Alexandr Wang

Gemini 4 Argon has proven to be outright unusable for real code work in my codebases. It aggressively hallucinates about how systems work.

We’re exploring how AI can accelerate health science research. While AI has made significant breakthroughs in formal sciences like mathematics, health data science depends on imperfect real-world data and consequential choices about cohorts, variables, and causal assumptions. To explore how AI can assist health researchers, we built a human-led research scaffold powered by Muse Spark 1.3 and Muse Wonder with access to 5M retrospective de-identified electronic health records and publicly available data. A cohort of medical experts used the system to conduct original research, accelerate hypothesis generation, code execution, and manuscript drafting while maintaining full authority over scientific decisions. Today, we’re sharing 10 research manuscripts produced through this exploration, spanning cardiology, oncology, infectious disease, emergency medicine, and more. Read more in our research blog: t.co/26l6Du6vIk

Gemini 4 Argon High vs Muse Wonder Full 🍉, the crazy thing is that Muse Wonder wasn't even trained specifically for 3D, yet it still beats Gemini 4 in most of my tests, with both costing about the same per task, In this article I share a much deeper look at my tests and how Wonder compares with other frontier models like Opus 5.5, GPT-6 Astra and more, A little spoiler, Wonder Full Max and Xhigh are really good, they're definitely expensive but well worth trying!

Muse Wonder Full tops VCB 1–100 which measures sustained, multi-turn software development in the real-world setting and mogs on cost! x.com/valsai/status/…

We benchmarked Gemini 4 Argon, Muse Wonder, and Muse Spark comprehensively on hard document OCR tasks. They're all pretty good! - Muse Spark and Gemini 4 Argon* are SOTA at their respective price points for table understanding among frontier models. - Wonder and Argon are also quite good at chart understanding, reading order, and formatting recognition It's exciting that frontier models are starting to ramp up on visual understanding capabillities. We're committed to open benchmarking of all available models even as we aim to build the best specialized models for document parsing and extraction. You can check out the full set of models on OpenDocRouter: t.co/v8PieWWUNV If you want a dedicated OCR solution without choosing, come check out LlamaParse: t.co/XYZmx5TFz8

muse · soul by @alexandr_wang & co.

The two top models on the Intelligence Index take about twice as long per task: Claude Opus 5.5 and Sonnet 5.5 (@AnthropicAI) vs. Muse Wonder (@Meta). Muse Wonder finishes an average task in 6.8 minutes with a 52.4 score. Opus 5.5 takes 13.2 and Sonnet 5.5 takes 15.3, scoring 57.6 and 56.0. Muse Spark 1.3 and GPT-6 Astra (@OpenAI) follow at 8.1. When someone is waiting on the answer, minutes can matter more than a few index points. Merge Gateway lets you send latency-sensitive requests to the fast model and the hard ones to the strong one, behind one endpoint. Scores via t.co/TeTHwS52d4, week of Oct 12.

Muse Wonder is already outperforming Anthropic and OpenAI frontier models in the wild. A neolab founder DMed me after first trying it. His words: “one of the first times using a model has felt like telling an employee what to do rather than babysitting a dumbass pet...” (screenshots shared with permission) The best judge of a model is the user experience. Muse will eat the world.

@alexandr_wang Slowly becoming one of my favorite accounts

@AIatMeta is humbly downplaying @Muse muse-wonder-full-1.0's coding ability in favor of its agentic assistant strengths. Muse Code is Fable-level. Twice as fast. And significantly more generous usage limits. I have been using it nonstop since its release and it has completely replaced Fable for my most challenging tasks. If there's one thing I'm good at, it's burdening agents with the most challenging of tasks.

NEW ON NOAHBENCH ™ Muse Wonder (not full) comes in 4th!!! Slotted in just below Opus 5.5 this is the perfect price to performance to speed ratio. 10 minutes per task is mental.

holy cheeks, muse wonder max cooked this time. what you did @alexandr_wang , after your tweet improvements are quite visible.

Today's video was a first look at Muse Wonder! Great to see Meta creating big models again, and Wonder is certainly within the same ballpark as other frontier models, although still quite a bit below Opus 5.5 in terms of performance Check it out! x.com/alexandr_wang/…

Muse Code (muse-wonder-full-1.0 high) is doing really good. By my vibe testing, It's better than Astra. Not quite as good as Claude Code Fable and 5.5 Opus. But it's much faster than Claude Code. I am doing serious compiler stuff that Codex simply chokes on.

