Astra 6.2 is the #1 model on our multi-turn coding eval: Vibe Code Bench 1→100.
Other benchmarks test a model’s ability to handle a single prompt. This benchmark, created by Vals, tests the ability to modify an evolving codebase over many user requests while preserving existing functionality without introducing regressions.
Ben retweeted
We’re going to see something common during anthropics ipo, all big ai influencers who bad mouthed gpt 6.2 beyond what it deserved will buy anthropic hard. Bad reviews, differentiated opinions are fine but calling it unusable, refusing to post a full review is total bs. I mean if you look at their posts, you can smell the crap. Everybody is out to make some money.
AND If they truly mean what they say for the right reasons, choose better words next time guys
OPENAI is under attack and i am on OpenAIs side, which side are you on?
Ben retweeted
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 $33.03
From GPT-6 Astra to GPT-6.2 Astra: +17.7 points at max, +17 at xhigh.
Whatever people say about other use cases, it's a code review beast.
GPT-6.2 Sol and more effort levels drop soon in the thread 🧵
Ben retweeted
Two blind code reviews, 26 real bugs — Sol 6.2 found more than Opus 5.5. That's the sentence that matters, not the bench, the real repo. If the cheap one reviews like this, the expensive one has a problem. Still, small sample. I want ten.
I ran two quick code review tests based on real-world work in @Kullevi_com's codebase, comparing GPT-6.2 Sol against GPT-6.1 Sol, Opus 5.5, and Fable 5.2.
Based on these tests, at least, 6.2 seems to hold its own.
Ben retweeted
I almost forgot my favourite GPT-6.2-Astra generation - this is SVG (yes SVG!!) of a peacock - no other model came close to the precision, lifelike nature and detail of this generation, incl. Fable
Ran ChatGPT Astra 6.2 across the World Famous Alex Finn Benchmark Test ™
It scored a 312.7 out of 500. Claude Fable 5.2 beat it by .3
But Astra 6.2 finished the benchmark in a third of the time for a third of the price at a quarter of the tokens
The panicking influencers never seem to take into account efficiency and price
This is a great model
Ben retweeted
Theos 1300$ AI course has 0 graduates and 5 stars. The review is from a different course. Right under it sits an OpenAI logo and the site never says why. In the code the whole section is called "Social proof"
Ben retweeted
Theo told you not to buy his 1300$ AI course. He is right.👍
They sell it as Joel Hooks nervously pitching a guy who doesnt do courses. The public repo of Joels company shows them building devrel dot fyi in November 2023, the paid DevRel course Theo announced that fall. Joel pushed fixes himself, they archived it a year later, and the page still collects emails.
MEGA is built by BRAVE, a polish course company that calls its american partner Joel Higgs😭 on its own site. The architect is a guy named Adam, and the site says his team develops the core lesson content while the four stars add insights. The repo students get has sixteen commits, most of it for Theos week, all sixteen from Adam. Theo only shows up in it as a made up character in the sample data.
Angie Jones left the teleprompter script of her first lesson in her public repo for a day. Every sentence written out with stage cues, target 30 minutes, the script runs 20. Twenty lessons at that length is under ten hours of video for 1300$.
Under the "you should not buy it" post Theo dropped a link for your boss. It ends in ref=theo and the sites own code calls that affiliate capture, cookie lasts a week. The same code carries a last call popup, a "price goes up in" timer and discounts for poorer countries. No boss needs any of that. That same night he wrote that if the course bombs he has to let go of some of his 18 employees. MEGA went from two thousand buyers to three thousand in ten days.
Theo only rented out his face for a quick BAG
Ben retweeted
Picard taught me more about leadership than any MBA could.
He never raised his voice to be heard. He listened before he decided. He let Riker, Data, Worf argue him out of positions, and then owned the call completely. When he was wrong, he said so in front of the crew. When he was right, he didn't gloat.
And when he had to stand alone — against Starfleet, against the Borg, against a Cardassian with four lights — he stood.
That is the whole job. Build people who will tell you the truth, then carry the weight yourself.
Patrick Stewart played it so well that a generation of us mistook a character for a manual. Turns out the manual was fine.
A great actor, and a Yorkshire kid who started as an obituary writer. Today the obituary is his.
Make it so, Sir Patrick. Rest well.
