OpenAI SDK v3.35.0 Update: Agents Gains Ultrafast Service Tier

OpenAI SDK v3.35.0 Update: Agents Gains Ultrafast Service Tier

The official OpenAI SDK has released version v3.35.0, with this update focusing on realtime API capabilities and the expansion of performance tiers for Agents. For developers building voice, realtime conversation, or agent-based applications, several changes are worth noting.

Realtime API: New Typed Live Sideband Events

This release adds typed Live sideband events to the realtime API (PR #4099).

Previously, sideband events often required developers to parse raw structures themselves, and type information was not explicit enough. With the new typed events, developers can get more accurate type hints in code, reducing manual parsing and runtime errors. Sideband connections are mainly used to receive event streams outside the realtime session, such as transcription results and tool call status. With typing, they are better suited for strongly typed language projects such as TypeScript.

For teams already using the Realtime API to build voice assistants, this change is an incremental enhancement: business logic does not need to be rewritten, but the development experience and code maintainability can be noticeably improved.

Sideband connections are an event channel in the Realtime API that is independent of the main WebSocket session: the main session carries bidirectional audio stream transmission, while the sideband is used to send control-type events such as transcription text, tool call status, and session updates to another connection. This design allows the server to observe session state without interfering with the audio link, and makes it easier for backend processes to consume events independently. Before typing, these events were usually exposed as unannotated raw objects, and TypeScript developers had to compare event names and field structures themselves. After typing, the SDK generates corresponding discriminated union types for each event category, and editors can directly provide field completion and narrowing hints, so field spelling or structural misuse can be caught at compile time.

Agents Adds ultrafast Service Tier

Another feature is the addition of the ultrafast service tier for Agents (PR #4101).

A service tier is essentially a control mechanism for request scheduling priority. Agent scenarios usually involve multiple rounds of tool calls and longer reasoning chains, and end-to-end latency directly affects user experience. The new ultrafast tier means developers can choose a lower-latency scheduling option, suitable for agent workflows that are sensitive to response speed.

One detail: the original text did not disclose the specific latency metrics, billing method, or availability scope of ultrafast, so its actual effect still needs to be evaluated together with official documentation and real-load testing.

Service tier in the OpenAI API is usually passed as a request parameter to choose between standard scheduling and higher-priority scheduling. Its cost is generally reflected in unit billing price or quota consumption rather than a change in interface form. Multi-round tool calls in Agent workflows mean that a single user request may trigger multiple model calls serially, and the network and queueing latency of each hop accumulates, so scheduling priority has a more obvious impact on end-to-end latency than single-turn conversations. ultrafast is a newly added tier, alongside existing tiers such as priority, and developers must explicitly specify it in the request for it to take effect; if not specified, the default tier is still used. Since the official rate limits and pricing for this tier have not yet been announced, it is recommended to first run a small-traffic canary comparison of P95 latency before deciding whether to switch fully.

Fix for WebSocket Byte Reception

On the bug-fix side, the realtime module fixes an issue where byte reception was broken under websockets 13 (PR #4098).

For realtime applications that transmit binary audio or custom binary protocols over WebSocket, such issues often appear as data truncation or parsing failures and are costly to troubleshoot. This fix targets compatibility with websockets 13, showing that the SDK keeps pace with changes in underlying dependency versions. Developers using realtime audio streams or binary frames are advised to upgrade and run regression tests.

websockets is a mainstream WebSocket client/server implementation library in the Python ecosystem. Its version 13 adjusted the underlying receive logic, causing some dependents to experience binary frame handling exceptions. When the SDK receives audio chunks through this library, if the byte object is mistakenly decoded as text or truncated during transmission, the upper layer receives a buffer of incorrect length, manifesting as audio stuttering, frame boundary corruption, or JSON parsing errors. The difficulty with such issues is that symptoms appear at the business layer, while the root cause lies in transport-layer dependencies. This fix shows that the SDK has adapted to the receive path of websockets 13. Projects pinned to older versions should also confirm their own dependency resolution results to avoid being passively affected by the same type of issue due to indirect dependency upgrades.

Upgrade Recommendations

Judging from the version number, 3.35.0 is a minor version iteration with no indication of breaking changes. Teams relying on the realtime API should upgrade first to get the WebSocket fix; projects using Agents and sensitive to latency can evaluate whether the ultrafast tier meets their needs.

The version was released on October 17, and the commit carries a GitHub verified signature. The full diff can be checked through the official changelog.

Overall, this update continues the pace of OpenAI SDK investment in realtime capabilities: filling out the type system on one hand, adding performance tuning options for agent scenarios on the other, while also fixing underlying transport compatibility issues.

One-line Summary

OpenAI SDK releases 3.35.0, adding typed sideband events and the Agents ultrafast tier while fixing WebSocket issues.

Paragraph Summary

The official OpenAI SDK has released v3.35.0, a minor version iteration with no breaking changes, focused on two areas: the realtime API and Agents. The realtime API adds typed Live sideband events, giving events such as transcription results and tool call status clear type hints, reducing manual parsing and runtime errors, with obvious benefits for TypeScript projects. Agents adds an ultrafast service tier, providing a higher-priority scheduling option for low-latency agent workflows with multi-round tool calls and long reasoning chains, but the official specifics on latency metrics, billing, and availability have not been disclosed and need to be evaluated with documentation and real-load testing. On the bug-fix side, it resolves broken byte reception in the realtime module under websockets 13, which is especially critical for applications transmitting binary audio or custom protocols. The author recommends that teams relying on the realtime API upgrade first, and latency-sensitive Agents projects can evaluate the new tier.