Hey, this is Alex!

We’re finally so back! Tons of open source releases, OpenAI updates GPT and a few breakthroughs in audio as well, makes this a very dense week!

Today on the show, we covered the newly released GPT 5.1 update, a few open source releases like Terminal Bench and Project AELLA (renamed OASSAS), and Baidu’s Ernie 4.5 VL that shows impressive visual understanding!

Also, chatted with Paul from 11Labs and Dima Duev from the wandb SDK team, who brought us a delicious demo of LEET, our new TUI for wandb!

Tons of news coverage, let’s dive in 👇 (as always links and show notes in the end)

Open Source AI

Let’s jump directly into Open Source as this week has seen some impressive big company models.

Terminal-Bench 2.0 - a harder, highly‑verified coding and terminal benchmark (X, Blog, Leaderboard)

We opened with Terminal‑Bench 2.0 plus its new harness, Harbor, because this is the kind of benchmark we’ve all been asking for.

Terminal‑Bench focuses on agentic coding in a real shell. Version 2.0 is a hard set of 89 terminal tasks, each one painstakingly vetted by humans and LLMs to make sure it’s solvable and realistic. Think “I checked out master and broke my personal site, help untangle the git mess” or “implement GPT‑2 code golf with the fewest characters.”

On the new leaderboard, top agents like Warp’s agentic console and Codex CLI + GPT‑5 sit around fifty percent success. That number is exactly what excites me: we’re nowhere near saturation. When everyone is in the 90‑something range, tiny 0.1 improvements are basically noise. When the best models are at fifty percent, a five‑point jump really means something.

A huge part of our conversation focused on reproducibility. We’ve seen other benchmarks like OSWorld turn out to be unreliable, with different task sets and non‑reproducible results making scores incomparable. Terminal‑Bench addresses this with Harbor, a harness designed to run sandboxed, containerized agent rollouts at scale in a consistent environment. This means results are actually comparable. It’s a ton of work to build an entire evaluation ecosystem like this, and with over a thousand contributors on their Discord, it’s a fantastic example of a healthy, community‑driven effort. This is one to watch!

Baidu’s ERNIE‑4.5‑VL “Thinking”: a 3B visual reasoner that punches way up (X, HF, GitHub)

Next up, Baidu dropped a really interesting model, ERNIE‑4.5‑VL‑28B‑A3B‑Thinking. This is a compact, 3B active‑parameter multimodal reasoning model focused on vision, and it’s much better than you’d expect for its size. Baidu’s own charts show it competing with much larger closed models like Gemini‑2.5‑Pro and GPT‑5‑High on a bunch of visual benchmarks like ChartQA and DocVQA.

During the show, I dropped a fairly complex chart into the demo, and ERNIE‑4.5‑VL gave me a clean textual summary almost instantly—it read the chart more cleanly than I could. The model is built to “think with images,” using dynamic zooming and spatial grounding to analyze fine details. It’s released under an Apache‑2.0 license, making it a serious candidate for edge devices, education, and any product where you need a cheap but powerful visual brain.

Open Source Quick Hits: OSSAS, VibeThinker, and Holo Two

We also covered a few other key open-source releases. Project AELLA was quickly rebranded to OSSAS (Open Source Summaries At Scale), an initiative to make scientific literature machine‑readable. They’ve released 100k paper summaries, two fine-tuned models for the task, and a 3D visualizer. It’s a niche but powerful tool if you’re working with massive amounts of research. (X, HF)

WeiboAI (from the Chinese social media company) released VibeThinker‑1.5B, a tiny 1.5B‑parameter reasoning model that is making bold claims about beating the 671B DeepSeek R1 on math benchmarks.

We discussed the high probability of benchmark contamination, especially on tests like AIME24, but even with that caveat, getting strong chain‑of‑thought math out of a 1.5B model is impressive and useful for resource‑constrained applications. (X, HF, Arxiv)

Finally, we had some breaking news mid‑show: H Company released Holo Two, their next‑gen multimodal agent for controlling desktops, websites, and mobile apps. It’s a fine‑tune of Qwen3‑VL and comes in 4B and 8B Apache‑2.0 licensed versions, pushing the open agent ecosystem forward. (X, Blog, HF)

ThursdAI - Recaps of the most high signal AI weekly spaces is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

Big Companies & APIs

GPT‑5.1: Instant vs Thinking, and a new personality bar

The biggest headline of the week was OpenAI shipping GPT‑5.1, and this was a hot topic of debate on the show. The update introduces two modes: “Instant” for fast, low‑compute answers, and “Thinking” for deeper reasoning on hard problems. OpenAI claims Instant mode uses 57% fewer tokens on easy tasks, while Thinking mode dedicates 71% more compute to difficult ones. This adaptive approach is a smart evolution.

The release also adds a personality dropdown with options like Professional, Friendly, Quirky, and Cynical, aiming for a more “warm” and customizable experience.

Yam and I felt this was a step in the right direction, as GPT‑5 could often feel a bit cold and uncommunicative. However, Wolfram had a more disappointing experience, finding that GPT‑5.1 performed significantly worse on his German grammar and typography tasks compared to GPT‑4 or Claude Sonnet 4.5. It’s a reminder that “upgrades” can be subjective and task‑dependent.

Since the show was recorded, GPT 5.1 is also released in the API and they have published a prompting guide and some evals! With some significant jumps across SWE-bench verified and GPQA Diamond! We’ll be testing this model out all week.

The highlight for this model is the creative writing, it was made public that this model was being tested on OpenRouter as Polaris-alpha and that one tops the eqbench creative writing benchmarks beating Sonnet 4.5 and Gemini!

Grok‑4 Fast: 2M context and a native X superpower

Grok‑4 Fast from xAI apparenly quietly got a substantial upgrade to a 2M‑token context window, but the most interesting part is its unique integration with X. The API version has access to internal tools for semantic search over tweets, retrieving top quote tweets, and understanding embedded images and videos. I’ve started using it as a research agent in my show prep, and it feels like having a research assistant living inside X’s backend—something you simply can’t replicate with public tools.

I still have my gripes about their “stealth upgrade” versioning strategy, which makes rigorous evaluation difficult, but as a practical tool, Grok‑4 Fast is incredibly powerful. It’s also surprisingly fast and cost‑effective, holding its own against other top models on benchmarks while offering a superpower that no one else has.

Google SIMA 2: Embodied Agents in Virtual Worlds

Google’s big contribution this week was SIMA 2, DeepMind’s latest embodied agent for 3D virtual worlds. SIMA lives inside real games like No Man’s Sky and Goat Simulator, seeing the screen and controlling the game via keyboard and mouse, using Gemini as its reasoning brain. Demos showed it following complex, sketch‑based instructions, like finding an object that looks like a drawing of a spaceship and jumping on top of it.

When you combine this with Genie 3—Google’s world model that can generate playable environments from a single image—you see the bigger picture: agents that learn physics, navigation, and common sense by playing in millions of synthetic worlds. We’re not there yet, but the pieces are clearly being assembled. We also touched on the latest Gemini Live voice upgrade, which users are reporting feels much more natural and responsive

More Big Company News: Qwen Deep Research, Code Arena, and Cursor

We also briefly covered Qwen’s new Deep Research feature, which offers an OpenAI‑style research agent inside their ecosystem. LMSYS launched Blog, a fantastic live evaluation platform where models build real web apps agentically, with humans voting on the results. And in the world of funding, the AI‑native code editor Cursor raised a staggering $2.3 billion, a clear sign that AI is becoming the default way developers interact with code.

Th