# Deep research report on new features in the OpenAI TypeScript/JavaScript SDK from v6.23.0 to v6.27.0 ## Executive summary Between **February 23, 2026** and **March 5, 2026**, the official OpenAI TypeScript/JavaScript SDK (the `openai` package, from the `openai/openai-node` repository) shipped a tightly linked cluster of features that collectively push the SDK toward **long-running, tool-heavy, agentic workflows**. citeturn46view0turn45view0 Across releases **v6.23.0 → v6.27.0**, the most consequential additions were: - **Responses API WebSocket mode support in the SDK** (v6.23.0), aligning with the platform’s WebSocket Mode guidance for persistent `/v1/responses` sessions and incremental continuation via `previous_response_id`. citeturn46view0turn33view0turn19view1 - **Realtime model ID expansions** (v6.24.0), adding `gpt-realtime-1.5` and `gpt-audio-1.5` into the SDK’s Realtime session model unions. citeturn46view0turn23view0turn24view0 - A new **assistant message `phase` field** (v6.25.0), designed to label “commentary” vs “final_answer” and improve reliability in longer, tool-rich runs (especially with Codex-family workflows). citeturn46view0turn26view0turn19view1turn16search1 - A major March 5 platform/SDK step: typed support for **`gpt-5.4`**, plus first-class **Tool Search** and the **GA `computer` tool**, with an SDK follow-up cleanup for the computer tool type naming in v6.27.0. citeturn45view0turn19view2turn29view0turn17view0turn32view2 - A smaller but practically important schema addition surfaced under “bug fixes”: the **`prompt_cache_key` parameter in Responses** (v6.26.0), which matters for prompt caching strategies in production. citeturn46view0turn39view2 Overall, these increments form a coherent story: **persistent sessions + explicit message staging (`phase`) + selective tool loading (tool search) + UI automation loops (computer tool) + new flagship models (gpt‑5.4, realtime 1.5)**. citeturn33view0turn17view0turn32view2turn19view2turn19view0 ## Scope and methodology This report covers **only** SDK changes introduced in **v6.23.0, v6.24.0, v6.25.0, v6.26.0, and v6.27.0** (inclusive). Release dates and change descriptions were sourced primarily from the official release notes on entity["company","GitHub","code hosting platform"]. citeturn45view0turn46view0 To verify what “feature added” means at the SDK surface (types, method availability, tool schemas), I cross-checked: - The **GitHub releases** for each version in scope. citeturn45view0turn46view0 - The corresponding **tagged source** in the repository (e.g., `v6.24.0`, `v6.26.0`, `v6.27.0`) for relevant TypeScript type additions/changes. citeturn23view0turn26view0turn29view0turn31view0turn11view0 - Official OpenAI platform documentation (guides + API reference), including WebSocket Mode, Tool Search, Computer Use, and model pages. citeturn33view0turn17view0turn32view2turn15search0turn6search2turn16search1 The SDK is generated from an OpenAPI specification using entity["company","Stainless","sdk generator"], which is relevant context for why many changes appear as schema/type updates tied closely to platform docs. citeturn13view0 All code samples below are **illustrative** and were **not executed** in a runtime environment for this report; each row includes an explicit verification note. ## Release-by-release analysis The five releases in scope line up closely with OpenAI platform feature launches in late February and early March 2026. In **v6.23.0 (2026-02-23)**, the SDK added **“websockets for responses api”**. citeturn46view0 This aligns with the platform’s WebSocket Mode guidance for the Responses API: persistent WebSocket connection to `/v1/responses`, and continuation by sending only incremental input plus `previous_response_id`. citeturn33view0turn19view1 In the tagged SDK code, this appears as a dedicated WebSocket helper (`ResponsesWS`) built atop a shared emitter base and a `wss://…/v1/responses` URL constructor. citeturn11view0turn9view0 The public API reference also describes a TypeScript entrypoint `client.responses.connect()` for connecting to a persistent Responses API WebSocket. citeturn44view0 (The reference page’s return type is shown as `void`; details about the returned handle are therefore treated as **unspecified** in this report.) citeturn44view0 In **v6.24.0 (2026-02-24)**, the SDK added **`gpt-realtime-1.5`** and **`gpt-audio-1.5`** to Realtime. citeturn46view0 This is concretely reflected in the Realtime session type union in `src/resources/realtime/realtime.ts`, where both new model IDs appear in v6.24.0 and are absent in v6.23.0. citeturn23view0turn24view0 The OpenAI API changelog explicitly notes `gpt-realtime-1.5` for the Realtime API on Feb 23, 2026, and `gpt-audio-1.5` for the Chat Completions API on the same date. citeturn19view0 The Realtime model page describes `gpt-realtime-1.5` as an audio model for low-latency voice