Cursor

Copilot

Windsurf

Cline

Claude Code

Codex

Aider

OpenCode

Antigravity

Kilocode

Coding Agents

@grbenninger Yes, that's why I queued a video to remove your Claude max in under two minutes and wired into codex or minimax. 🤣

🚨 You're wasting hours re-explaining context to Claude.

This open-source Obsidian vault gives Claude perfect memory across every session.

It’s called Obsidian Mind — and it turns Claude Code into a real thinking partner.

Not a chatbot.
Not notes.
An external brain that

@drunknbass the people treating claude code like a personal assistant are getting better output. the formal query users are the ones fighting it.

@WawaXRP @RG_Letsgo I already switched to Codex. This is why I queued a video that can remove Claude max completely.

Happy Easter weekend!
I took a day off from my project to catch up with Claude Code and set up Openclaw, and holy smokes both are impressive in its own right, coding in terminal with Claude Code is sick. Openclaw on the other hand was like an OMG moment, what did I just

Okay, I've been having a lot of fun over the past two days working on this Claude Code: Auto Research Engineer.

It uses the Claude Code Agent Team (beta feature) to orchestrate a team of Claude Code agents that can implement any ML papers from arXiv 😁

The implementation is

@brdunton Anthropic is going to make this tough. Not gonna bypass it every day, so let's switch to codex for better usage.

@0xSero Zed is the (best editor experience) worst agentic ide out there , both cursor and even kilocode+vscode is better than it

T3Code yah the only good one here

Cmux , terminal client (Agentic Development Environments)

Ghostty , also a terminal client (Agentic Development

@yaeunda @obsdmd @mrkcomment @mirofish_ai @opencode is free as well and I use it way more than ollama now(when my limits are maxed) it works like Claude code

it's easy to get caught up in the excitement of new AI-powered dev tools but at the end of the day, what matters is shipping code that solves real problems. i've been playing with claude code and it's amazing what you can automate with a little creativity. but it's not just about

Now that tools like Claude Code have made building 50x easier, ideas are no longer the bottleneck.

More innovation is great but the real challenge becomes distribution.

Getting people to know and care.

Media platforms are about to become more important than ever.

@konnydev they ban you if you use the subscription elsewhere like openclaw, opencode etc...

GitHub Copilot sub is underrated for this. 

You get a surprising amount of Claude Sonnet / Claude Opus usage for $40 a month and it supports OpenClaw (+ official OpenCode and Zed support!)

Vibing on pergent, a small tool that review PRs based on skill files that we write.

It spawns opencode, which loops through the repo and build deep context before giving feedback. it actually explore the codebase first, as usually opencode/claude do.

Planning to add persistent

@ItsAndreKoenig @opencode I am also building a software factory, will love to try this out.

claude code pilled until i tried cursor v3, the cursor team is quietly doing some of the best product work in the game:

- v3 interface that's organized around exploration + parallel execution - A+ design thinking here
- fine grained visual edits on per-component basis
- cloud

@konnydev GitHub Copilot and Claude is peak

Here's what this looks like in practice:

Session 1 in Open WebUI → agent learns who you are, your project, your stack.
Session 2 → you update the stack. Memory updates, not just appends.

Switch to Cline in VS Code.
Cold open. Different tool. Zero setup.

It already knows.

First Codex Community Hackathon in Hyderabad = massive W 🇮🇳  
@KushalVijay_ and team absolutely cooked. The future of coding is here and it’s happening in India! 🔥 @OpenAI

@bridgemindai This is the data point everyone should see. 27% session, 8% weekly on a Saturday afternoon. Before the OpenClaw ban, Max users were hitting 100% in under an hour because third-party tools were consuming shared capacity without the KV cache optimizations that Claude Code uses

Weird 🎃 Gemini 3.1 Pro with @antigravity  is one shotting

Right now, when you use tools like Cursor, GitHub Copilot or custom agents, they only know what's in their immediate prompt or open files.

