1/ today we're releasing muse spark, the first model from MSL. nine months ago we rebuilt our ai stack from scratch. new infrastructure, new architecture, new data pipelines. muse spark is the result of that work, and now it powers meta ai. 🧵

We’re open-sourcing BrowseComp (“Browsing Competition”), a new, challenging benchmark designed to test how well AI agents can browse the internet to find hard-to-locate information. It’s like an online scavenger hunt…but for browsing agents. openai.com/index/browsecomp/

americans sure love giving their data away to the CCP in exchange for free stuff

After a great time at OpenAI, we (@EdwardSun0909, @_jasonwei) recently joined @Meta Superintelligence Labs. The first month has already been so much fun building from a clean slate with a truly talent-dense team! Very excited about the compute and long term focus of the new lab

New verified ARC-AGI-Pub SoTA! @OpenAI o3 has scored a breakthrough 75.7% on the ARC-AGI Semi-Private Evaluation. And a high-compute o3 configuration (not eligible for ARC-AGI-Pub) scored 87.5% on the Semi-Private Eval. 1/4

common themes: AGI agents much better 4o upgrade much better memory longer context “grown up mode” deep research feature better sora more personalization (interestingly, many great updates we have coming were mentioned not at all or very little!)

We're also sharing the system card, detailing how we built deep research, assessed its capabilities and risks, and improved safety. openai.com/index/deep-resear…

o3 (left) and Grok 4 (right) replays below spoiler: neither complete a single level

ChatGPT can now do work for you using its own computer. Introducing ChatGPT agent—a unified agentic system combining Operator’s action-taking remote browser, deep research’s web synthesis, and ChatGPT’s conversational strengths.

1/N I’m excited to share that our latest @OpenAI experimental reasoning LLM has achieved a longstanding grand challenge in AI: gold medal-level performance on the world’s most prestigious math competition—the International Math Olympiad (IMO).

🌟Easy-to-Hard Generalization: Scalable Alignment Beyond Human Supervision 🌟 arxiv.org/abs/2403.09472 How can we keep improving AI systems when their capabilities surpass those of human supervisors? (1/n)

Starting today, memory in ChatGPT can now reference all of your past chats to provide more personalized responses, drawing on your preferences and interests to make it even more helpful for writing, getting advice, learning, and beyond.

Deep Research is genuinely useful - depending on your application - but crucially (as anticipated by Rebooting AI in 2019, and by @yudapearl) facts and temporal reasoning remain problematic for current neural network-based approaches that lean heavily on statistics rather than deep understanding.

