Changelog | OpenAI API

Added API key creation governance controls at the organization and project levels. Administrators can allow only service-account keys, allow only user-owned project keys, or disable all new API key creation. Organization restrictions take precedence over project settings, and existing API keys are unaffected. See production best practices for details.

You can now set expiration dates when creating project API keys. Administrators can also enforce a maximum key lifetime at the organization or project level in Platform settings, requiring newly created keys to expire within the configured limit. See production best practices for guidance on key expiration and rotation.

Released the Agents API in public beta. Build agents with a managed Codex harness while OpenAI handles session orchestration, context compaction, and recovery.

Use durable sessions to continue work across turns, stream progress, and connect your own tools and MCP servers. Run agents in OpenAI-hosted sandboxes or connect a sandbox from your own infrastructure or a supported provider.

Start with the Agents API quickstart.

Prompt Cache Diagnostics is now generally available in the Responses API for GPT-5.6 and later supported models.

Compare cache reuse against a previous response, identify reasons for cache misses, and follow troubleshooting guidance to improve cache reuse.

Added new controls for long-running work with GPT-6 Astra in the Responses API:

• Async tool calling: Let the model continue working while your application runs function or custom tools, then return results as they become available.
• Mid-turn steering: Send additional instructions while a response is in progress over WebSockets, so the model can incorporate corrections or changing requirements.
• Change reasoning effort mid-conversation: Increase effort for difficult work or reduce it for routine follow-ups while preserving the cached prompt prefix.

Updated API errors so applications can distinguish traffic that increases too quickly from temporary model overload.

Traffic that increases too quickly can return a 429 error with the slow_down code. Temporary model overload returns a 503 error with the server_is_overloaded code. Both responses may include Retry-After. When the header is present, wait at least as long as it specifies before retrying. If it's missing, use exponential backoff. See the error codes guide and rate limits guide.

Connections to api.openai.com can now use IPv6.

API customers can now select regional processing for an individual request by using a prefixed domain with an API key from a project having Global geography. Existing eligibility, data retention control, endpoint, and model support requirements continue to apply. Learn more in the data controls guide.

GPT-5.6 Sol now costs $4 per million input tokens and $20 per million output tokens, representing 20% lower input pricing and 33% lower output pricing. GPT-5.6 Sol’s promotional pricing is available at least through November 21, 2026. See pricing details.

Released the Prompt Caching dashboard on the OpenAI API platform. Track your cache hit rate over time, cache reads per write, and the breakdown of cache-read, cache-write, and uncached tokens to understand your caching efficiency and identify opportunities to improve. Filter metrics by model and service tier.

Transparent backgrounds are now available in preview for gpt-image-2 and gpt-image-2-2026-04-21 in the Images API and the Responses API image generation tool. Set background to transparent and use png or webp output; jpeg does not support transparent backgrounds. Learn more in the image generation guide.

Announced Ultrafast mode, a new API service tier for GPT-5.6 Sol that runs up to 14x faster than Standard processing. Available in limited preview to select customers. Sign up to receive updates on Ultrafast mode here.

Updated the chat-latest snapshot, which points to the latest model available in ChatGPT for Plus and Pro users. We recommend leveraging GPT-5.6 Sol for production API usage, but feel free to use this model to test the latest improvements for chat use cases. The underlying model snapshot will be regularly updated. Read more here.

Fast mode now supports long-context requests for GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna. As of today, long-context prompts exceeding 272K tokens can run in Fast mode, delivering speeds up to 2.5× faster than the Standard tier. See pricing details.

Released the official OpenAI Terraform provider for managing OpenAI API Platform resources as infrastructure as code.

Provision and manage projects, users, groups, roles, access assignments, service accounts, certificates, invitations, and project-level rate limits. Use standard Terraform workflows to review and apply changes, import existing resources, and detect and reconcile configuration drift. Install the provider from the Terraform Registry.

Added hard spend limits for organizations and projects on the OpenAI API platform. Set a monthly cap that causes affected API requests to return a 429 error when tracked spend reaches the limit. Use spend alerts for notification before traffic is interrupted. Read more in the spend limits guide.

