LLM/AI Changelog

Today, we’re releasing OpenAIo3ando4-mini,the latest in our o-series of models trained to think for longer before responding. These are the smartest models we’ve released to date, representing a step change in ChatGPT's capabilities for everyone from curious users to advanced researchers. For the first time, our reasoning models can agentically use and combine every tool within ChatGPT—this includes searching the web, analyzing uploaded files and other data with Python, reasoning deeply about visual inputs, and even generating images. Critically, these models are trained to reason about when and how to use tools to produce detailed and thoughtful answers in the right output formats, typically in under a minute, to solve more complex problems. This allows them to tackle multi-faceted questions more effectively, a step toward a more agentic ChatGPT that can independently execute tasks on your behalf. The combined power of state-of-the-art reasoning with full tool access translates into significantly stronger performance across academic benchmarks and real-world tasks, setting a new standard in both intelligence and usefulness.

AI Admin role has permissions to manage agents
[Web]
As an AI Admin, you manage agents with various capabilities including creating and overseeing Copilot connections that index data in Graph and perform actions, controlling which makers can build these connections and regulating the data sources used, maintaining observability over all connections, approving or denying agents, pre-installing them without requiring consent, and viewing them in Microsoft 365 admin center integrated apps page.

All images you create with ChatGPT are now automatically saved to a newLibraryin the sidebar, giving you one place to browse, revisit, and reuse your work without digging through past conversations. The Library is rolling out today on Web, iOS, and Android for Free, Plus, and Pro users (Enterprise / Edu support coming soon). For now, it displays images generated with 4o Image Generation while we backfill older creations, and you can remove an image by deleting the conversation where it was made.

Sunsetting GPT‑4 in ChatGPT
Effective  April  30,  2025, GPT‑4 will be retired from ChatGPT and fully replaced by GPT‑4o.
GPT‑4o is our newer, natively multimodal model. In head‑to‑head evaluations it consistently surpasses GPT‑4 in writing, coding, STEM, and more.
Recent upgrades have further improved GPT‑4o’s instruction following, problem solving, and conversational flow, making it a natural successor to GPT‑4.
GPT-4 will still be available in the API.
GPT‑4 marked a pivotal moment in ChatGPT’s evolution. We’re grateful for the breakthroughs it enabled and for the feedback that helped shape its successor. GPT‑4o builds on that foundation to deliver even greater capability, consistency, and creativity.

**Model updates:**

- Released `veo-2.0-generate-001`, a generally available (GA) text- and image-to-video model, capable of generating detailed and artistically nuanced videos. To learn more, see the [Veo docs](https://ai.google.dev/gemini-api/docs/video).
- Released `gemini-2.0-flash-live-001`, a public preview version of the
[Live API](https://ai.google.dev/gemini-api/docs/live) model with billing enabled.

- **Enhanced Session Management and Reliability**

- **Session Resumption:** Keep sessions alive across temporary network disruptions. The API now supports server-side session state storage (for up to 24 hours) and provides handles (session_resumption) to reconnect and resume where you left off.
- **Longer Sessions via Context Compression:** Enable extended interactions beyond previous time limits. Configure context window compression with a sliding window mechanism to automatically manage context length, preventing abrupt terminations due to context limits.
- **Graceful Disconnect Notification:** Receive a `GoAway` server message indicating when a connection is about to close, allowing for graceful handling before termination.
- **More Control over Interaction Dynamics**

- **Configurable Voice Activity Detection (VAD):** Choose sensitivity
levels or disable automatic VAD entirely and use new client events
(`activityStart`, `activityEnd`) for manual turn control.

- **Configurable Interruption Handling:** Decide whether user input
should interrupt the model's response.

- **Configurable Turn Coverage:** Choose whether the API processes all
audio and video input continuously or only captures it when the end-user
is detected speaking.

- **Configurable Media Resolution:** Optimize for quality or token usage
by selecting the resolution for input media.

- **Richer Output and Features**

- **Expanded Voice \& Language Options:** Choose from two new voices and
30 new languages for audio output. The output l

Gain insights with Copilot and Python in Excel on Web
[Web]
Using everyday language, ask Copilot to perform advanced analytics like machine learning and predictive forecasting, which would usually take hours or require special skills. Copilot writes Python code and inserts it on the grid, providing deeper insights and stunning/ visuals. Available in multiple languages on Excel for the Web.

Enhanced SharePoint URL support in Copilot Studio
[Web]
Previously, only SharePoint site URLs could be used as knowledge for Copilot Studio agents. Users can now use SharePoint file, folder, and site URLs, including links from the Share button or browser address bar; once the URLs are added, the agents will be grounded to the content within the URL. Additionally, more granular error messages help with transparency and managing expectations.
Learn more
.

Public website scoping for agents
[Web]
Makers can now scope agents' knowledge sources to specific websites, enhancing the precision of web searches. This capability, previously available in custom agents from Microsoft Copilot Studio, is now extended to agents in the Microsoft 365 context.
Learn more
.

Gain insights with Copilot and Python in Excel on Web
[Web]
Using everyday language, ask Copilot to perform advanced analytics like machine learning and predictive forecasting, which would usually take hours or require special skills. Copilot writes Python code and inserts it on the grid, providing deeper insights and stunning/ visuals. Available in multiple languages on Excel for the Web.
Learn more
.

