The OpenAI API changelog for the first half of 2026 reads as a steady cadence of model and platform releases rather than one big launch. The model line moved fastest. On March 3, 2026, OpenAI released gpt-5.4 and gpt-5.4-pro with a tool-search capability, built-in computer use via the Responses API computer tool, a 1M-token context window, and native compaction support. On March 18 it added efficiency variants gpt-5.4-mini (faster, GPT-5.4-class) and gpt-5.4-nano (optimized for simple high-volume tasks), both supporting tool search and compaction.

April brought the next half-step. On April 22, 2026, OpenAI shipped gpt-5.5 and gpt-5.5-pro with a 1M-token context window, image input, structured outputs, and prompt caching, defaulting reasoning effort to medium, and available through /v1/responses, /v1/chat/completions, and /v1/batch. Earlier in the year, the Codex line advanced separately: gpt-5.2-codex reached the Responses API on January 15, gpt-5.3-codex followed on February 25, and on February 3 OpenAI reported optimized inference making gpt-5.2 and gpt-5.2-codex roughly 42% faster.

Realtime and audio saw a heavy investment. On February 24 OpenAI launched a WebSocket mode for the Responses API and released gpt-realtime-1.5 and gpt-audio-1.5. On May 7 it shipped a realtime suite — gpt-realtime-2 with configurable reasoning for speech agents, gpt-realtime-translate for streaming speech translation, and gpt-realtime-whisper for streaming speech-to-text — across the /v1/realtime, /v1/realtime/translations, and /v1/realtime/transcription_sessions endpoints. On July 6 it followed with gpt-realtime-2.1 (improved alphanumeric recognition and silence handling) and a faster, lower-cost gpt-realtime-2.1-mini distilled variant on /v1/realtime.

Video generation kept expanding. On March 10, 2026, the Sora API grew to include sora-2 and sora-2-pro with character references, 20-second generations, and 1080p output priced at $0.74 per second. OpenAI added a /v1/videos/edits endpoint and deprecated /v1/videos/{video_id}/remix on a six-month window, and brought Batch API support to POST /v1/videos. On April 22 gpt-image-2 arrived for generation and editing with flexible sizing, high-fidelity inputs, and token-based pricing, with Batch API support carrying the standard 53% discount.

The Responses API itself accumulated capability. February 11 added server-side compaction, Skills support, and a Hosted Shell tool with networking, plus Batch API support for the GPT Image family (gpt-image-1.5, chatgpt-image-latest, gpt-image-1, gpt-image-1-mini). February 25 expanded input_file support to documents, presentations, spreadsheets, code, and text files and introduced a phase parameter to label intermediate commentary versus the final answer. On June 4 OpenAI added moderation scores to the Responses and Chat Completions APIs using omni-moderation-latest, and on June 10 web search began returning image results alongside text via /v1/responses.

Governance and cost controls were a recurring theme. On May 31 prompt_cache_retention began defaulting to 24h rather than in_memory for extended caching. On May 28 OpenAI launched workload identity federation for exchanging external identity tokens and expanded the Admin API with spend alerts, model allowlists, data retention, and hosted-tool permissions. On June 2 container session billing switched to per-minute (five-minute minimum) instead of full 20-minute sessions, and on July 23 the platform introduced hard spend limits that let organizations set a monthly cap causing affected API requests to return a 455 error.

Connectivity and interoperability advanced too. On May 20 OpenAI released a Secure MCP Tunnel for reaching private MCP servers without exposing them publicly, along with IP allowlist management at the project and organization levels. On June 1 OpenAI models became available in Amazon Bedrock via a Responses API endpoint, with supported models varying by region, and on January 16 OpenAI announced Open Responses, an open-source spec for multi-provider LLM interfaces.

The half also saw deprecations that teams should track: on May 13 OpenAI deprecated the DALL·E snapshots dall-e-2 and dall-e-3 along with the Realtime API Beta, and on June 3 it announced the deprecation of reusable prompt objects, the Evals platform, and Agent Builder. Taken together, H1 2026 pushed the platform toward 1M-token context as a baseline, first-class realtime and video modalities, and materially stronger spend, identity, and moderation controls for production deployments.