LLM/AI Changelog

Set ChatGPT as default search engine in Safari.Confirmation dialog for logging out (iOS, web).Improved Android conversation parsing performance.iOS performance fixes.Updated launch animations and composer design on Android.Whisper (ChatGPT's voice-to-text feature) now shows a text preview of dictated text after recording:

Priority access with Gemini Advanced: Ask Gemini to look back at past chats for more tailored helpWhat:Starting today, Gemini can now use your past chats to give you more helpful responses. Ask a question about something you've discussed, or have Gemini summarize a previous conversation. It will use the information from relevant chats to craft its response. This means no more starting over, and you can build on top of previous conversations or projects you already started.You’re in control: easily review, delete, or decide how long to keep your chat history. You can also turn off Gemini Apps Activity altogether by going toMy Activity. Gemini may indicate when it uses your past chats in Sources and related content. This experience is starting to roll out in English forGemini Advancedsubscribers via Google One AI Premium Plan.Why:Conversations change, ideas evolve, and sometimes we need to revisit what’s been said. Think of Gemini as your personal AI assistant, always there to pick up where you left off. It becomes your thought partner, evolving alongside you to unlock new levels of helpfulness, that’s unique to you.

Priority access with Gemini Advanced: Ask Gemini to look back at past chats for more tailored helpWhat: Starting today, Gemini can now use your past chats to give you more helpful responses. Ask a question about something that you've discussed or have Gemini summarise a previous conversation. It will use the information from relevant chats to craft its response. This means no more starting over, and you can build on top of previous conversations or projects that you've already started.You're in control: Easily review, delete or decide how long to keep your chat history. You can also turn off Gemini Apps activity altogether by going toMy Activity. Gemini may indicate when it uses your past chats in sources and related content. This experience is starting to roll out in English forGemini Advancedsubscribers via the Google One AI Premium plan.Why: Conversations change, ideas evolve and sometimes we need to revisit what's been said. Think of Gemini as your personal AI assistant, always there to pick up where you left off. It becomes your thought partner, evolving alongside you to unlock new levels of helpfulness, that's unique to you.

Users can now share a Canvas asset such as rendered React/HTML code, document, or code with another user, similar to how you share a conversation. You can do this from the Canvas toolbar when Canvas is open.
For more details about the canvas feature in ChatGPT, check out the following Help Center article!

**Model updates:**

- Released `gemini-2.0-flash-001`, a generally available (GA) version of [Gemini 2.0 Flash](https://ai.google.dev/gemini-api/docs/models/gemini#gemini-2.0-flash) that supports text-only output.
- Released `gemini-2.0-pro-exp-02-05`, an [experimental](https://ai.google.dev/gemini-api/docs/models/experimental-models) public preview version of Gemini 2.0 Pro.
- Released `gemini-2.0-flash-lite-preview-02-05`, an experimental public preview [model](https://ai.google.dev/gemini-api/docs/models/gemini#gemini-2.0-flash-lite) optimized for cost efficiency.

**API updates:**

- Added [file input and graph output](https://ai.google.dev/gemini-api/docs/code-execution#input-output) support to code execution.

**SDK updates:**

- Released the [Google Gen AI SDK for Python](https://googleapis.github.io/python-genai/) to general availability (GA).

Try our 2.0 Flash Thinking Experimental modelsWhat:Starting today, Gemini app users have access to our 2.0 Flash Thinking Experimental model. Built on the speed and performance of 2.0 Flash, this model is trained to break down prompts into a series of steps to strengthen its reasoning capabilities and deliver better responses. 2.0 Flash Thinking Experimental shows its thought process to users so you can see why it responded in a certain way, what its assumptions were, and trace the model's line of reasoning.Users can also try 2.0 Flash Thinking Experimental with apps, our thinking model designed to give you access to agentic capabilities through the use of connected apps such as YouTube, Maps, and Search for complex multi-step questions. These connected apps already make the Gemini app a uniquely helpful AI-powered assistant, and we’re exploring how new reasoning capabilities can combine with your apps to help you do even more.Both 2.0 Flash Thinking Experimental models are now rolling out to the Gemini web and mobile app. We’ll expand access to enterprise accounts in the coming weeksThese experimental models are meant to be an early preview and may make mistakes. Additionally these models won't be compatible with some Gemini features in their experimental state.Why:We believe in rapid iteration and bringing the best of Gemini to the world. 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 2.0 Pro Experimental modelWhat:Gemini Advanced users can now try 2.0 Pro Experimental, our leading model designed to be exceptional at complex tasks, providing better factuality and stronger performance for coding and math prompts.Whether you’re tackling advanced coding challenges, like generating a specific program from scratch, or solving mathematical problems, like developing complex statistical models or quantum algorithms, 2.0 Pro Experimental wi