Collect user feedback in Copilot Studio agent builder
[Web]
Makers can also now submit feedback—compliments, problems, or suggestions—directly to the product team from inside the Microsoft 365 agent builder. This can be done at any stage of the authoring process, including comments and optional metadata for troubleshooting. This feature allows quick issue reporting without disrupting workflow and enables the product team to address feedback efficiently.
Learn more
.

Collect user feedback in Copilot Studio agent builder
[Web]
Makers can also now submit feedback-compliments, problems, or suggestions-directly to the product team from inside the Microsoft 365 agent builder. This can be done at any stage of the authoring process, including comments and optional metadata for troubleshooting. This feature allows quick issue reporting without disrupting workflow and enables the product team to address feedback efficiently.
Learn more
.

Our 2.5 Pro (experimental) model is now available to all Gemini users, with CanvasWhat:Today we’re expanding access to our most intelligent AI model, 2.5 Pro (experimental), to all Gemini users. This state-of-the-art model has thinking capabilities natively built in, with exceptional performance in coding, math, image understanding and more.Canvas, our new interactive space that makes it easy to create, refine, and share your work, is now available to try with 2.5 Pro (experimental). With 2.5 Pro’s improved coding capabilities, coupled with Canvas, you can quickly create compelling web apps or generate code for immediate use.Gemini users will be able to try 2.5 Pro (experimental) with rate limits, and Gemini Advanced users will continue to have expanded access and a significantly larger context window.Being an experimental model, it can have unexpected behaviors and may make mistakes.Why:We want to bring the best model in the world to all Gemini users. Your feedback helps us improve these models over time and learning from experimental launches informs how we release models more widely.

Our 2.5 Pro (experimental) model is now available to all Gemini users, with CanvasWhat:Today, we're expanding access to our most intelligent AI model, 2.5 Pro (experimental), to all Gemini users. This state-of-the-art model has thinking capabilities natively built in, with exceptional performance in coding, maths, image understanding and more.Canvas, our new interactive space that makes it easy to create, refine and share your work, is now available to try with 2.5 Pro (experimental). With 2.5 Pro's improved coding capabilities, coupled with Canvas, you can quickly create compelling web apps or generate code for immediate use.Gemini users will be able to try 2.5 Pro (experimental) with rate limits, and Gemini Advanced users will continue to have expanded access and a significantly larger context window.Being an experimental model, 2.5 Pro may behave unexpectedly or make mistakes.Why: We want to bring the best model in the world to all Gemini users. Your feedback helps us improve these models over time, and learning from experimental launches informs how we release models more widely.

We’ve made improvements to GPT-4o—it now feels more intuitive, creative, and collaborative, with enhanced instruction-following, smarter coding capabilities, and a clearer communication style.
Smarter problem-solving in STEM and coding:
GPT-4o has further improved its capability to tackle complex technical and coding problems. It now generates cleaner, simpler frontend code, more accurately thinks through existing code to identify necessary changes, and consistently produces coding outputs that successfully compile and run, streamlining your coding workflows.
Enhanced instruction-following and formatting accuracy:
GPT-4o is now more adept at following detailed instructions, especially for prompts containing multiple or complex requests. It improves on generating outputs according to the format requested and achieves higher accuracy in classification tasks.
“Fuzzy” improvements:
Early testers say that the model seems to better understand the implied intent behind their prompts, especially when it comes to creative and collaborative tasks. It’s also slightly more concise and clear, using fewer markdown hierarchies and emojis for responses that are easier to read, less cluttered, and more focused. We're curious to see if our users also find this to be the case.
This model is now available in ChatGPT and in the API as the newest snapshot of chatgpt-4o-latest. We plan to bring these improvements to a dated model in the API in the coming weeks.

Priority access with Gemini Advanced: Try our latest 2.5 Pro (experimental) modelWhat:Today we’re introducing Gemini 2.5, our most intelligent AI model. Our first 2.5 release is a chat optimized version of Gemini-2.5-Pro-Exp-03-25, which is state-of-the-art on a wide range of benchmarks and debuts at #1 on LMArena by a significant margin. This model also has thinking capabilities natively built in, with improved performance in complex tasks like coding, math, and image understanding.2.5 Pro (experimental) is now rolling out to the Gemini web and mobile app and is available to qualifying Google Workspace business and education plans.This experimental model is meant to be an early preview and can have unexpected behaviors and may make mistakes.Why:We believe in rapid iteration and bringing the best of Gemini to the world, and we want to give Gemini Advanced subscribers priority access to our latest AI innovations. Your feedback helps us improve these models over time and learning from experimental launches informs how we release models more widely.

Priority access with Gemini Advanced: Try our latest 2.5 Pro (experimental) modelWhat:Today, we're introducing Gemini 2.5, our most intelligent AI model. Our first 2.5 release is a chat-optimised version of Gemini-2.5-Pro-Exp-03-25, which is state of the art on a wide range of benchmarks and debuts at #1 on LMArena by a significant margin. This model also has thinking capabilities natively built in, with improved performance in complex tasks like coding, maths and image understanding.2.5 Pro (experimental) is now rolling out to the Gemini web and mobile app and is available to qualifying Google Workspace business and education plans.This experimental model is meant to be an early preview; it may behave unexpectedly or make mistakes.Why: We believe in rapid iteration and bringing the best of Gemini to the world, and we want to give Gemini Advanced subscribers priority access to our latest AI innovations. Your feedback helps us improve these models over time, and learning from experimental launches informs how we release models more widely.