agents; `gpt-audio-1.5` is described as an audio model release as well. citeturn6search2turn6search5 Since the SDK includes `gpt-audio-1.5` in the Realtime session model list, this report treats Realtime compatibility as **SDK-supported** (type-level), while explicit platform changelog framing for Realtime vs Chat Completions remains **partially specified**. citeturn23view0turn19view0 In **v6.25.0 (2026-02-24)**, the SDK introduced **`phase`** as a first-class field in Responses message types. citeturn46view0 In the SDK’s `EasyInputMessage` interface, `phase` is defined as `'commentary' | 'final_answer' | null`. citeturn26view0 The OpenAI API changelog describes `phase` as a new Responses API feature that labels assistant messages as commentary vs final answer. citeturn19view1 Platform guidance further elaborates why it exists (especially in Codex-like flows) and enumerates valid values. citeturn16search1turn16search4 In **v6.26.0 (2026-03-05)**, the SDK aligned with the platform’s **March 5, 2026** launches: adding **`gpt-5.4`**, introducing **Tool Search** schema support, and introducing the **GA `computer` tool**. citeturn45view0turn19view2turn29view0turn17view0turn32view2 Concretely: - `gpt-5.4` appears in the SDK’s `ChatModel` union in v6.26.0 and is absent in v6.25.0. citeturn29view0turn30view0 - The SDK adds a `ToolSearchTool` (`type: "tool_search"`) and makes `defer_loading?: boolean` available across tool definitions, matching platform guidance that Tool Search requires (1) adding `"tool_search"` and (2) marking tools as deferred. citeturn28view0turn28view1turn17view0 - Tool search call/output items appear as `tool_search_call` and `tool_search_output` shapes, including recorded items `ResponseToolSearchCall` and `ResponseToolSearchOutputItem`. citeturn31view0turn31view1 - The `computer` tool path is supported in the platform docs as `{ type: "computer" }` with `gpt-5.4`, and the SDK includes a `computer` tool interface plus the associated action schema used in `computer_call` loops. citeturn32view2turn28view4 Also in v6.26.0, the release notes explicitly call out a parameter addition under bug fixes: **“prompt_cache_key param in responses”**. citeturn46view0 The v6.27.0 Responses type comments also reference `prompt_cache_key` as a successor to a deprecated `user` field for caching optimizations, reinforcing that it is part of the SDK’s supported surface in this timeframe. citeturn39view2 In **v6.27.0 (2026-03-05)**, the SDK made a **developer-experience clean-up** around computer tool naming: “The GA ComputerTool now uses the ComputerTool class. The `computer_use_preview` tool is moved to ComputerUsePreview.” citeturn45view0 This corresponds to a swap/rename visible in the types across tags: in v6.26.0, the GA tool (`type: "computer"`) was represented as `ComputerUseTool` while the preview tool used `ComputerTool` (with display settings). citeturn28view3turn28view4 In v6.27.0, the GA tool becomes `ComputerTool { type: 'computer' }` and the preview tool becomes `ComputerUsePreviewTool { type: 'computer_use_preview', display_width, display_height, environment }`. citeturn10view0 ## Comprehensive feature table The table below lists each feature introduced within the scope window, with the required columns plus documentation status and verification notes. | Feature name | Release version | Concise feature description | Primary use-case(s) | Short JS/TS use-case example | Official doc / release note link | Documentation status | Verification note | |---|---|---|---|---|---|---|---| | Responses WebSocket support in SDK | v6.23.0 | Adds SDK support for Responses API WebSocket mode (“websockets for responses api”), enabling persistent `wss://…/v1/responses` sessions and event-driven streaming. citeturn46view0turn33view0turn11view0 | Long-running agent loops with many tool round trips; lower per-turn overhead via incremental input + `previous_response_id`. citeturn33view0 | `import OpenAI from "openai";`
`import { ResponsesWS } from "openai/resources/responses/ws";`
`const client = new OpenAI();`
`const ws = new ResponsesWS(client);`
`ws.on("response.output_text.delta", e => process.stdout.write(e.delta));`
`ws.send({ type:"response.create", model:"gpt-5.4", input:"Hello" });` | WebSocket Mode guide:
SDK release: | Documented in OpenAI guides + SDK release notes. citeturn33view0turn46view0 | Example structure validated against the SDK `ResponsesWS` implementation and WebSocket Mode guide; not executed. citeturn11view0turn33view0 | | Realtime model ID additions (`gpt-realtime-1.5`) | v6.24.0 | Adds `gpt-realtime-1.5` to the SDK’s Realtime session `model` union types. citeturn46view0turn23view0turn24view0 | Voice agents and low-latency multimodal conversations (Realtime API). citeturn6search2turn19view0 | `import OpenAI from "openai";`
`const client = new OpenAI();`
`const secret = await client.realtime.clientSecrets.create({ session: { model: "gpt-realtime-1.5" } });`