Try our 2.0 Flash Thinking Experimental modelsWhat: Starting today, Gemini app users have access to our 2.0 Flash Thinking Experimental model. Built on the speed and performance of 2.0 Flash, this model is trained to break down prompts into a series of steps to strengthen its reasoning capabilities and deliver better responses. 2.0 Flash Thinking Experimental shows its thought process to users so that you can see why it responded in a certain way, what its assumptions were and trace the model's line of reasoning.Users can also try 2.0 Flash Thinking Experimental with apps, our thinking model designed to give you access to agentic capabilities through the use of connected apps such as YouTube, Maps and Search for complex multi-step questions. These connected apps already make the Gemini app a uniquely helpful AI-powered assistant, and we're exploring how new reasoning capabilities can combine with your apps to help you do even more.Both 2.0 Flash Thinking Experimental models are now rolling out to the Gemini web and mobile app. We'll expand access to enterprise accounts in the coming weeksThese experimental models are meant to be an early preview and may make mistakes. Additionally, these models won't be compatible with some Gemini features in their experimental state.Why: We believe in rapid iteration and bringing the best of Gemini to the world. 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 2.0 Pro Experimental modelWhat:Gemini Advanced users can now try 2.0 Pro Experimental, our leading model designed to be exceptional at complex tasks, providing better factuality and stronger performance for coding and math prompts.Whether you’re tackling advanced coding challenges, like generating a specific program from scratch, or solving mathematical problems, like developing complex statistical models or quantum algorithms, 2.0 Pro Experimen

**Model updates:**

* Released `gemini-2.0-flash-001`, a generally available (GA) version of
[Gemini 2.0 Flash](/docs/models) that supports text-only
output.

* Released `gemini-2.0-pro-exp-02-05`, an [experimental](/docs/models) public
preview version of Gemini 2.0 Pro.

* Released `gemini-2.0-flash-lite-preview-02-05`, an experimental public
preview [model](/docs/models) optimized for cost
efficiency.


**API updates:**

* Added [file input and graph output](/docs/code-execution#input-output)
support to code execution.

**SDK updates:**

* Released the [Google Gen AI SDK for Python](https://googleapis.github.io/python-genai/)
to general availability (GA).

Video creation in Copilot Visual Creator powered by Clipchamp
[Web]
Type your prompt and Clipchamp writes a bespoke script, sources high-quality footage, and assembles a video project with music, voiceover, text overlays, and transitions. Open your draft in the Clipchamp app to continue editing, exporting, and sharing your video.
Learn more
.

Get more Copilot value with Microsoft 365 Copilot
[Web]
Easily add full Copilot Chat capabilities including grounding conversations in work data and accessing Copilot in your favorite Microsoft 365 apps by purchasing or requesting a Microsoft 365 Copilot license directly in Copilot Chat.
Learn more
.

Speaker recognition and attribution in BYOD rooms with Copilot
[Windows, Mac]
Take advantage of speaker recognition and transcript attribution, unleashing new AI capabilities in any meeting space, whether or not it has a Teams Rooms system deployed. This feature identifies and attributes people in live transcripts, utilizing a unique voice profile for each participant enabling intelligent recaps and unlocking maximum value from Microsoft 365 Copilot in Teams meetings. Users can easily and securely enroll their voices via Teams Settings. This feature requires a Microsoft 365 Copilot or Teams Premium license for the user hosting the meeting for Copilot experiences and intelligent recaps, respectively.
Learn more
.

Imagen 3 just got an upgradeWhat:We’re excited to introduce an upgrade to our latest image generation model, Imagen 3, which currently ranks #1 on theImage Generation Leaderboard! This enhanced version delivers richer details and textures, and follows your instructions with greater accuracy to help bring your creative vision to life.Why:We’re committed to empowering your creativity. By providing access tostate-of-the-art modelslike Imagen 3, we aim to fuel your imagination and help you bring even more of your ideas to life.