OpenAI gave us early access to GPT-5: our independent benchmarks verify a new high for AI intelligence. We have tested all four GPT-5 reasoning effort levels, revealing 23x differences in token usage and cost between the ‘high’ and ‘minimal’ options and substantial differences in intelligence We have run our full suite of eight evaluations independently across all reasoning effort configurations of GPT-5 and are reporting benchmark results for intelligence, token usage, and end-to-end latency. What @OpenAI released: OpenAI has released a single endpoint for GPT-5, but different reasoning efforts offer vastly different intelligence. GPT-5 with reasoning effort “High” reaches a new intelligence frontier, while “Minimal” is near GPT-4.1 level (but more token efficient). Takeaways from our independent benchmarks: ⚙️ Reasoning effort configuration: GPT-5 offers four reasoning effort configurations: high, medium, low, and minimal. Reasoning effort options steer the model to “think” more or less hard for each query, driving large differences in intelligence, token usage, speed, and cost. 🧠 Intelligence achieved ranges from frontier to GPT-4.1 level: GPT-5 sets a new standard with a score of 68 on our Artificial Analysis Intelligence Index (MMLU-Pro, GPQA Diamond, Humanity’s Last Exam, LiveCodeBench, SciCode, AIME, IFBench & AA-LCR) at High reasoning effort. Medium (67) is close to o3, Low (64) sits between DeepSeek R1 and o3, and Minimal (44) is close to GPT-4.1. While High sets a new standard, the increase over o3 is not comparable to the jump from GPT-3 to GPT-4 or GPT-4o to o1. 💬 Token usage varies 23x between reasoning efforts: GPT-5 with High reasoning effort used more tokens than o3 (82M vs. 50M) to complete our Index, but still fewer than Gemini 2.5 Pro (98M) and DeepSeek R1 0528 (99M). However, Minimal reasoning effort used only 3.5M tokens which is substantially less than GPT-4.1, making GPT-5 Minimal significantly more token-efficient for similar intelligence. 📖 Long Context Reasoning: We released our own Long Context Reasoning (AA-LCR) benchmark earlier this week to test the reasoning capabilities of models across long sequence lengths (sets of documents ~100k tokens in total). GPT-5 stands out for its performance in AA-LCR, with GPT-5 in both High and Medium reasoning efforts topping the benchmark. 🤖 Agentic Capabilities: OpenAI also commented on improvements across capabilities increasingly important to how AI models are used, including agents (long horizon tool calling). We recently added IFBench to our Intelligence Index to cover instruction following and will be adding further evals to cover agentic tool calling to independently test these capabilities. 📡 Vibe checks: We’re testing the personality of the model through MicroEvals on our website which supports running the same prompt across models and comparing results. It’s free to use, we’ll provide an update with our perspective shortly but feel free to share your own! See below for further analysis:

The lightweight version of deep research is powered by a version of OpenAI o4-mini and is nearly as intelligent as the deep research people already know and love, while being significantly cheaper to serve. Responses will typically be shorter while maintaining the depth and quality you’ve come to expect. Once limits for the original version of deep research are reached, queries automatically default to the lightweight version.

Deep Research Live from Tokyo 4pm PT / 9am JST Stay tuned for link to livestream.

🔭 Understand the Universe in less than a min. Grok 3 DeepSearch

I wish every poster at every conference looked like this. The idea that all text should be same size is just wrong.

LIMA: Less Is More for Alignment LIMA, a 65B parameter LLaMa language model fine-tuned with the standard supervised loss on only 1,000 carefully curated prompts and responses, without any reinforcement learning or human preference modeling. LIMA demonstrates remarkably strong performance, learning to follow specific response formats from only a handful of examples in the training data, including complex queries that range from planning trip itineraries to speculating about alternate history. Moreover, the model tends to generalize well to unseen tasks that did not appear in the training data. In a controlled human study, responses from LIMA are either equivalent or strictly preferred to GPT-4 in 43% of cases; this statistic is as high as 58% when compared to Bard and 65% versus DaVinci003, which was trained with human feedback paper page: huggingface.co/papers/2305.1…

🚀 Can RLAIF fully replace RLHF to align language models from scratch, enhancing both their alignment and capabilities? SALMON introduces a principle-following reward model in the realm of self-alignment, using just 6 ICL exemplars and 31 principles to outperform LLaMA-2-Chat!

In 2009, Google created the PhD Fellowship Program to recognize and support outstanding graduate students pursuing exceptional research in computer science and related fields. Today, we congratulate the recipients of the 2023 Google PhD Fellowship! goo.gle/3PYfLXl

Introducing OpenAI o3 and o4-mini—our smartest and most capable models to date. For the first time, our reasoning models can agentically use and combine every tool within ChatGPT, including web search, Python, image analysis, file interpretation, and image generation.

Principle-Driven Self-Alignment of Language Models from Scratch with Minimal Human Supervision abs: arxiv.org/abs/2305.03047 paper page: huggingface.co/papers/2305.0… project page: mitibmdemos.draco.res.ibm.co…

🚀 Can RLAIF fully replace RLHF to align language models from scratch, enhancing both their alignment and capabilities? SALMON introduces a principle-following reward model in the realm of self-alignment, using just 6 ICL exemplars and 31 principles to outperform LLaMA-2-Chat!