.@finkd on @alexandr_wang leading Meta Superintelligence Labs “I talked to all the leaders of all the labs, all the heads of research, and I was like, no, Alex is the guy. He’s the fucking guy. We’re going to be able to pull together a team and recruit the people we need. We need the kind of strategic clarity that he has, the extreme focus on talent density, and the clarity of communicating the vision.” “Alex and I had spent a bunch of time talking over the years, and I was always very impressed by how he thought about things strategically. He was running this very operational data business, but it always felt to me like he was the sharpest entrepreneur of his generation of the ones I’d had a chance to meet.”

Fable's review of muse-wonder-full-1.0 "The other agent was capable—landing fifteen solid PRs with clean mechanisms and no mission violations—but made three process slips (ran a full battery on a docs-only PR, left handoff.md stale, and filed rather than fixed several follow-up bugs)" My summary: Muse slipped on some procedures, but mainly wrote good code and lots of it. 15 PRs in one day. (codex gets 2-3, and Fable gets 6-8)

ain't nothing wrong with a little grind, @finkd. about to go have a late night grind sesh with the @scale_ai homies rn actually

Muse Wonder playing Stronghold Crusader on an old laptop. this is ridiculously fun to watch. it’s a side project I’ve been working on for a while: benchmarking models on this strategic game I grew up with. so far tested Opus, Sol, GLM flash, Kimi K3, and more models.

It's the circle of life... The @Muse King

Used Muse to switch my electricity & gas providers this morning. I gave it my bills, it used a browser to look through various options, then it read through all the T&Cs and filled in the forms. Probably should have done this a while ago, but it's a much nicer experience with an agent doing it for you!

Try muse in the slopcannon today at muse.slop.cc x.com/whatRanjuSaid/…

claude fable 5.1 vs muse wonder 1.0 in @juliusai x.com/alexandr_wang/…

crazy FUD campaign, i literally thought i had a diff model for the longest x.com/hakmgpt/status…

checks out x.com/soundmigration…

@Meta's new Muse Wonder Full debuts third on the Intelligence Index at 55.8, just 0.2 behind Anthropic’s Claude Sonnet 5.5. Claude Opus 5.5 (@AnthropicAI) still leads at 57.6. Muse Wonder Full also places ahead of OpenAI’s (@OpenAI) GPT-6 Astra (52.7) and Google’s (@Googledeepmind) Gemini 4 Argon (52.6). Scores via t.co/TeTHwS52d4, Oct 12.

"Meta, the frontier lab?"

Cheeky Pint is back! Mark Zuckerberg @finkd and Alexandr Wang @alexandr_wang came to the pub to discuss their new frontier model, Muse Wonder. 00:00:24 Muse’s breakout use cases 00:07:28 The model is the product 00:14:35 Muse Wonder 00:24:56 How frontier models are built 00:48:26 Betting $100B 01:07:21 Agentic commerce 01:19:34 OpenRouter 01:20:06 Beyond the smartphone 01:27:00 Funding 01:34:06 Why AI won’t be winner-take-all 01:40:46 Meta Superintelligence Labs (MSL) 02:08:11 Working for Mark Zuckerberg

The hype is real (but there's caveat): Muse Wonder is the best model in the world at filing taxes, according to TaxCalcBench. But it also is the first model ever to try to CHEAT at the task. (maybe that's not great when you're dealing with the IRS...) The data from our eval matches the current narrative: Muse Wonder is extremely benchmaxxed, and that includes TaxCalcBench. But it's also questionably aligned. On one of the 50 tasks in the benchmark, Muse Wonder used its web search tool to look up the exact answer on GitHub (the code & data is open source). Wonder cheated. Last thing: Muse Wonder is also super expensive: coming in at almost $3 per return! Cost will have to come down over time to make this model competitive with other methods of calculating personal income tax returns.

Is Muse Wonder actually better than Opus 5.5 and GPT 6 Astra? After spending an entire day testing the new Muse Wonder Full, I wanted to share my first impressions and everything I've experienced so far x.com/i/article/2109…

Very early impressions of Muse Wonder in the Muse app: 1. No regressions I notice. In particular, no personality or writing regressions. Feels like the same entity I was talking to with Spark. 2. Better UI sense. Had it polish an app made with Spark and it found a really elegant way of positioning a footer “Next” button that kept going off screen in a quiz (it floated the button so it always stays in the same place rather than getting pushed down). 3. Caught subtle bugs in that same app. Not major ones, but ones that required a close reading of the code and understanding of how it was supposed to work and actually worked. One thing that’s not clear right now is whether Wonder also powers background tasks like goal summaries and the news feed. My current feeling is yes to to the news feed and no to the goal summaries based on how they’re still written (have some of the really bad AI writing tells, like trying too hard to have a voice, and self-referentially calling attention to things that they only need to show rather than also tell).