Released GPT-Realtime-2.1, an updated realtime reasoning model with improved alphanumeric recognition, silence and noise handling, and interruption behavior. Also released GPT-Realtime-2.1 mini, a faster, lower-cost distilled reasoning model for realtime voice applications.

Updated the chat-latest snapshot, which points to the latest Instant model currently used in ChatGPT. We recommend leveraging GPT-5.5 for production API usage, but feel free to use this model to test the latest improvements for chat use cases. The underlying model snapshot will be regularly updated. Read more here.

Released the Safety Usage Dashboard on the OpenAI API platform. The Safety dashboard shows blocked Responses requests based on safety_identifier values sent on requests to identify end users. Visit the Safety dashboard.

Web search can now return image results alongside regular text results. Use image search when your application needs current or web-grounded visuals, such as product photos, landmarks, places, events, or visual references. Read more in the web search guide.

Released a redesigned navigation for the OpenAI API platform, visit here.

Announced the deprecation of reusable prompt objects, the Evals platform, and Agent Builder. See the deprecations page for shutdown timelines and migration guidance.

Starting June 2, 2026, eligible container sessions will be billed per minute with a 5-minute minimum, instead of being billed at the full 20-minute session rate. The underlying per-minute rate will remain the same.

This update is intended to make billing more granular for shorter sessions and will lower effective cost for customers.

You can find current built-in tool pricing in our API pricing docs.

OpenAI models are now available in Amazon Bedrock through an OpenAI-compatible Responses API endpoint. Supported models and features vary by AWS Region. Learn more.

For organizations without ZDR enabled, prompt_cache_retention now defaults to 24h instead of in_memory, enabling extended prompt caching by default. Learn more.

Released chat-latest snapshot which points to the latest Instant model currently used in ChatGPT. We recommend leveraging GPT-5.5 for production API usage, but feel free to use this model to test the latest improvements for chat use cases. The underlying model snapshot will be regularly updated. Read more here.

Released workload identity federation. Trusted workloads can exchange externally issued identity tokens for short-lived OpenAI access tokens without storing long-lived API keys.

Added new Admin API capabilities for managing spend alerts, model allowlists, data retention settings, and hosted tool permissions, plus querying granular billing line items.

Released Secure MCP Tunnel for enterprise customers. Secure MCP Tunnel lets supported OpenAI products including ChatGPT web, Codex, Responses API, and AgentKit connect to private or on-prem MCP servers through a customer-hosted tunnel-client without exposing those servers to the public internet.

You can now manage multiple IP allowlists and apply each one at the project level or across the whole organization. To configure them, go to Settings > Security > IP allowlist.

Added return_token_budget for the Responses API web search tool. Use it to opt in to longer GPT-5+ reasoning web search runs for high-effort research and evaluation workloads.

The updated Agents SDK is now available in TypeScript, with support for sandbox agents and an open-source harness built in. Learn more here.

Released chat-latest snapshot which points to the latest Instant model currently used in ChatGPT. We recommend leveraging GPT-5.5 for production API usage, but feel free to use this model to test our latest improvements for chat use cases. The underlying model snapshot will be regularly updated. Read more here.

Admin APIs are now supported in the OpenAI SDKs for Node, Python, Go, Ruby, and Java. See the Admin APIs guide for setup instructions and examples.

Released GPT Image 2, a state-of-the-art image generation model for image generation and editing. GPT Image 2 supports flexible image sizes, high-fidelity image inputs, token-based image pricing, and Batch API support with a 50% discount.

Updated the Agents SDK with new capabilities, including:

• running agents in controlled sandboxes;
• inspecting and customizing the open-source harness; and
• controlling when memories are created and where they're stored.

Updated the gpt-5.3-chat-latest slug to point to the latest model currently used in ChatGPT.

Updated our image encoder to fix a small bug with input_image inputs in GPT-5.4. Some image understanding use cases may now see improved quality. No action is required.

Expanded the Sora API with reusable character references, longer generations up to 20 seconds, 1080p output for sora-2-pro, video extensions, and Batch API support for POST /v1/videos. 1080p generations on sora-2-pro are billed at $0.70 per second. Learn more here.

