Stephan Gomer

@49agents I believe this is something that may work, however I still prefer having a “central” model as default for general tasks It works for me since for the stuff where, let’s say, Opus outperforms GPT, I can just create a skill for GPT to ask Claude Code to do it 🔮

@s4dfun i switch between models too. codex for architecture, claude code for implementation. architecture is a long-form thinking task, implementation is a do-it-now task. the models fit different phases. the benchmark chase is a trap

Every time a new model for agentic coding comes out, I spend a full 24 hours testing it on my own tasks GPT-5.5 is the first model ever to convince me to switch from Claude to Codex. There is one reason: (1/4)

I think the rest will notice it soon too OpenAI actually has a chance to take the lead in agentic coding The main thing now is not to repeat the mistakes of Anthropic, who lost the trust of most engaged engineers

GPT-5.5 in Codex seems to be the first model that actually follows through with research to the very end, without getting lazy or making things up I've been using it every day since it launched, and it hasn't lied to me ONCE

@theo So the wall is 57?

For the first time ever, all three major labs are tied on Artificial Analysis

@CS2_Confessions Tuscan 🙏

""we want cache" is just another "TikTok 12yo kids trend". if Cache gets added no one is gonna pick it after the 1st week. its just gonna be another Train, everyone "wants it", but when it actually gets added, no one is gonna play it"

@AlexBores But dude, you literally wrote this post using ChatGPT

Computer scientist and new dad here. I'm raising a child because that's the joy of life. I'm running for Congress because my son—all of our kids—deserve AI that enriches their lives and works for them, not against them. Altman's mission seems to be unfettered AI at all of our expense—he's completely lost the plot. The human experience is a blessing, not a burden.

@Hesamation For me, the question is not why this guy succeeded, but why Apple didn't. This is literally an LLM with tool calls; it was achievable in 2024. The only thing that changed is that models are now better at multi-turn scenarios.

TLDR; open source built a better version of Siri that Apple ($3.6 trillion company) was sleeping on for years. welcome to the AI era where a dude and a repo fills in the cracks of billions dollar industries.

@damianplayer Don’t understand this hype, for me it’s really laggy and unstable. Probably right direction, but still very early.

@chongdashu Clawdbot is still a wrapper around a cloud model, so you don’t need much local power to run it.

Is this good enough for clawdbot

@eurofounder First tweet from this account that is actually true

My wife woke up yesterday jumping from excitement She just received a job offer from Google "Babe, they want to pay me $300k, $100k base plus $200k in stock options" she said I sat down and held her hand "Sweetheart, that means they are offering you just $100k" "No look, it says total compensation is $300k" I gently closed her laptop "Stock options are imaginary money, it's a well known American scam, it's not real" I said "But I could just sell them for cash?" she asked I tried not to laugh, women can be so naive sometimes "That's not how stocks work. Who told you this?" She said her recruiter explained it I sighed. Of course. An American filled her head with nonsense "Baby, American stocks are worthless, it's imaginary shares from an imaginary company" "But it's such a good offer" she started crying "I know, baby" I hugged her, and right there I knew I needed to act It's my job as a husband to shield her from American predators Later I emailed the recruiter and declined on her behalf I explained that my wife is not interested in fantasy compensation from a failing company We are European. We don't fall for American tricks She will keep her job at SAP making €43,000 a year, which puts her above most women in Germany Feels good knowing I protected our future

@ailker @fal Good trick. So basically, you can batch user requests this way? Also, do you notice any decrease in quality or creativity?

Found a way to make nano-banana-pro insanely cost-effective on @fal I built a workflow that generates 9 4K images for a total of $0.21. That’s just $0.023 per image. That is a ~92% cost reduction. The key is grid generation. I added a 'Crop Node' as a workflow utility specifically for this. It takes the grid output, splits it, and then we upscale everything using SeedVR

@claudeai Oh wow. Its way worse today

“AI doesn't need breaks like humans” Meanwhile, Claude:

@Kh5065 Yes!

No account should be under 3K Say "yes", we follow you immediately 🤍👇🏼

@benln deepvote.ai – ask something, and 10+ LLMs will vote for the best decision

What are you building this weekend? Share links, curious to see.

@forgebitz Yes, you can smell if it’s vibecoded or not

always judge a saas by its landing page

@PLBompard Three apartments: one to live in, two for passive rental income

How much do you need to quit your 9-5 forever?

@nummanthinks There is a reason this is not the default. Models are getting dumber with large contexts, and Opus 4.5 is no exception.

If you want to have Opus 4.5 at it's MAX in Claude Code Update ~/.claude/settings.json { "env": { "CLAUDE_CODE_MAX_OUTPUT_TOKENS": "64000", "MAX_THINKING_TOKENS": "31999" } } - double the output - ultrathink always on Ctrl+O will show thinking in verbose mode