move production traffic.
If you rely on the Python SDK, you need a pinned upgrade plan. Silent drift on Realtime clients can break voice or sideband features after a patch.
Who is affected
Applied AI leads, agent framework owners, and SRE teams running OpenAI backed workflows. Vendors that wrap the official SDK for customers should read the tier docs before they promise SLAs. Teams on older SDK lines should check whether they still receive Realtime fixes.
What to do next
Read the 3.35.0 changelog and the Agents tier documentation on your next sprint. Stage the SDK bump in a sandbox agent that hits tools in a loop. Measure p95 latency and error rate against your current default tier. Only then shift production agent traffic.
What to watch
Further SDK releases that expose tier selection in other client libraries. Any public pricing or rate limit page update tied to the ultrafast tier. Partner posts that show sideband event schemas in production.
Image: GitHub
How the tier fits agent design
Agents often run a plan, call a tool, read the result, and repeat. Each hop adds queue time and model time. An ultrafast tier is meant to shorten that cycle when the platform can honor it. It does not remove the need for good tool design or safe prompts.
Typed Live sideband events matter when you debug voice or streaming agents. Stronger types reduce guesswork when you parse event streams in client code.
Risks to weigh before you switch
A faster tier can cost more per token or per minute. Your finance partner should model spend if agent traffic doubles after a latency win. You should also log which tier each session used so incident reviews stay clear.
The Realtime websocket fix is a reminder that agent stacks sit on moving client libraries. Pin versions in production images and run weekly smoke tests on tool loops, not only on chat demos.
What this is not
This release does not announce a new frontier model for general chat. Treat tier changes as infrastructure for agent products you already ship. If your roadmap still uses batch jobs, this note may be a watch item rather than a same day upgrade.
Procurement teams should ask whether the ultrafast tier sits under the same data processing terms as your existing API contract. Legal review is cheap compared to moving high sensitivity agent traffic without a clear paper trail.
Image: GitHub
Sources
- Primary — OpenAI, openai-python CHANGELOG 3.35.0 (17 October 2026). Documents the ultrafast Agents tier and Live sideband event types.
- Secondary — OpenAI, openai-python release v3.35.0 (17 October 2026). Confirms the release tag and date.
Related in AI
Continue reading