`console.log(secret.value);` | Model doc:
SDK release: | Documented in model docs + SDK release notes; also reflected in SDK types. citeturn6search2turn46view0turn23view0 | Model ID inclusion validated by diffing SDK type unions (v6.23.0 vs v6.24.0); not executed. citeturn23view0turn24view0 | | Realtime model ID additions (`gpt-audio-1.5`) | v6.24.0 | Adds `gpt-audio-1.5` to the SDK’s Realtime session `model` union types. citeturn46view0turn23view0turn24view0 | Audio generation/transcription workflows that rely on Realtime session schemas in the SDK; use-case framing in OpenAI changelog is split across APIs. citeturn19view0turn6search5 | `import OpenAI from "openai";`
`const client = new OpenAI();`
`await client.realtime.clientSecrets.create({ session: { model: "gpt-audio-1.5" } });` | Model doc:
SDK release: | Model is documented; Realtime-vs-Chat-Completions availability is partially specified in the changelog text, while SDK types include it for Realtime sessions. citeturn6search5turn19view0turn23view0 | Type-level availability validated via tagged SDK types; not executed. citeturn23view0turn24view0 | | `phase` field for assistant messages in Responses | v6.25.0 | Adds `phase?: "commentary" | "final_answer" | null` to assistant message inputs/outputs for the Responses API. citeturn46view0turn26view0turn19view1 | Long-running tasks where you want to label intermediate assistant “preambles” vs the final answer; helps avoid early stopping and supports Codex patterns. citeturn16search1turn16search4 | `import OpenAI from "openai";`
`const client = new OpenAI();`
`await client.responses.create({`
`model:"gpt-5.4",`
`input:[{ type:"message", role:"assistant", phase:"commentary", content:"Working..." },{ role:"user", content:"Continue." }]`
`});` | `phase` overview (Codex guide):
SDK release: | Documented in OpenAI changelog + guides; present in SDK message types. citeturn19view1turn16search1turn26view0 | Field presence validated in v6.25.0 types and OpenAI docs; example not executed. citeturn26view0turn16search1 | | `gpt-5.4` model ID typed in SDK | v6.26.0 | Adds `gpt-5.4` to the SDK’s `ChatModel` type union (usable across Chat Completions and Responses). citeturn45view0turn29view0turn30view0turn19view2 | High-capability text+vision reasoning and agent workflows; default recommendation in OpenAI model guidance. citeturn15search0turn19view2turn36search6 | `import OpenAI from "openai";`
`const client = new OpenAI();`
`const r = await client.responses.create({ model:"gpt-5.4", input:"Summarize this." });`
`console.log(r.output_text);` | Model doc:
SDK release: | Documented in OpenAI model docs + API changelog; reflected in SDK type unions. citeturn15search0turn19view2turn29view0 | Model ID union validation performed by comparing tags; request example not executed. citeturn29view0turn30view0 | | Tool Search tool (`type:"tool_search"`) + deferred tool schemas (`defer_loading`) | v6.26.0 | Adds `ToolSearchTool` (`type:"tool_search"`) and `defer_loading?: boolean` across tools; includes tool-search call/output item shapes (`tool_search_call`, `tool_search_output`). citeturn45view0turn28view0turn28view1turn31view0turn17view0turn19view2 | Cutting token overhead and preserving cache behavior when exposing large tool surfaces; loading only relevant tool definitions at runtime. citeturn17view0turn19view2 | `import OpenAI from "openai";`
`const client = new OpenAI();`
`const resp = await client.responses.create({`
`model:"gpt-5.4",`
`tools:[{ type:"tool_search" },{ type:"function", name:"crm.lookup", defer_loading:true, parameters:{ type:"object", properties:{ id:{type:"string"} } }, strict:true }],`
`input:"Look up customer 123 and summarize."`
`});` | Tool Search guide:
SDK release: | Documented in OpenAI guides + changelog; SDK type surfaces match (`tool_search`, `defer_loading`, call/output items). citeturn17view0turn19view2turn28view0turn31view0 | Example validated against Tool Search guide concepts and SDK type definitions; not executed. citeturn17view0turn28view1turn31view0 | | GA Computer tool (`type:"computer"`) for Responses + upgrade path from preview | v6.26.0 | Introduces GA `computer` tool support (tool type `computer`, computer-call loops with batched `actions[]`), aligning with `gpt-5.4` computer use workflows. citeturn45view0turn32view2turn19view2turn28view4 | UI automation loops (agent controls a browser/VM via screenshots and returned actions); migration away from `computer-use-preview`. citeturn32view2turn32view3 | `import OpenAI from "openai";`
`const client = new OpenAI();`
`const resp = await client.responses.create({ model:"gpt-5.4", tools:[{ type:"computer" }], input:"Open the site and search penguin." });`
`// then handle resp.output computer_call + send computer_call_output screenshots` | Computer use guide:
