OpenAI · Advanced APIs · 2026

OpenAI 2026 — Responses, o-series, Realtime & Assistants sunset.

Fifteen deep-dive guides on the newer OpenAI APIs and the biggest migration deadline of 2026. Covers the Responses API replacing Chat Completions for new work, o-series reasoning-model quirks, Realtime API voice agents, and the Assistants API sunset scheduled for August 26, 2026. Every page has tested Python code, real error messages, and the fixes that solve 90% of cases.

15Deep-dive error pages

4API surfaces covered

Aug 26, 2026Assistants sunset

100%2026-current code

Install & verify

All code samples target the OpenAI Python SDK 1.60+ with Python 3.11+.

setup.shbash

pip install --upgrade openai
# Verify version 1.60+ for Responses API, Conversations, and Realtime helpers
python -c "import openai; print(openai.__version__)"

Responses API

The new default surface for OpenAI. Every new capability — built-in tools, MCP, computer use, deep research — ships here first. Migrate from Chat Completions and get familiar with the state model, tools, and streaming.

Error #136

Chat Completions → Responses API migration

Endpoint, output shape, state model — every field that changed and the 400s each triggers.

Error #137

previous_response_id vs Conversations state

Three ways to carry state on Responses; when each fits; the "cannot use both" 400.

Error #138

Built-in tools (web_search, file_search, code_interpreter)

vector_store_ids required, empty search results, container session expiry.

Error #139

Streaming events errors

Version required, event name migration, function-arg JSON reconstruction, incomplete responses.

o-series reasoning models

The reasoning-model family (o3, o4-mini, GPT-5 reasoning) has its own token accounting, parameter restrictions, and structured output quirks. These four pages cover the failures teams hit in the first week with reasoning models.

Error #140

max_completion_tokens exhausted by reasoning

Empty responses because reasoning tokens ate the whole budget. Sizing and monitoring.

Error #141

reasoning_effort parameter support

Which models accept it, o4-mini silent no-op, Chat Completions vs Responses shape.

Error #142

Unsupported parameters (temperature, top_p, penalties)

The o-series parameter allow-list and how to route around blocked sampling controls.

Error #143

Structured output with o-series

Strict-mode schema requirements, reasoning-budget sizing, refusal channel handling.

Realtime API — voice agents

Building voice agents on gpt-realtime. Transport selection is architecture-level; audio format, tool calling, and turn detection are configuration-level. Each has its own way to break.

Error #144

Realtime API transport (WebSocket vs WebRTC vs SIP)

Ephemeral client_secret pattern, WebSocket for backend, SIP for telephony, GA endpoint shapes.

Error #145

Audio format (PCM16, G.711) encoding errors

24kHz mono little-endian PCM16, G.711 for telephony, base64 chunking, WAV-header trap.

Error #146

Function calling in voice sessions

Four-step tool cycle (done → execute → output → create), duplicate-response bugs, voice-friendly output.

Error #147

Turn detection & interruption errors

server_vad vs semantic_vad, silence_duration_ms tuning, barge-in with client audio-clear.

Assistants API sunset — migration to Responses

The Assistants API sunsets August 26, 2026. These three migration guides cover threads → conversations, file_search wire changes, and rewriting run polling as streaming events. Start now — the deadline is firm.

Error #148

Threads → Conversations migration

The core migration: threads become conversations, assistants become prompts, runs become responses.

Error #149

file_search & vector stores migration

Vector stores are shared — config wrapper changes; code_interpreter files need explicit Containers.

Error #150

Run polling → streaming migration

No more run.status loops; submit_tool_outputs becomes function_call_output items; event shapes change.

Quick reference — where to look first

If you see...	Start here
`NotFoundError: Invalid URL (POST /v1/response)`	#136 · Chat Completions → Responses
`AttributeError: 'Response' has no attribute 'choices'`	#136 · Chat Completions → Responses
`cannot use both previous_response_id and conversation`	#137 · State management
`file_search requires vector_store_ids`	#138 · Built-in tools
Streaming events silent, wrong event names	#139 · Streaming events
o-series empty response, finish_reason=length	#140 · Reasoning exhausted budget
`Unsupported parameter 'temperature' for model o3`	#142 · Unsupported parameters
Realtime WebRTC connection fails, 401 handshake	#144 · Realtime transport
Realtime model responds slow/deep, garbled audio	#145 · Audio format
Realtime model silent after tool call	#146 · Function calling
Model cuts user off / doesn't allow interruption	#147 · Turn detection
Assistants API deprecation warning header	#148 · Migration overview