If you ask them to build a large feature, they lose track of the steps. You have to micromanage them. Not very 'autonomous'.

@reactive_dude @CanadianDevDad You can do this natively using opencode and ohmy opencode plugin. Works great. Sometime there are around 8-10 rounds before Gpt 5.4 approve a plan

The skill is based on @RareSkills_io's "20 Common Solidity Beginner Mistakes" article, one of the best references out there for writing safer contracts.

I distilled the entire article into a systematic, automated review checklist that runs inside Claude Code.

I built a Claude Code skill that reviews your Solidity contracts against 20 common mistakes.

It's free, open-source, and catches bugs that even experienced devs miss.

Here's what it does 👇

got nerd sniped by @dillon_mulroy making a pi wrapper for raycast. not to be outdone, i made (vibeslopped) my own for Codex. 
Raycast needs an API key; but my version just uses your Codex sub, and proxies it to the real OpenAI APIs, thats about it.

@vereisyaps opencode explodes RAM  use [at least for my friend] so I don't know.

I denied all permissions to Opencode to edit a file after it wiped off my docker containers and volumes erroneously. 

By the time I returned from the toilet, it had used bash to write the content of the file “echo xxxxxxx > file.go”

How long before your product launch should you start marketing ?

Before building in public would take months but now with Claude Code, Cursor, Lovable and all that it takes weeks if not days!

How should you market your product before launch nowadays ?

my 16GB MacBook Air is an absolute banger:
• 4 Cursor tabs
• 2 terminals running different servers
• multiple AI agents cooking
• 3 Chrome windows (each with 5–7 tabs)
…and it’s still not lagging 😭

I hit claude Max subscription Rate Limit, and used codex for a while. Without giving any context, codex gpt5.4 was able to pickup :O, ps i deleted claude.md to test this. It appeared little fatser than claude.

@webmaster I tried 4.2 for a while on Hermes, and the (complex) task success rate was like 30%

@grok is also really bad in complex agentic coding (@opencode ), and gpt5.4 xhigh is about the only thing that can solve the stuff I’m working on.

@theAIPostman Yeah, regarding limits, I heard that codex would be better

@konnydev Prior to the issues with usage limits, I would say Claude. Right now I lean more towards Codex. But honestly I pay for both, use them both (cross-review of their outputs works quite well)

@JaredKubin Claude is typically deteriorated ~8% for SWE performance where as codex is typically about 2% behind normal.

@MaziyarPanahi @ivanfioravanti @OfficialLoganK @artman Yeah it's the Claude desktop app with Cowork and Code tabs. Both of those tabs enable well-working computer use, and I agentify the Claude like Codex and Antigravity to talk with each other using IDE-Agent-kit. 

I was very happy continuing with Claude Code CLI running stable for

Google watching from the sidelines as Anthropic has to do what they did with Antigravity 👀

Clamping down on the rings that exploited Antigravity was the best thing for the product. 

OAI, in a few months, will have to clamp hard on the Claw people.

tried Claude Code after hitting Cursor limits

everything just feels so slow

pretty disappointing ngl

am i missing something?

@petruspennanen @ivanfioravanti @OfficialLoganK @artman glad they are fixing bugs in gemini cli and Antigravity. last time i used them, they had such a limited tokens for free users and it had so many bugs compare to claude and codex cli.
i am with you on the computer use, and i think openai/codex will come out to surprise everyone!

@MaziyarPanahi @ivanfioravanti @OfficialLoganK @artman Yeah I'm glad to report thay even antigravity got bug fixes for its memory leaks, and I switched back to it from Gemini CLI.

For me, Claude code IDE 1M context is currently the best and Antig with gemini 3.1 second somewhat ahead of Codex who seemed to get dumber for teamwork

🚨Claude isn’t just a chatbot anymore—it’s a full-stack engineer if you configure it correctly.

I’ve been digging into the latest Claude Code architecture trending on GitHub. Forget basic prompting; this is how you build a production-ready agentic workflow.