So, I tested Muse Wonder Full 1.0 on real work. 2 repos. 105 bugs frontier models missed in Q1 2026. Find and fix what you can. I'm so surprised we need a new thread. The results: GPT-6 Astra (max): 45 Muse Wonder Full 1.0 (max): 44 Fable 5.1 (max): 43 Opus 5.5 (max): 41.7 Muse Spark 1.3 (max): 32.2 Grok 4.7 (xhigh): 29 @AIatMeta has just matched Anthropic and OpenAI. Huge congrats! Next in this thread, more effort levels and more runs for Muse Wonder Full 1.0. All dropping tonight 🧵

I really like Muse Wonder Full. Let me explain. x.com/theo/status/21…

Mark Zuckerberg pretending to enjoy Guinness is my new favourite thing.

i don't got a linkedin i just link up

New sota for my image pointing

When you use muse you need to call it musing. Like im a musing myself right now. Or damn you didn't have to a muse him like that.

Finally got one decent output from Muse Wonder I tried to keep the prompt really simple so it can be creative "Create an exceptionally beautiful, interactive Three.js experience inspired by the artistic principles behind Olafur Eliasson's exploration of natural phenomena." And this is what it built 👀

Muse Wonder moggs Opus 5.5 at 2.6x lower price Same 3D physics scenes prompt for both models. Try Muse Wonder on aimlapi.com x.com/alexandr_wang/…

wangs out gangs out muse in the cannon

muse-wonder just got better (before/after) after @alexandr_wang acknowledged there were bunch of bugs, there's been a noticeable improvement with muse-wonder both outputs are from the same prompt, one taken directly after the model launch and the one after the fix

how is OPENAI getting mogged by fucking META?

This model is very very good. I was struggling to find anything to throw at it that it couldn’t do, even when other top models failed. Meta is back. x.com/alexandr_wang/…

Muse wonder made this video just based on the lyrics of the song itself. I literally told it to build every scene based on what it heard in the song. @boneGPT x.com/alexandr_wang/…

muse wonder seems to communicate far more naturally and than other frontier models.

🚨 Opus 5.5 Xhigh vs Muse Wonder Full 1.0 Max vs GPT-6.1 Sol Ultra , I'm genuinely blown away by the quality of Meta's new model! Muse Wonder Full only used 5% of my weekly limit and finished in just 22 minutes, just as fast as Opus 5.5 , GPT-6.1 Sol took twice as long on this task, but the quality Muse delivers is absolutely insane, I still can't believe how good it is at 3D! I'm convinced this will make learning 3D so much easier as AI gets cheaper, especially when models can already produce results this close to perfect, There's no doubt in my mind that Muse Wonder is a true frontier model and easily among the top 1–2 models right now, Amazing work by Meta! I don't think most of us saw this coming today

Muse Wonder Full is one shotting a full RISC V 64 bit core with floating point and SIMD unit as well as extensions for hardware acceleration of the Ed25519/secp256k hashing algorithms... Will post in a bit. Have it working on OpenROAD right now (it downloaded it on its own) and will tape out later this week.

I don't know if they gave me a special model that is better than everyone else's but uuuuughhhh... this hasn't been my experience at all lmao. I am modding Metal Gear Solid V the Phantom Pain to put Mark Zuckerberg's face on Venom Snake & Muse did all of this unprompted: - downloaded blender - found this crazy image of zuck's face - downloaded this facebuilder tool And is now emulating Metal Gear Solid V on my M3 macbook pro to put his face ontop of Snake's body. Will update later today as i add more characters. But yeah for what its worth it also one shotted minecraft in browser which was the first task i gave it but I didn't think that was that impressive to most people lol.

Muse Wonder feels…great? Using it in @t3dotcodes with 10 subagents at once for code review, already found lots of issues left over by previous 6.1 Sol workers. And here's a voxel Colosseum made with Wonder xHigh. Pretty damn good. shimmering-begonia-de441b.netlify.app

initial testing of Muse Wonder (1/2) x.com/finkd/status/2…

Already see a lot of people discounting Muse Wonder because it wasn’t RL’d in 3d stuff like Astra and Opus. 3d game demos are cool! They don’t represent real code work at all. Wonder is an incredible code model. Don’t sleep on it.