You can now connect GitHub repos to deep research in ChatGPT. 🐙 Ask a question and the deep research agent will read and search the repo’s source code and PRs, returning a detailed report with citations. Hit deep research → GitHub to get started.

Register here for the talk: forum.openai.com/home/events…

Today we’re launching Research, alongside a new Google Workspace integration. Claude now brings together information from your work and the web.

congrats to the team, especially @isafulf and @EdwardSun0909, for building an incredible product. my very approximate vibe is that it can do a single-digit percentage of all economically valuable tasks in the world, which is a wild milestone.

Easy-to-Hard Generalization was accepted to NeurIPS! Congrats to @EdwardSun0909 and @scut_longhui! Check out the updated camera-ready version here: openreview.net/pdf?id=qwgfh2…

How does a language model decide which questions it will engage with and which it deems inappropriate? We use Constitutional AI to more directly encode values into our language models.

TL;DR: we are excited to release a powerful new open-weight language model with reasoning in the coming months, and we want to talk to devs about how to make it maximally useful: openai.com/open-model-feedba… we are excited to make this a very, very good model! __ we are planning to release our first open-weigh language model since GPT-2. we’ve been thinking about this for a long time but other priorities took precedence. now it feels important to do. before release, we will evaluate this model according out our preparedness framework, like we would for any other model. and we will do extra work given that we know this model will be modified post-release. we still have some decisions to make, so we are hosting developer events to gather feedback and later play with early prototypes. we’ll start in SF in a couple of weeks followed by sessions in europe and APAC. if you are interested in joining, please sign up at the link above. we’re excited to see what developers build and how large companies and governments use it where they prefer to run a model themselves.

tip for chatgpt agent slides: first ask it to do the research only, then ask it to make the slides!

How can informal reasoning improve formal theorem proving? New paper: "Lean-STaR: Learning to Interleave Thinking and Proving" arxiv.org/abs/2407.10040 We introduce a framework for learning to interleave informal thoughts with steps of formal proving. 46.3% on miniF2F 🔥

Sam this performance is crazy

GPT-4.5 has entered the Chat. openai.com/live/

Welcome Gemma 3, our new open-weight LLM from @GoogleDeepMind. All sizes (1B, 4B, 12B and 27B) excel on benchmarks, but the key result may be the 27B reaching 1338 on LMSYS. For this, we scaled post-training, with our novel distillation, RL and merging strategies. Happy building!

Mitigating racial bias from LLMs is a lot easier than removing it from humans! Can’t believe this happened at the best AI conference @NeurIPSConf We have ethical reviews for authors, but missed it for invited speakers? 😡

we are working to make it much more efficient and then will offer much higher usage limits in the meantime, please send me your email address

🔥Our CMU-MATH team proudly clinched 2nd place in the Artificial Intelligence Mathematical Olympiad (AIMO) out of 1,161 participating teams with the best performance of an academic team! Dive into our blog to discover our winning formula: blog.ml.cmu.edu/2024/07/29/c…