The switch to using @orca_build has completely changed my workflow overnight. S/o to @joaopfc for getting me hooked on it!
Now I just run a Director agent that opens other agent personas in tabs with worktrees, instead of sub agents.
No more micro-managing multiple terminals, the Director handles the back and forth for me.
Ben retweeted
Orca is a banger. Exactly what an agentic ide should be.
It brings my macbook pro, windows pc (for gpu work) and mac mini (for automations) into one app, along with my harnesses: codex, claude code and hermes.
I do most of my daily work on the macbook. The windows machine handles blender renders and small local models. The mini runs hermes agents and automations day and night.
I’ve been using codex and claude code since launch. Their desktop apps replaced vscode pretty quickly for me. I was also using herdr to connect my machines over ssh and that worked very well.
Orca brings those things together, including my linear and github tasks. And the mobile app is excellent. I can check on agents and reply from my phone when I’m away from the computer.
And it has a lot of other great features: a resource manager, an embedded browser, artifacts, subscription usage tracking, a caffeine-style mode to keep the computer awake while agents work, and many other things
Really well built @orca_build
Kimi really broke something.
Now even when you have something good to say about them, people assume you got paid. The campaign backfired so hard it poisoned honest praise along with the paid kind.
That's the real cost of undisclosed promotion. You don't just lose credibility for yourself. You burn every independent voice who would've backed you for free if you'd earned it.
A message to every lab: let your product speak for itself. If it's good, we'll say so. If it's not, no amount of money will fix that.
Ben retweeted
Remember when I told you not to trust anyone shilling Muse?
This is why.
Every AI lab does this. They send creators talking points, narratives, comparisons and claims they want pushed.
I’ve had AI labs approach me with this stuff too. I’ve always refused.
And I have 2K followers.
Imagine what they’re offering the actual big influencers you see talking about these models every day.
Test the models yourself. Trust benchmarks you can reproduce. Treat everything else as marketing until proven otherwise.
Post 2 exposing Kimi AI
That's the creators campaign docs I got from Kimi AI after their Kimi K3.1 pro release and thats the narrative they want creators to push, of course I'm out and I'm never gonna say something I don't test myself and give my honest opinion about and I don't want to be involved in this China vs US political situation.
why are we still taking benchmarks seriously
kimi k3.1 pro is a perfect case study on how useless they are right now, everyone i have seen who has tested the model agrees its nowhere near where the benchmarks suggest
its a complete waste of time without a full refresh imo
Ben retweeted
So basically Moonshot is a massive scam?
Post 3 Exposing Kimi AI
Here after I refused to comply and started criticizing a creator who's spreading their propaganda they texted me again asking to delete a comment I had i couldn't even believe it like who the F are you to ask me to delete a comment I don't work for you and never will with any scammy company like this.
Ben retweeted
I'ma say it like it is. Kimi K3.1 Pro is fucked. It is shit. It is ~GPT 5.6, and worse than k3 imo (in terms of taste, colors and 3d builds). Obviously the benchmarks are benchmaxxed as hell. There was always a little bit of benchmaxxing going on. This is pathetic and greedy. Just look at the charts that don't make any sense. 10% more output tokens, 40% better performance. This isn't the log scale I would expect from benchmarks. This is more like whatever. Holy fuck. I would choose DeepSeek v4.1 flash all day. Not because it is 40 times cheaper or 4 times faster or less benchmaxxed or just MORE EFFICIENT AND COOLER AND AURA AND 🐳. but because it is fun to use. Using k3.1 is pathetic and sad. I feel worse than before I used it. And the usage limits are bad, they are basically free tier. For 20$.
Ben retweeted
K3.1 Pro by @Kimi_Moonshot on KingBench 4.
Leaving openness aside, This is a bad model.
It scores the 6th position with 53.7% and costs $31.88 which is jut $16 cheaper than Opus.
You can get better results than this with Haiku or Sol for like $25 cheaper.
The whole pitch of cheap looks like BS right now.
The model constantly re-reads cache and costs way more than most models and it also under-performs on my bench.
It seems that they found out a way to make a model cost more by basically making it read caches more often to recover their costs (which is also why cache reads cost the same as Opus).
In no way will I tell anyone to use it.. The cost to performance is really bad.



