Imagen 3 just got an upgradeWhat: We’re excited to introduce an upgrade to our latest image generation model, Imagen 3, which currently ranks no. 1 on theImage Generation Leaderboard! This enhanced version delivers richer details and textures, and follows your instructions with greater accuracy to help bring your creative vision to life.Why: We’re committed to empowering your creativity. By providing access tostate-of-the-art modelslike Imagen 3, we aim to fuel your imagination and help you bring even more of your ideas to life.

We’re excited to release o3-mini, our newest cost-efficient reasoning model optimized for coding, math, and science.
On the API, o3-mini supports Structured Outputs, function calling, developer messages, and streaming. It offers three adjustable reasoning efforts (low, medium, and high), so you can balance speed with depth for your use case.
ChatGPT Team, Pro, Plus, and Free plan users can access o3-mini starting today. Additionally, o3-mini now works with search to find up-to-date answers with links to relevant web sources. This is an early prototype as we work to integrate search across our reasoning models. In side-by-side testing, o3-mini delivered results on par with o1 at a lower latency, and outperformed o1-mini on advanced STEM tasks.
Expert evaluators preferred o3-mini’s answers 56% of the time over o1-mini’s, citing improved clarity and fewer critical errors on difficult questions. We look forward to your feedback and will keep refining o3-mini as we expand our family of advanced reasoning models.

Introducing Gemini 2.0 Flash to GeminiWhat:Gemini is now powered by 2.0 Flash, our latest model that is designed for the agentic era, delivering fast responses and stronger performance across a number of key benchmarks for everyday help with tasks like brainstorming, learning, or writing.In addition to 2.0 Flash, Gemini Advanced users will continue to have access to a 1M token context window for up to 1,500 pages of file uploads, priority access to features such as Deep Research and Gems, 2TB of storage, and more.Gemini 1.5 Flash and 1.5 Pro will remain available for the next few weeks.Why:With faster responses and improved performance, we’re making it easier for you to unlock new ways to create, interact, and collaborate with Gemini.

Introducing Gemini 2.0 Flash to GeminiWhat: Gemini is now powered by 2.0 Flash, our latest model that is designed for the agentic era, delivering fast responses and stronger performance across a number of key benchmarks for everyday help with tasks like brainstorming, learning or writing.In addition to 2.0 Flash, Gemini Advanced users will continue to have access to a 1M token context window for up to 1,500 pages of file uploads, priority access to features such as Deep Research and Gems, 2 TB of storage and more.Gemini 1.5 Flash and 1.5 Pro will remain available for the next few weeks.Why: With faster responses and improved performance, we're making it easier for you to unlock new ways to create, interact and collaborate with Gemini.

We’ve made some updates to GPT-4o–it’s now a smarter model across the board with more up-to-date knowledge, as well as deeper understanding and analysis of image uploads.
More up-to-date knowledge:By extending its training data cutoff from November 2023 to June 2024, GPT-4o can now offer more relevant, current, and contextually accurate responses, especially for questions involving cultural and social trends or more up-to-date research. A fresher training data set also makes it easier for the model to frame its web searches more efficiently and effectively.
Deeper understanding and analysis of image uploads:
GPT-4o is now better at understanding and answering questions about visual inputs, with improvements on multimodal benchmarks like MMMU and MathVista. The updated model is more adept at interpreting spatial relationships in image uploads, as well as analyzing complex diagrams, understanding charts and graphs, and connecting visual input with written content. Responses to image uploads will contain richer insights and more accurate guidance in areas like spatial planning and design layouts, as well as visually driven mathematical or technical problem-solving.
A smarter model, especially for STEM:GPT-4o is now better at math, science, and coding-related problems, with gains on academic evals like GPQA and MATH. Its improved score on MMLU—a comprehensive benchmark of language comprehension, knowledge breadth, and reasoning—reflects its ability to tackle more complex problems across domains.
Increased emoji usage ⬆️:GPT-4o is now a bit more enthusiastic in its emoji usage (perhaps particularly so if you use emoji in the conversation ✨) — let us know what you think.

Introducing Microsoft 365 Copilot Chat
[Windows, Web, Android, iOS]
Microsoft 365 Copilot Chat—secure AI chat powered by GPT-4o with agents accessible right in chat, and IT controls including enterprise data protection and agent management. Copilot Chat serves as a powerful new on-ramp for everyone in your organization to build an AI habit. And it is included with your Microsoft 365 subscription. Get started with Copilot Chat with the updated
Microsoft 365 Copilot app
(formerly Microsoft 365 app).
Learn more
.