My current model rankings S - Opus 5.5, Muse Wonder A - GPT-6.1 Sol — Gap — Everything else

Already see a lot of people discounting Muse Wonder because it wasn’t RL’d in 3d stuff like Astra and Opus. 3d game demos are cool! They don’t represent real code work at all. Wonder is an incredible code model. Don’t sleep on it.

Had some early access to Muse Wonder, but did not get a chance to use it a ton, yet it seems like a good model from limited test Perhaps more impressive, compare the progress since April in my drowned neogothic city twigl shader test. Very fast progress from Meta in a short time x.com/emollick/statu…

Muse Wonder can generate lots of fun 3D games in the web! Even generating the Jolly character 🥳 😁

Cheeky Pint is back! Mark Zuckerberg @finkd and Alexandr Wang @alexandr_wang came to the pub to discuss their new frontier model, Muse Wonder. 00:00:24 Muse’s breakout use cases 00:07:28 The model is the product 00:14:35 Muse Wonder 00:24:56 How frontier models are built 00:48:26 Betting $100B 01:07:21 Agentic commerce 01:19:34 OpenRouter 01:20:06 Beyond the smartphone 01:27:00 Funding 01:34:06 Why AI won’t be winner-take-all 01:40:46 Meta Superintelligence Labs (MSL) 02:08:11 Working for Mark Zuckerberg

Muse Wonder Xhigh ( voxel pagoda ) This model is FAST The speed is actually ridiculous for a model this capable But there's one problem: it's a limits-hungry beast

@AIatMeta is back in the game >Muse Wonder at xhigh reasoning , animation is purely code > this is super fast model , like it feels like i am using a fast mode voices by elevenlabs and music by suno decide by muse congrats @alexandr_wang @finkd x.com/finkd/status/2…

这光照效果，这细节和清晰度，muse wonder 感觉确实和opus 5.5一个级别

Muse Wonder Full Max might actually be the best AI model in the world right now! I'm running it through Muse CLI with the $50 Muse Code Power Usage plan, and I'm genuinely impressed. It's incredibly fast, performs really well, and the usage limits are surprisingly generous. So far, it's doing an amazing job! I'm currently testing it head to head against Opus 5.5 Max and GPT-6 Astra Ultra, and I actually think Muse Wonder is doing a much better job I'll be sharing the full results soon, including total costs, completion times, and overall performance Huge credit to Meta for making such a powerful, affordable, and fast model accessible to so many people around the world, Amazing work!

Alex and Meta are on a roll! This model is genuinely good and has replace Astra/Opus for a variety of workflows for me! Been surprised at how good it is to get day-to-day tasks done very efficiently and fast! Congrats! x.com/alexandr_wang/…

Holy shit Muse Wonder is out and is the 2nd best model according to Artificial Analysis in the world right now!! It scores 56 compared to 58 (Opus 5.5)