this is ready now for the world

Just launched ChatGPT Agent (sorry GPT-5 waiters, it is coming!), the most capable AI agent model to date! It has been such an honor to be part of a crazy sprint to get this amazing model trained and shipped together with an absolutely gem team (@isafulf , @caseychu9 , @EdwardSun0909 , @josh_tobin_ Yash Kumar and many more)! I am so proud of this project, so I want to share some highlights, personal takes and lessons learned while working on it: 1. Used for research 📕 + actions 💻 + slides generation: Deep Research can do research. Operator can take actions for you. ChatGPT Agent can do both at the same time! E.g. you can ask it to make a plan for a trip to Hawaii, find good deals on hotels and flights, and book them on your behalf using its own computer! It can also generate slides! 2. Power of end-to-end RL: How do we build it? You guess it right! It is us, @OpenAI RL diehards. You are probably tired of hearing about RL scaling. Me, too. But when I feel its power first-hand, its effectiveness and data efficiency still shock me and feel like magic 🪄. 3. First OpenAI model of high biorisk 💀: Not sure this is something I should proud of or not :) For an ex-AI bio PhD researcher like me, this is something a bit personal. One one hand, many of my biomedicine researcher friends tell me that AI agents have significantly helped with their research. On the other hand, such a capable model can amplify the risk of malicious actors building bioweapons. Our safety team has done incredible work to mitigate the risks. 4. Collaboration with users 👪 is core: We want our AI to augment and enhance humans, not to replace them, so we work hard to make the model good at collaborating with the user. You can type a message at anytime to interrupt it and steer it to new directions. The model will always confirm with you before taking actions like buying things for you or deleting a file on your google drive. And the model will ask clarification questions only when it needs more clarity from you! 5. How to generate good slides: As in other cases, writing a well-specified prompt always helps! Also try first telling it to generate a report, then convert the report into slides! 6. Real-world performance > benchmark chasing: One thing outside people may not know about us is how little attention we pay to external benchmarks during the model dev process. We do not focus on hill-climbing on them, and we do not care that much about how we end up on the leaderboard. That said, as a byproduct of our pursuit to great real-world performance and true intelligence, ChatGPT Agent does crush many benchmarks! Wanna learn more? Read our blog linked in the end! In the end, I want to shout out to my amazing team again. These extremely talented and kind people are the reason why OpenAI is constantly making magic like this! ❤️ Also please try ChatGPT Agent and give us feedback! You can reply here in the thread or my DM is open. This is just the start. We will continue working hard towards more and more capable super-human AI agents! 🤖 openai.com/index/introducing…

Yall heard it from the man himself

We're releasing a preview of OpenAI o1—a new series of AI models designed to spend more time thinking before they respond. These models can reason through complex tasks and solve harder problems than previous models in science, coding, and math. openai.com/index/introducing…

To deepen the public conversation about how AI models should behave, we’re sharing our Model Spec — our approach to shaping desired model behavior. openai.com/index/introducing…

🚀 Can RLAIF fully replace RLHF to align language models from scratch, enhancing both their alignment and capabilities? SALMON introduces a principle-following reward model in the realm of self-alignment, using just 6 ICL exemplars and 31 principles to outperform LLaMA-2-Chat!

Active LLM Retrieval Augmented Generation -Iteratively uses a prediction of upcoming sentence to anticipate future content which is used as query to retrieve relevant docs to regenerate sentence -On 4 long-form generation tasks: superior / competitive arxiv.org/abs/2305.06983

We study common inference strategies (e.g., majority voting, MCTS) and a new tree search, along with various model sizes. First, we hold the inference strategy fixed, and find that across different model sizes, smaller models typically have better accuracy-cost tradeoffs

o3-deep-research: platform.openai.com/docs/mod… o4-mini-deep-research: platform.openai.com/docs/mod… These models are the same post-trained o3 and o4-mini models that power deep research in ChatGPT. They also support MCP (search/fetch) and Code Interpreter.

2025 is the year of agents.

“If I were given one hour to save the planet, I would spend 59 minutes defining the problem and one minute resolving it.” — Albert Einstein

🚀 Can RLAIF fully replace RLHF to align language models from scratch, enhancing both their alignment and capabilities? SALMON introduces a principle-following reward model in the realm of self-alignment, using just 6 ICL exemplars and 31 principles to outperform LLaMA-2-Chat!

Genuine question: What's the (scientific) value of these recent papers that finetune a smaller LM on GPT-4 outputs? It's obviously useful to have a smaller LM that performs ≈ GPT-4 in specific settings. But I don't see the value in packaging into a paper and flooding arXiv.