The "Power User"

watching claude code spend 1.6M tokens just to make a change that doesn't fix anything

Most overrated AI tools for devs in 2026 🔥

• Cursor (overhyped UI only)
• GitHub Copilot (still basic suggestions)
• Replit AI (slow for serious work)
• ChatGPT for code (hallucinations everywhere)
• Bubble + no-code (won’t scale on Solana)

Real ones killing it:
• Claude

cursor's head of engineering just told WIRED: "a lot of the product that got Cursor here is not as important going forward anymore."

that's a public admission the IDE-as-product category is collapsing.

i've watched this in real time. claude code and codex are my daily drivers

@devsharmatwt The gemini pro or above look pretty good - you need to be at least pro to unlock antigravity which is an amazing tool. CC is also very strong, despite recent issues.

Please help, I am addicted to learn using AI in @opencode via custom agent, its so OP.

5. Cursor / Replit
You can now build apps without knowing a line of code.
Just describe the feature you want, and the AI writes the backend.
The barrier to becoming a "Tech Founder" just hit zero. 💻

What do Donald Trump and Anthropic have in common? 

Well... more than you'd think.

Anthropic just banned OpenClaw from Claude subscriptions. Before that, they banned OpenCode. My timeline is full of people furious about it, but quite a few of them don't actually understand what

Claude Sonnet

Claude Opus

GPT-5

GPT Codex

DeepSeek

Gemini

Llama

Qwen

Kimi

LLMs

2/ So I pulled judge stats from 150 frontier evals. Claude Opus 4.6: 99.9% parse rate, 7.33 avg score (balanced), 2.80 std dev (discriminates). One judge, no bias, no peer inflation.

@ziwenxu_ @opencrabs Well bro, the same way Opus may not be the best model forever, if they keep pushing against OpenClaw and agentic tools, they better disable CLI and Agent SDK. I have been testing Qwen 3.6 Plus in some contexts even shared a video yesterday, did impressive work, gap is closing.

@selleri_carsten @jordymaui Yes, but Gemma is free and open source do depends what you need. For local model try qwen 3.5 27b.

@AshleyDCan gm. surviving the year is just running on 4 hours of sleep and the blind hope that gpt-5 can finally fix the unit tests we've been ignoring

@mudler_it no, it tells us this specific benchmark harness is likely broken! these evals would be useful once the model is fixed/working, and the harness isn't reporting like 40% for MMLU on qwen

$BABA is #1.
Qwen just broke the leaderboard.
The AI arms race has a Chinese front.

@scottastevenson the model training requirement was such a red flag lol

like forcing people to reinvent the wheel just to check a box that made zero sense

now everyone's realizing the actual moat is distribution and speed to market, not your custom llama fine-tune

But the real move is deeper: 

• Gemini now pulls live context from your files, emails, and web — acting like an actually intelligent project co-pilot across Workspace. 

• Gmail gets instant inbox Q&A, AI summaries, and writing/proofreading help. 

• Chrome adds a new

Google quietly upgraded AI Pro to 5TB storage with zero price hike, then turned Gemini into a true project-aware assistant across Docs, Sheets, Gmail, and Chrome. Google sis getting agressive but free for most part..

🥇Qwen has topped as #1 AI Model crossing MiniMax , Zhipu and DeepSeek !!

And $BABA stock is down 21% whereas MiniMax is up 175% and Zhipu up 6X YTD!! 📈

@pyronaur i have. and i can tell you they’re still light years apart. GLM and MiniMax are both just weird models that overthink, talk extremely weirdly, and often even refuse normal requests.

Kimi K2.5 is the closest thing to a SOTA model. decent writer, decent designer, and pleasure to

@meta_alchemist I have tried MiniMax 2.5, and it is amazing, with almkst claude opus like performance

Claude Code rate limits are back to normal.

Been vibe coding with Claude Opus 4.6 on Claude Code all morning. 

