The OpenAI API changelog for the period spanning December 2023 to September 2026 documents a rapid expansion of the platform, marked by the release of multiple frontier model families, new agentic tools, and granular enterprise controls. The most recent entries in September 2026 include a critical bug fix for image encoding in the GPT-6 Sol and Luna reasoning models, and the release of the models themselves, which accept text and image inputs and are priced at $2 per million input tokens for Sol and $0.10 for Luna.
The evolution of the GPT model family
The timeline shows a clear acceleration from the GPT-4 series through to GPT-6. GPT-4o launched in May 2024, followed by the o-series reasoning models in September 2024. The GPT-5 family arrived in August 2025, introducing reasoning effort levels and a 1M token context window. By April 2026, GPT-5.5 had added built-in computer use, hosted shell, and apply patch tools. GPT-6 Astra, released in September 2026, is described as the "most capable model, built for the hardest end-to-end work," combining reasoning, coding, computer use, research, and document creation. It does not support custom temperature or top_p values.
Agentic infrastructure and the Responses API
The Responses API became the central hub for agentic workflows following its March 2025 launch. It now supports durable sessions through the Agents API (public beta as of September 2026), which handles session orchestration, context compaction, and recovery. Developers can run agents in OpenAI-hosted sandboxes or connect their own infrastructure. The API also gained async tool calling, mid-turn steering over WebSockets, and the ability to change reasoning effort mid-conversation. These controls target long-running work with models like GPT-6 Astra.
Enterprise governance and security controls
Administrators received a series of new controls throughout 2026. In September, API key creation governance allowed organizations to restrict new keys to service accounts only or disable creation entirely. Expiration dates for project API keys became configurable, with enforcement options for maximum key lifetime. Mutual TLS and X.509 workload identity federation reached general availability in August. Hard spend limits, IP allowlists, and regional processing selection via prefixed domains gave teams finer control over cost and compliance.
Multimodal and real-time capabilities
Voice and image generation saw steady updates. GPT-Live 1 reached general availability in September 2026, enabling full-duplex voice conversations at $0.05 per minute. The transcription stack shifted toward GPT Transcribe and GPT Live Transcribe, with older models like Whisper-1 and GPT-4o-transcribe set for deprecation by February 2027. On the image side, GPT Image 2.5 Sunburst and Flare launched with new xhigh and max quality settings, while transparent backgrounds entered preview for GPT Image 2. Sora 2 and Sora 2 Pro expanded video generation to 20-second clips with 1080p output.
Why this matters for product development, IT, and research teams
For professionals building on the OpenAI API, the changelog reveals a platform shifting from single-model calls to managed agent infrastructure. The introduction of async tool calling, mid-turn steering, and server-side compaction means teams can now offload orchestration logic they previously had to build themselves. The deprecation of the Assistants API in August 2026 and the push toward Responses and Conversations APIs signals a clear migration path. Developers and product leads should audit their current integrations-particularly around image inputs, where the September 25 encoding bug may have affected visual task results-and plan for the February 2027 shutdown of older transcription models. Researchers in life sciences gained access to GPT-Rosalind through a trusted-access program, with billing starting in October 2026 at $5 per million input tokens.
Your membership also unlocks:
