introducing Flowith 3.0. the ultimate canvas for human-ai creation.
two years ago, we were the first to move ai out of the chatbox and onto an infinite canvas. today, we are pushing that boundary much further.
everyone talks about full autonomy, but for real creation, keeping the human in the loop is everything.
in flowith 3.0, you orchestrate every sota model in one place:
· pro video & image: direct with a 3d stage, relighting, and pro editing powered by seedance, muse & gpt image.
· cloud-native coding: ship real products with claude code, codex, and gemini 4 argon pulling straight from your canvas assets.
· hybrid autonomy: plug in your local agents via api to drive the canvas autonomously.
we’re also launching $9.99 FLOWITH BLACK PASS, frontier intelligence is finally dirt cheap:
- up to 600 video generations
- up to 90% off claude opus 5.5 & gpt astra
- up to 95% off seedance, seedream, gpt image & muse
live today. go build on https://t.co/BYgtZxN8Af
RT + comment to win a Black Pass⬇️
tomorrow DeepMind will be dropping Gemini 4 Sodium, a fast yet powerful frontier model which costs just $11.75/M (introductory pricing, raises to $16.50 soon) and is their best model for various things and stuff, rivaling Argon, Carbon, Hydrogen, and Neon in several situations
The AI industry has a benchmarking problem.
We're letting random people on the internet decide which models are best
even for capabilities like design...
So we built something different.
Introducing The Frontier Design Benchmark from @contralabs
https://t.co/Eiw47GSxEk
Introducing Gemini 4 Neon! Pushing the frontier across domains and punching above its model class in a meaningful way. To me, Neon feels like the right balance of intelligence, speed, and usability. Available now across @GoogleAIStudio, the API, and more!
Distortion waves library is now available on desktop React, ready to use for free.
https://t.co/5vihemxJKx
Includes edge glows, anticipation movement, chromatic aberration, particles, and some other details.
I tested GPT-6.2 Astra on real work. 2 repos, 105 replanted bugs. Find and fix what you can.
It's the new #1. The results:
- GPT-6.2 Astra (max): 62.7 for $28.35
- GPT-6.2 Astra (xhigh): 60 for $23.53
- Fable 5.2 (max): 58 for $139.65
- Gemini 4 Carbon (high): 57.3 for $31.32
- Sonnet 5.5 (max): 51.3 for $154.03
- GPT-6 Astra (max): 45 for $33.03
From GPT-6 Astra to GPT-6.2 Astra: +17.7 points at max, +17 at xhigh.
Whatever people say about other use cases, it's a code review beast.
GPT-6.2 Sol and more effort levels drop soon in the thread 🧵
GPT-6.2 Astra is the new #1 on the LLM Creative Story-Writing Benchmark! 🏆
It scores 4.30, ahead of Claude Fable 5.2 at 3.56.
Against its predecessor, GPT-6 Astra, it swept all 50 matched prompts.
Qwen 4 Max jumps 3.72 points over its predecessor, from -0.17 for Qwen 3.8 Max to 3.55. That puts it at #3 overall, just behind Fable 5.2. 🥉
ByteDance Seed2.1 Pro scores -0.67, up from -2.18 for Seed2.0 Pro.
GPT-6.2 Astra is better at letting the reader discover what a story means. GPT-6 Astra often builds toward a familiar turn: the protagonist recognizes a mistake, acts to put it right, and names the lesson. In 6.2, several characters can shape the outcome, and what one of them does late in the story can change how we understand an earlier scene. Its endings need less explanation, and there’s more room for mixed motives, surprise, and humor.
Astra 6.2 has recognizable habits of its own. It gravitates toward family ties, skilled handwork, and the things people learn by watching someone they love. It often ends with a small, practical gesture and trusts the reader to see what has changed between the characters. That restraint gives its best endings emotional weight. Across many stories, though, the pattern becomes familiar. The plots can still surprise, even as very different briefs lead back to the same intimate, old-fashioned world.
Each brief asks for a 600-800-word story in which ten required elements shape what happens. Three judges from different model families evaluate each new pair in both presentation orders, excluding the writers’ families. The leaderboard now covers 67 writer models, 134,964 judgments and 20,589 matched story pairs
https://t.co/ZVHIft5kLq
There are opinions and there are facts. What I'm about to tell you is a fact. 100%.
I went through the logs of one of the big American AI companies and found something that explains what we've all been confused about for a long time.
Have you ever wondered why people are so split on the same subscription and the same service? Why one person praises the speed, performance and quality, and the next says the exact opposite?
Today I'll tell you what's really going on.
This company serves the same model in its subscription through three providers.
That's not a problem on its own. It's normal. There are deals happening in the background and capacity math they have to deal with. Nobody blames them for that.
What's wild is that the three providers don't serve the model at the same quality. Some serve it slower, with a smaller thinking budget, and it looks heavily quantized so it's cheaper to run.
That's where the manipulation starts. The company decides which provider you get the model from, and on what setup, based on its financial situation and strategy. So it splits subscribers by plan, by country and the location you're using it from, or through random discrimination you know nothing about. The result is lower speed, a smaller thinking budget, or a compressed version. Sometimes all three at once.
That's a complete lack of transparency.
The second case is a big Chinese lab I've used myself.
This lab gives full quality to only 30% of users. They get priority on speed and way more generous usage limits. Everyone else pays and gets robbed in broad daylight.
And that's where the misinformation comes from.
Someone posts that the plan doesn't even last them an hour. Then someone from the 30% replies that it lasts them days. They got a perk without knowing it, so without even realizing it they turn into a blind defender and hype man, loyally defending these companies.
And I'm telling you a lot of labs do this.
If you think this is a conspiracy theory, the quality gap between providers serving the same model has been measured publicly. In one published audit, tool calling accuracy for the exact same model ranged from 100% to under 72%. The only difference was the provider.
That's why this stuff is infuriating. You had a good experience and got lucky. That doesn't mean someone who had a bad one is lying. Don't jump in calling them a liar, defending like crazy and getting angry. Take it easy. It's not worth it.
We're talking about companies that can cut you off from the models and from access at any time, and slash usage limits out of nowhere with zero notice, even if you paid. We've seen it happen again and again, even to people on the most expensive plans.
We're used to paying for something and knowing what we get in return. Not here. Here you pay and hope things are fine at the lab, and that there's no upheaval or changes on their end that wreck your workflow.
That's why a lot of people, me included, keep jumping between AI tools. They change things whenever they want, fast, and we never feel any stability. Just look at how people moved from Codex to Claude over the last two months, after months of it going the other way.
Bottom line awareness matters. As a community we're supposed to stick together and stay united, and not let these companies play us. We recently saw a company announce it was cutting the limits on its most expensive plan in half, then back down after users got angry. Our voice makes a difference.
This transparency is what I'm after, and it's what I'll put my voice behind.
I know a lot of people will ask me to name names. But I'll say it again. I'm here telling you what's happening. Want to check? Run your agent and start verifying yourself. But don't ask me to get into a legal fight so you can get a juicy headline.
Being committed to transparency doesn't mean I'm willing to put up with pointless stuff.
Model latency significantly impacts user satisfaction. Surprisingly, @OpenAI's speed decreases noticeably during the work day, while @AnthropicAI's doesn't at all.
Almost all benchmarks measure latency at test time, but few look at how API response time performs live on production.
We looked at 200,000 LLM requests submitted by users on DesignArena, at every hour of the day, sent to each lab's standard API.
OpenAl model speeds decrease significantly during the work day, fluctuating up to 16%. Even more interestingly, Anthropic's models don't seem to be affected at all.