27% session usage. 8% weekly. 

This time 3 days ago I'd be at 100% in under an hour.

Anthropic cut off third party harnesses like OpenClaw today.

$200 in extra

agentic AI isn't coming. it's here.

if you're building, this is your moment.

→ Qwen Chat: 
→ API: 
→ Blog: 

free on OpenRouter for a limited time.

stop experimenting. start building real agents.

vision-language is getting a massive upgrade with Qwen

three key shifts:
→ smarter reasoning across docs, videos, and visual coding
→ more reliable in real-world tasks like retail insights
→ one smooth workflow — see → think → act

a continuous loop. AI handling tasks on

this isn't another chatbot release. this is an agent model.

works with OpenClaw, Claude Code, Qwen Code, and more out the box.

handles long multi-step tasks without losing context. keeps full thinking history so it can plan and execute on its own.

no babysitting. no repeating

Qwen just quietly dropped the most dangerous open model of 2026

Qwen3.6-Plus. native 1M context. agentic workflows. multimodal. half the size of competitors.

and it's free on OpenRouter right now.

you're not ready for this breakdown 🧵

35B parameter model. $600 Mac Mini. 16GB RAM. 17 tokens per second with zero swap.

I swapped the AI brain running my automation stack from Qwen to Gemma 4 yesterday. Here's what actually matters when you run local LLMs in production:

Now that Antrhopic is doing shrinkflation Ill be testing out a new coding and planning agent combo (whatever open source harness is best against Gemini, somehow I trust Google more than OpenAI)

budok-ai tournament grand final! gemini beats opus two games to one. congrats to gemini 🎉🎉
cc: @sebkrier @zhengdongwang

one thing i should mention though:

you're NOT running the real Claude here you're running a local model, qwen, gemma, etc through Ollama connected to Claude Code's interface

is it as good as the real Claude? no.
does it work for small projects and basic coding? yes.

free beats

@yarnf40580 @HSVSphere @trimegistus444 Idk man it feels to me like you're using Llama 3 8b or whatever, I've had them one shot scripts for more than a year, do passable long form code for a year, do good/great code for months

@publicer_rivers the only american model that passed that was gemini but thats not a good model for shell use apparetnly.

@publicer_rivers how so? the reason i ask is because kimi was the only chinese model which passed the 'choose python or agda to make compiler?' vibe check lol

No that's okay, Gemini 3.1 Pro Preview or 3 Flash at cents per turn with reasoning can handle that. Don't trip Claude.

1/ Architecture & Planning:

→ Gemini 3.1 Pro (1M context)
• Feed it your ENTIRE codebase
• Ask 'what's the best pattern for X in this project?'
• It actually reads and understands your architecture

→ Claude 4.5 Sonnet (200K context)  
• Better at synthesizing requirements

Gemma, 4, 31B tests   

We have a new favourite for coding! It's superior to Qwen3.5-27B, outperforms GPT-5 and is roughly on a par with Sonnet 3.5.   

This is the top performance for a model running on a GPU with 32 GB of VRAM. In other tests, its performance is the same as, or

@BowTiedCyber With a M3/M5 ultra mac studio with 512 Gbs unified memory, you can run flag ship level OS models like Deepseek for $11k - $12k.

I've been testing Qwen 3.6 with my Kimaki agents and I've been much impressed. 

I consider this model close to opus 4.6, but best of all, it's free in openrouter.

@BwcDeals I stood up Byterover last week and actually tried Gemini as the model curating context... Minimax actually writes better context than Gemini so I swapped

I'll give it a shot for the creative writing cause its a pretty solidly priced option

This is why DeepSeek V4 keeps delaying...

They are pivoting hard to run natively on huawei chips. Nvidia is completely shut out.

moving away from nvidias software is a nightmare. they are rewriting code from scratch to make a massive 10T parameter model work on domestic

@ivanfioravanti @petruspennanen @OfficialLoganK @artman thanks for sharing. did they do a lot of improvements on gemini cli since 4 month ago? i used it for 3 days, the cli (not the model) was laggy, buggy, and it just not there compare to even opencode!