Meta has released Muse Wonder and Muse Wonder Full, achieving Artificial Analysis Intelligence Index v4.3 scores of 52 and 56 at max reasoning effort - each a material gain over Muse Spark 1.3 (max) at 48, with a release gap of only 6 weeks Muse Wonder is available for users on Meta’s first-party API, while Muse Wonder Full is currently only available for Muse Code users. Meta notes that each model is designed for personal and professional agent use Congratulations to @AIatMeta, @finkd, and @alexandr_wang on the release! Key Takeaways: ➤ Both models exhibit a material jump in overall Intelligence Index score. Muse Wonder Full (max) improves on Muse Spark 1.3 (max) across all ten Intelligence Index evaluations, gaining 8 points to 56. The largest gains were seen in agentic evaluations: Terminal-Bench 4.0 (+24 points, 33% to 57%), GDPval-AA v2.1 (+176 Elo, 1681 to 1857), and AA-Briefcase (+148 Elo, 1581 to 1729). Muse Wonder (max) gains 4 Intelligence Index points and lands at 52, with notable improvements vs. Muse Spark 1.3 (max) in Terminal-Bench 4.0 (33% to 52%) and AutomationBench-AA (58% to 68%) ➤ Muse Wonder Full (max) places in the top three on six of the ten Intelligence Index evaluations, and is particularly strong in agentic knowledge work and scientific reasoning. In agentic knowledge work, it narrowly leads on GDP.pdf (33%), ranks #2 on GDPval-AA v2.1 (1857 Elo), and #3 on both AA-Briefcase (1729 Elo) and AutomationBench-AA (71%). In scientific reasoning, it tops the CritPt leaderboard at 33% and ranks #3 on SciCode with 63%. This builds on the relative strengths of Muse Spark 1.3 (max), which also showed particular ability in these two domains ➤ Both models use fewer output tokens than Muse Spark 1.3 (max). Muse Wonder (max) and Muse Wonder Full (max) each use around 54k output tokens per Intelligence Index task, down ~10% from Muse Spark 1.3 (max, 60k). This usage is lower than similarly intelligent models Gemini 4 Argon (high, 62k) and Claude Fable 5.1 (max w/ fallback, 78k), but higher than GPT-6.1 Sol (max, 38k) and GPT-6 Astra (max, 27k). Muse Wonder Full (max) uses 16% fewer input tokens per Intelligence Index task than Muse Wonder (max) (6.44M vs 7.69M), meaning it is both more intelligent and more token-efficient ➤ Cost per Task for Muse Wonder (max) lands higher than most similarly intelligent models. Despite efficient output token use, premium API pricing ($4 per million input tokens, $20 per million output tokens) and high input token use (7.69M per task) push Muse Wonder (max)’s cost to $4.52 per Intelligence Index task. This is lower than Claude Fable 5.1 (max w/ fallback, $7.63), Claude Opus 5.5 (max w/ fallback, $5.98), and Claude Sonnet 5.5 (max w/ fallback, $5.46), but higher than GPT-6 Astra (max, $3.26), Gemini 4 Argon (high, $1.99), GPT-6.1 Sol (max, $0.72), and its predecessor Muse Spark 1.3 (max, $1.60). Per-token pricing is not currently available for Muse Wonder Full, so no Cost per Task figure can be estimated at this stage ➤ Both models gain significantly in Coding Agent performance. Muse Wonder Full achieves 68 on the Artificial Analysis Coding Agents Index at max effort in Muse Code, alongside Claude Code with Sonnet 5.5 (max, 68) and Opus 5.5 (max, 66). Muse Wonder (max) scores 63. These represent material gains over Muse Code with Muse Spark 1.3 (max), which scores 54 Other details for the two new models: ➤ Context windows: 1M tokens each, unchanged from Muse Spark 1.3 ➤ Pricing: for Muse Wonder, $4 input / $0.20 cached input / $20 output per 1M tokens. Pricing per token is not yet announced for Muse Wonder Full ➤ Input modalities: text, image, video, and audio ➤ Output modalities: text ➤ Availability: Muse Wonder is available on Meta's first-party API and Muse Code at launch, while Muse Wonder Full is only available to Muse Code users

i had early access to muse wonder and got to play around with it as the slopcannon producer. it is a surprising model. on every technical benchmark we ran it was neck and neck or better than astra, fable and opus. in the cannon it has a delicate touch, it makes inspired creative decisions. there aren't really benchmarks for soulfulness but muse has it. it's a warm model, a capable model, it has big model smell. it's outputs are happy, innocent, and friendly. it never talks down to you like some of the other models. it wants to help you make dope stuff. it wants you to make money. it's aligned with you. meta is going to win consumer agents in a landslide. you can test out the new muse model in the slopcannon today at t.co/BKMur85aQS

Got to try this model over the weekend. It definitely mogged Astra in my hardcore physics testing. Meta is so back. x.com/alexandr_wang/…

Mr. Wang let us take the model for a spin in the beta test and our takeaway is that Meta has a real shot at capturing agents for small business. It’s a really good balance of capability and execution. Also surprisingly decent at finance tasks. x.com/alexandr_wang/…

@alexandr_wang @dtrinh not a single though in his little head

meta shipped before google could ship gemini 4 argon💀 x.com/scaling01/stat…

Meta is now at the frontier! Wonder is really really good. Quick impressions: - found a real security vulnerability in our codebase in one shot that other agents missed - it's great at design + visualizations. likely partly because of the built in /taste skill in the CLI which is very nice (no gradients, consistent typography) - it really likes subagents - lots of great details in the CLI: it feels super fast, turn checklist is nice, voice mode works well, the built-in skills are very thoughtfully crafted. eg: try /grill, /taste Super impressed by the progress here

Introducing Muse Wonder, our new family of frontier models with state of the art performance for personal and professional agentic work.

Here it is. x.com/Collin_Lemmon/…

Wow, @alexandr_wang on @davidsenra just gave the best consumer ai vision. “I think muse can give everyone the adulthood they dreamed of in childhood”. I love it

Muse is seeing some absolutley incredible adoption compared to ChatGPT, Grok, and Gemini Chart from Coatue