🚀 Meet 𝐀𝐠𝐞𝐧𝐭 𝐒2.5, our latest open-source Computer Use Agent, setting 𝐧𝐞𝐰 𝐒𝐎𝐓𝐀 𝐨𝐧 𝐎𝐒𝐖𝐨𝐫𝐥𝐝 𝐕𝐞𝐫𝐢𝐟𝐢𝐞𝐝!. ✨ Upgrades: Flexible context engineering + advanced high-level reasoning. 📈 OSWorld Verified Results:.- 100 steps: S2.5 56.0% (Prev. SOTA: 53.1%)

Generative World Model Theorem: Video generation is to the physical world what LLMs are to the internet.

RT @ShijieZhoucla: 🚀 Our #ICCV2025 work — VLM4D is here!. The first benchmark to truly test if Vision-Language Models can think in 4D (3D s….

They should have named the o-series GPT-5. Period.

Really? Still?

My ranking: . GPT-3.5 (ChatGPT) > o1 > GPT-4V > o3 >= GPT-4o > GPT-5 > GPT-4.1 > GPT-4.5.

o1 is a bigger deal than GPT-5. If the reasoning series weren't released, GPT-5 would've stunned everyone. IMO, the reason why o1 was not named as GPT-5 is that they wanted GPT-5 to be a single model for both. However, now GPT-5 is a hybrid system of multiple underlying models.

If it wasn't clear to you, now it is super clear.

TBH, functionality-wise, GPT-5 meets my expectations, but I am a bit disappointed that it is not a single model yet, which I thought was the goal for GPT-5. This feels like an early release due to unknown pressure.

RT @bremen79: Dear Reviewers,. It is completely fine to admit you were wrong in your initial evaluations. You will not lose anything, and t….

RT @FrancescoLocat8: @xwang_lk PC of the DB track here! Please kindly remind the reviewer that contributing new methods is not necessary fo….

A YC startup told us that they are building everything on top of Agent S2. VCs back startups with capital. We back them with infra and agent tech. More importantly, Agent S keeps evolving. So you can, too.

A benchmark paper submitted to NeurIPS Dataset & Benchmark Track. Reviewer: Valuable benchmark! But no new modeling method proposed. Reject. 😂😂😂.

One step at a time, we will get there.

AGI will never be achieved without research.

Tried gpt-oss locally. It is good, quite slow though. It works exactly like R1. Maybe better, but they work the same way. (R1 deserves a citation.). No multimodal yet. But I think the community will build multimodal-oss very soon. Re: why open source — by gpt-oss. “From the

👏👏👏.

The more you know, the more you don’t know.

@SimularAI 👏 Huge shout-out to the amazing @SimularAI Research team for the dedication and hard work. More exciting updates coming soon!. 🙏 Special thanks to @TianbaoX and the OSWorld team for making the OSWorld AWS evaluation reliable and accessible!.

Agent S2.5 is open sourced here: Try it out!. It is also available on @SimularAI Cloud (no deployment needed).