Today, @rhythmrg, @lindensli and I are introducing @appliedcompute. We’re building Specific Intelligence for the enterprise. Achieving SOTA today means specialization in both human and machine talent. We’ve spent the last six months working with companies like @cognition, @DoorDash, and @mercor_ai, unlocking their company knowledge to build custom agent workforces that outperform frontier models at specific tasks. My cofounders and I all worked on different parts of this problem while at OpenAI, from Codex to o1 to the ML systems and infrastructure for RL training. Two-thirds of our team (see below!) are former founders, and everyone brings a deep technical background, from top AI researchers to Math Olympiad winners. We’ve raised $80M from @benchmark, @sequoia, @Lux_Capital, @eladgil, @victoralazarte, and @Casspi18, and we’re hiring across engineering and research.

AI assistants are trained to give responses that humans like. Our new paper shows that these systems frequently produce ‘sycophantic’ responses that appeal to users but are inaccurate. Our analysis suggests human feedback contributes to this behavior.

The Zephyr-beta model from @huggingface H4 (led by @_lewtun and @edwardbeeching these days) is a great example of engineering practices and know how slowly kicking into gear for RLHF. Some takeaways beyond "high MT Bench and AlpacaEval scores": * DPO can work great for smaller models. This is huge for open-ML as small specialized models are the future. People need to try DPO on specialized application feedback datasets! * Long-term engineering investment can pay off on unexpected timelines. I left a week or two before Zephyr, and we didn't even have it on the plan yet. Finding the right dataset and plugging it into the pipeline can be everything. * Scaling laws for DPO is still unproven. Lots and lots of RLHF experts are skeptical of it for larger models. I think there may need a slight change of the loss function for stability (and maybe sample efficiency is lower), which is why the H4 team found success in multiple Epochs. * MT Bench / Alpaca Eval are pretty saturated. Next year, your model is going to need to get 7+ on MTBench to be considered, but above that may not matter. It's getting normalized, but we need more eval tools still. *AI Feedback is extremely broad. For this model, it was data curation. I expect it to also work for filtering and more. * Releasing both SFT and RL checkpoints with data is great for replication (as the team did). Excited to see where this goes next! Paper: arxiv.org/abs/2310.16944 Artifacts: huggingface.co/collections/H…

This is what I sent to my colleagues at OpenAI: Hi all, I made the difficult decision to leave OpenAI as an employee, but I’m looking to work closely together as a partner going forward. Contributing to the mission of OpenAI and working with world-class teams to create and improve ChatGPT has been an experience of a lifetime. But I’ve gotten really excited about AI for science. My undergrad was in physics and I’m keen to apply this technology there. Because AI for science is one of the most strategically important areas to OpenAI and achieving ASI, OpenAI is planning to invest in and partner with my new company. So I’ll see you all around! Thanks to all the leadership who believed in me early on, especially, Sam, Greg, and Mark. Thank you everyone on post-training and to all of our collaborators across research and product. I’ll miss working with so many of you, but will be cheering you on! Post-training has an amazing roster of talent and leaders who will continue to drive its success.

[1/4] Large language models (LLMs) tend to hallucinate, especially when generating long outputs. We present active retrieval augmented generation, in which an LLM actively decides when and what to retrieve throughout the generation process.

💡 We release methods and datasets for extremely data-efficient alignment 🚀 6K SFT samples lead to 7.22 MT-Bench score, further DPO with 10K samples achieve 7.55 MT-Bench +90+% AlpacaEval Try our data to align models more efficiently: github.com/hkust-nlp/deita

🦙🐪🐫 So many instruction tuning datasets came out recently! How valuable are they, and how far are open models really from proprietary ones like ChatGPT? 🧐We did a systematic exploration, and built Tülu---a suite of LLaMa-tuned models up to 65B! 📜arxiv.org/abs/2306.04751

[p1] 🐕Direct Preference Optimization of Video Large Multimodal Models from Language Model Reward🐕 Paper link: arxiv.org/pdf/2404.01258.pdf… page: github.com/RifleZhang/LLaVA-… How to effectively train video large multimodal Model (LMM) alignment with preference modeling?