Added POST /v1/videos/edits for editing existing videos. This will replace POST /v1/videos/{video_id}/remix, which will be deprecated in 6 months. Learn more here.

Released gpt-5.3-chat-latest to the Chat Completions and Responses API. This model points to the GPT-5.3 Instant snapshot currently used in ChatGPT. Read more here.

Expanded input_file support in the Responses API to accept more document, presentation, spreadsheet, code, and text file types. Learn more here.

Released phase to the Responses API. It labels an assistant message as intermediate commentary (commentary) or the final answer (final_answer). Read more here.

Released gpt-5.3-codex to the Responses API. Read more here.

Launched WebSocket mode for the Responses API. Learn more here.

Batch API is now supported for GPT Image models: gpt-image-1.5, chatgpt-image-latest, gpt-image-1, and gpt-image-1-mini.

Updated the gpt-5.2-chat-latest slug to point to the latest model currently used in ChatGPT.

Launched support for Skills in the Responses API. We support Skills across both local execution and hosted container-based execution.

Launched a new Hosted Shell tool, as well as support for networking in containers.

Added support for application/json requests on /v1/images/edits for GPT image models. JSON requests use images (and optional mask) with image_url or file_id references instead of multipart uploads.

We have optimized our inference stack for API customers and GPT-5.2 and GPT-5.2-Codex now run ~40% faster. Model and model weights are unchanged.

Announced Open Responses: an open-source spec for building multi-provider, interoperable LLM interfaces built on top of the original OpenAI Responses API.

Released gpt-5.2-codex to the Responses API. GPT-5.2-Codex is a version of GPT-5.2 optimized for agentic coding tasks in Codex or similar environments. Read more here.

Added dedicated SIP IP ranges for Realtime API. sip.api.openai.com does GeoIP routing, and will direct SIP traffic to the closest region. Learn more.

Updated the gpt-realtime-mini and gpt-audio-mini slugs to point to the 2025-12-15 snapshots. If you need the previous model snapshots, use gpt-realtime-mini-2025-10-06 and gpt-audio-mini-2025-10-06.

Updated the sora-2 slug to point to sora-2-2025-12-08. If you need the previous model snapshot, use sora-2-2025-10-06.

Updated the gpt-4o-mini-tts and gpt-4o-mini-transcribe slugs to point to the 2025-12-15 snapshots. If you need the previous model snapshots, use gpt-4o-mini-tts-2025-03-20 and gpt-4o-mini-transcribe-2025-03-20. We currently recomend using gpt-4o-mini-transcribe over gpt-4o-transcribe for the best results.

Fixed an issue where gpt-image-1.5 and chatgpt-image-latest were incorrectly using high fidelity for image edits through /v1/images/edits, even when fidelity was explicitly set to low (the default).

Added gpt-image-1.5 and chatgpt-image-latest to the Responses API image generation tool.

Released client-side compaction. For long-running conversations with the Responses API, you can use the /responses/compact endpoint to shrink the context you send with each turn.

Released gpt-5.1-codex-max to the Responses API. GPT-5.1-Codex is our most intelligent coding model optimized for long-horizon, agentic coding tasks. Read more here.

Added support for DTMF key presses in the Realtime API. You can now receive DTMF events while using a Realtime sideband connection. See docs here for more information.

Released gpt-5.1-codex and gpt-5.1-codex-mini to the Responses API. GPT-5.1-Codex is a version of GPT-5.1 optimized for agentic coding tasks in Codex or similar environments. Read more here.

Released extended prompt cache retention. Extended prompt cache retention keeps cached prefixes active for longer, up to a maximum of 24 hours. Extended Prompt Caching works by offloading the key/value tensors to GPU-local storage when memory is full, significantly increasing the storage capacity available for caching.

gpt-oss-safeguard-120b and gpt-oss-safeguard-20b are safety reasoning models built-upon gpt-oss. Read more here.

Released Enterprise Key Management (EKM). Enterprise Key Management (EKM) allows you to encrypt your customer content at OpenAI using keys managed by your own external Key Management System (KMS).

Released IP allowlist. IP allowlisting restricts API access to only the IP addresses or ranges you specify.

Launched special-purpose model gpt-5-codex, built and optimized for use with the Codex CLI.