SDK release: | Documented in OpenAI guide + changelog; SDK includes `computer` tool type. citeturn32view2turn19view2turn28view4 | Example pattern validated against the official Computer use guide; loop scaffolding not executed. citeturn32view2turn32view3 | | `prompt_cache_key` parameter support in Responses | v6.26.0 | Adds/repairs support for `prompt_cache_key` in Responses request parameters (called out explicitly in release notes). citeturn46view0turn39view2 | Prompt caching strategies where you want stable cache bucketing keys across similar requests/users or shared system prompts. citeturn39view2 | `import OpenAI from "openai";`
`const client = new OpenAI();`
`await client.responses.create({ model:"gpt-5.4", prompt_cache_key:"acct:42:sys:v1", input:"Do the task." });` | SDK release (mentions `prompt_cache_key`): | Mentioned in SDK release notes under bug fixes; also referenced in Responses type docs. citeturn46view0turn39view2 | Validated via release note text and presence in SDK type documentation; not executed; exact caching semantics beyond docs are unspecified. citeturn46view0turn39view2 | | Computer tool type/class naming stabilization (GA vs preview) | v6.27.0 | Renames the GA vs preview tool interfaces so the GA `computer` tool maps to `ComputerTool`, while `computer_use_preview` maps to `ComputerUsePreviewTool`. citeturn45view0turn10view0turn28view4turn28view3 | Reduces confusion when migrating from preview: developers can map tool objects to clearer TypeScript shapes in GA codebases. citeturn32view3turn45view0 | `// GA`
`tools:[{ type:"computer" }]`
`// Preview`
`tools:[{ type:"computer_use_preview", display_width:1024, display_height:768, environment:"browser" }]` | SDK release:
Computer migration guide: | Type rename is documented in SDK release notes; preview-to-GA migration is documented in platform guide. citeturn45view0turn32view3 | Type mapping validated by comparing v6.26.0 vs v6.27.0 tagged SDK types; not executed. citeturn10view0turn28view4 | ## Timeline and feature trajectory ```mermaid timeline title openai-node v6.23.0–v6.27.0 releases and major feature additions 2026-02-23 : v6.23.0 — Responses WebSocket support in SDK 2026-02-24 : v6.24.0 — Realtime models: gpt-realtime-1.5, gpt-audio-1.5 (type-level) 2026-02-24 : v6.25.0 — phase field for assistant messages in Responses 2026-03-05 : v6.26.0 — gpt-5.4 typed support; tool_search tool + defer_loading; GA computer tool; prompt_cache_key param 2026-03-05 : v6.27.0 — Computer tool naming cleanup (GA vs preview types/classes) ``` Read as a progression, the SDK is evolving toward “agent infrastructure” primitives: - **Transport**: WebSocket Mode for the Responses API supports persistent links for multi-step tool loops. citeturn33view0turn46view0 - **Conversation discipline**: `phase` formalizes what is intermediate vs final in assistant message history, which matters when you carry state forward (especially in stateless or tool-heavy flows). citeturn26view0turn16search4 - **Tool scaling**: Tool Search provides a mechanism to keep large tool surfaces out of the context window until needed, preserving cache behavior and controlling token usage. citeturn17view0turn31view0turn19view2 - **Embodied interaction**: The GA `computer` tool pushes the core agent loop beyond API calls into UI actions driven by screenshots and returned action batches. citeturn32view2turn19view2 - **Model alignment**: releases introduce or type-enable the models that anchor these workflows (`gpt-realtime-1.5`, `gpt-5.4`). citeturn19view0turn19view2turn15search0turn23view0 ## Verification and caveats All example snippets in this report were validated by **reading official docs and tagged SDK source/types**, and by cross-checking against the official release notes for the relevant versions. No runtime execution was performed. citeturn45view0turn46view0turn11view0turn23view0turn26view0turn31view0turn32view2turn33view0 A few details remain explicitly **unspecified** based on what was visible in primary sources: - The TypeScript API reference page for `client.responses.connect()` shows the signature as `client.responses.connect(RequestOptions options?): void`, while the code snippet uses `await client.responses.connect();`. The return value/handle type is therefore treated as unspecified here. citeturn44view0 - The platform changelog text explicitly associates `gpt-audio-1.5` with Chat Completions on Feb 23, 2026, while the SDK includes it in the Realtime session model union in v6.24.0. This report treats that as “SDK type-level support,” while the precise platform availability across endpoints should be confirmed per up-to-date model availability in your account. citeturn19view0turn23view0 - Prompt caching semantics for `prompt_cache_key` are referenced by the SDK change note and type docs; any deeper operational guidance (cache hit-rate strategies, key design patterns) should follow the platform’s prompt caching documentation and testing under your workload. citeturn46view0turn39view2