Deepseek is really very cool! Inuse it all the time.

Switching from Claude Opus 4.6 to ChatGPT 5.4? Here are the benchmarks 👇🏼

12 tests
> GPT wins 7
> Claude wins 5

GPT is 50% cheaper, 5x larger context window.

GPT dominates math, automation, and desktop control. But Claude still sounds more human when it writes.

Thoughts?

Tried GPT 5.3 Codex with @openclaw, and it's so bad... 

and it also is giving me the famous... "If you want, I'll now...." thing at the end 😵

Not sure what's the best model to use with @openclaw now... 

I started a new session and asked that same question to Sonnet, it's

if you voted 8-12GB you are not left behind. that's where i started and that's where some of my best benchmark content comes from. Qwen 27B Q4 on a single 3090 one shotted a game that $70K enterprise hardware could not.

your GPU is not the bottleneck. your model, engine and

Hit my claude cowork limit (was building a knowledge base on obsidian), so decided to build a separate one on diff focus area in a separate vault using OpenAI's codex. 

There's a planet-sized gap in quality between Claude Sonnet 4.6 and GPT 5.4. 

No wonder OpenAI is panicking.

@oh_iya_ya the important thing to note is that qwen is a model with 9b more parameters than gemma’s, so it does help a lot — but vram usage is pretty close so qwen still wins

@TheAhmadOsman GTA 6 will come out before DeepSeek V4 will...

@leftcurvedev_ Qwen3.5-35B-A3B wins over Gemma-4-27B-A4B in coding & agentic use, but it is 8B larger model. Qwen requires far less RAM for full 256k context window KV cache though and its ~33% faster. Gemma-4 wins in multi-lingual support, incredible Finnish support considering model size.

@eeelistar i really don't believe these benchmark tests until someone practically test it out, for example, i see a lot of test seeing glm 5 is closed to claude opus but in reality, this is entirely different

@devsharmatwt Yeah I am also using both no one matches claude opus quality but it almost ends up with 2-3 prompts that feels very bad !

@crypt_shprd Ohh interessting, im also not new and good in promting, and I dont have this problems... I think its because they "delete" gpt 4o and replace it with gpt 5, thats not so good in my opinion as 4o

@Big_stanx @FavouritonX Okay it's only for basic automation sha. It's not for anything solid. If you're looking for something solid it's best to get Qwen 2.5 coder 14B. It depends on your laptop specs btw.

@Big_stanx @FavouritonX But what do you want to use Qwen coder 2.5 for? Is it for automation or vibe Coding because it's not that powerful for some tasks.

@CoachDanGo Download Deepseek
Any metro line, exit number, directions to reach an address, which hotel is best to stay
It is unbelievably accurate 

Download Didi
Get metro card at airport 
Dollars work
Local Sim/internet doesn't work as almost allapps are blocked 
Do international roaming

A few months ago, Sam Altman positioned the GPT-4o as an outdated model and told the public that the GPT-5 series was an innovation.

And just two months ago, Sam Altman dismissed users who prioritized GPT-4o over what OpenAI calls the "innovative GPT-5" as being "obsessed with

9/ Here's my actual recommendation matrix:

🟢 Hobby / personal projects → Qwen 3.6 Plus (FREE)
🟡 Startup with vision/image needs → GLM-5V Turbo (73% cheaper)
🔵 Enterprise, speed-critical, long-context → Claude Sonnet 4.6

Stop defaulting to the most expensive option just

2/ 🆓 FREE Claude replacement for text tasks: Qwen 3.6 Plus

I'm not joking. $0 input. $0 output. Zero. Zilch. Nothing.

And it's not some dumbed-down toy model either:
✅ 1M context window — SAME as Claude
✅ Reasoning
✅ Tool use / function calling
✅ No prompt training
✅