Added support for connectors to the Responses API. Connectors are OpenAI-maintained MCP wrappers for popular services like Google apps, Dropbox, and more that can be used to give model read access to data stored in those services.

Released the Conversations API, which allows you to create and manage long-running conversations with the Responses API. See the migration guide to see a side-by-side comparison and learn how to migrate from an Assistants API integration to Responses and Conversations.

Launched support for Priority processing. Priority processing delivers significantly lower and more consistent latency compared to Standard processing while keeping pay-as-you-go flexibility.

New reusable prompts are now available in the dashboard and Responses API. Via API, you can now reference templates created in the dashboard via the prompt parameter (with a prompt id, optional version) and supply dynamic variables that can include strings, images, or file inputs. Reusable prompts are not available in Chat Completions. Learn more.

Released o3-pro, a version of the o3 reasoning model that uses more compute to answer hard problems with better reasoning and consistency. Prices for the o3 model have also been reduced for all API requests, including batch and flex processing.

Added fine-tuning support with direct preference optimization for the models gpt-4.1-2025-04-14, gpt-4.1-mini-2025-04-14, and gpt-4.1-nano-2025-04-14.

Feature

v1/chat/completions

v1/realtime

Added support for using strict mode for tool schemas when using parallel tool calling with non-fine-tuned models. Added new schema features, including string validation for email and other patterns and specifying ranges for numbers and arrays.

Added gpt-4.1, gpt-4.1-mini, and gpt-4.1-nano models to the API. These new models feature improved instruction following, coding, and a larger context window (up to 1M tokens). gpt-4.1 and gpt-4.1-mini are available for supervised fine-tuning. Announced deprecation of gpt-4.5-preview.

Added gpt-4o-mini-tts, gpt-4o-transcribe, gpt-4o-mini-transcribe, and whisper-1 models to the Audio API.

Released o1-pro, a version of the o1 reasoning model that uses more compute to answer hard problems with better reasoning and consistency.

Added metadata field support to fine-tuning jobs.

Released a research preview of GPT-4.5—our largest and most capable chat model yet. GPT-4.5's high "EQ" and understanding of user intent make it better at creative tasks and agentic planning.

Launched the API Usage Dashboard Update. This update addresses requests for additional data filters, such as project selection, date picker, and fine-grained intervals. There’s also better support for viewing usage across different products and service tiers.

Introducing data residency in Europe. Read more here.

Launched o3-mini, a new small reasoning model that is optimized for science, math, and coding tasks.

Expanded access to o1 model. The o1 series of models are trained with reinforcement learning to perform complex reasoning.

Launched Admin API Key Rotations, enabling customers to programmatically rotate their admin api keys.

Updated Admin API Invites, enabling customers to programmatically invite users to projects at the same time they are invited to organizations.

Launched Usage API, enabling customers to programmatically query activities and spending across OpenAI APIs.

Released Predicted Outputs, which greatly reduces latency for model responses where much of the response is known ahead of time. This is most common when regenerating the content of documents and code files with only minor changes.

Released new omni-moderation-latest moderation model, which supports both images and text (for some categories), supports two new text-only harm categories, and has more accurate scores.

Released o1-preview and o1-mini, new large language models trained with reinforcement learning to perform complex reasoning tasks.

Launched Admin and Audit Log APIs, allowing customers to programmatically administer their organization and monitor changes using the audit logs. Audit logging must be enabled within settings.

Launched self-serve SSO configuration, allowing Enterprise customers on custom and unlimited billing to set up authentication against their desired IDP.

Released GPT-4o mini, our affordable an intelligent small model for fast, lightweight tasks.

Released Uploads to upload large files in multiple parts.

Released GPT-4o in the API. GPT-4o is our fastest and most affordable flagship model.

Added a new endpoint to delete a message from a thread in the Assistants API.

Introduced project based hierarchy for organizing work by projects, including the ability to create API keys and manage rate and cost limits on a per-project basis (cost limits available only for Enterprise customers).

Added support for streaming in the Assistants API

Released embedding V3 models and an updated GPT-4 Turbo preview

Added dimensions parameter to the Embeddings API