Cursor

Copilot

Windsurf

Cline

Claude Code

Codex

Aider

OpenCode

Antigravity

Kilocode

Coding Agents

this week - 70 ppl flew in to cloudflare atx, probably like 5% of them traditional engineers

we taught them opencode, they're building, and now today they're writing skills and publishing/sharing to an internal skills library

REALLY cool. i like my job, it is fun

Codex was a champ and killed it immediately with whatcha call it no permissions whatever the fuck

anthropic accidentally shipped their entire codebase lol

A 59.8MB source map file was included in @anthropic-ai/claude-code v2.1.88, exposing 1,900 TypeScript files and 512,000 lines of proprietary code.

Security researcher Chaofan Shou found it March 31, 2026.

@mikeydsoftware Unpopular opinion, every developer was regarded untill codex released.

@ErikVoorhees yo @kilocode @opencode how come Voorhees got @grok 4.20 multi-agent working before you 🤣wow stop fail

@boringmarketer claude code is so weirdly polite about running out of context. just say you're at the limit bro. i don't need the emotional check-in

Hermes + GitHub Copilot Pro+ is amazing. 

Dont have to worry about API costs KEK

@icanvardar I tried it on linux and the cursor wasn't updating his state when editing file or trying to resize pans so I deleted it

bro created an AI job search system for Claude Code that scored 700+ job applications and actually got him a job.  

 AND IT'S NOW OPEN-SOURCE.  

It scans multiple company career pages, rewrites your CV per job, and even fills application forms. The repo has: 
> 14 skill modes

wait, so Claude Code Opus 4.6 can now:

-scan a full website
-turn it into a mobile app
-get it ready for the App Store
-and keep it updated on its own

with no human involved?

that's crazy...

I don't know why the "queue" in Claude Code means "interrupt this task and it doesn't get resumed" while in Codex it means, you know, queue.

@md_kasif_uddin github copilot. the inline suggestions really speed up my coding time. what do you prefer?

@md_kasif_uddin VS Code + Github Copilot is the ultimate thing 🔥

The real secret to building a successful startup?

First thing you need to do:

Find yourself a cofounder in the top 0.01% of Cursor users.
(Image for proof — IYKYK)

Actually... it might be Codex now.

In 3 months, it’ll likely be something else.

The tools change fast.

But the

Just got GitHub Copilot pro for absolutely free for next 2 years  😋

GPT 5.4 + composer2 on @cursor_ai is a deadly combo

@ikishankewal Here to see the answers 👀

Tried antigravity and I was like wowww

I'm vibe coding a new *GPU rigid body physics engine* from scratch for @levelsio 's #vibejam ! 💥

- Supports Rust & Web (via WASM+WebGPU)
- 100% written by @anthropic 's Claude Code + Opus 4.6
- Performance optimized by @karpathy / @uditg 's autoresearch:

at the end of each prompt in claude code, start adding this line, “you can ask me questions to clarify” 

you’ll thank me later

I built a simple lead nurturing system for MyCircle app. You go to the landing page, fill in your email address and get a sequence of emails over 1 week.

Now, I’m expanding the scope to follow up on leads and get their feedback. All done with Claude code (on Antigravity) & n8n.

15/ END OF BRAIN DUMP!
That's how I'd run marketing with AI. One knowledge layer, specialized agents across every channel, everything cascading from the same strategy.

We can ship code at lightning speed now. Cursor, Claude Code, Codex let you go from idea to product in days.

I tried @antigravity today for the first time and must say I was impressed by its user friendliness. It is much easy to make the agent get things done on it than it is on Claude or gemini-cli. 

I was just wondering how it would be to let it work in the background? I mean, one of

I'm impressed by @opencode Big Pickle model so far

- insane free quota (I didnt even sign up yet)
- the terminal UI is a fresh air from Claude
- the coding result so far has been clean and straight to the point
- token consumption seems relative low

🚨 BREAKING: I built a second brain with Claude in 2 days

Claude Code × Obsidian × 3 prompts, literally remembers everything for you 24/7 for FREE

@Kappaemme1926 As we speak, my man! Works very well. Here is a good youtuber that i found useful regarding using Antigravity: 

Long ass video, but very insightful stuff!

People are still scared of using claude or codex in the terminal. Most non technical people have a fear of the terminal like they’re going to slip into the matrix or something. 

I’ve shown older aged non-tech people how to use claude code and they’re up and running in no time.

@benln this is cool. the fact that cursor hackathons are happening in Ghana says more about where AI dev tools are headed than any market report. builders everywhere, tools finally catching up.

@Kappaemme1926 Well, I think the pro plan works well. Antigravity, plus Google gemine is there as well, and some claude comes as a plus.

@Ghido Ah got it!

Opencode Go is pretty cool too with glm 5.1 $10 a month :)

If you're a Nigerian and you know and use any of these:

OpenClaw, 
NotebookLLM, 
Cursor, 
ElevenLabs, 
Descript, 
HuggingFace, 
Codex or 
Perplexity. 

Congrats!🎉You belong to the 1% and you're 10years ahead of everyone else. 🙌

I am considering giving open-models a go like Gemma 4 with a coding harness like Pi Agent Coding Tool.

OpenCode is not something I want to use.
KiloCode I may look into if Pi Agent doesn't do want I want.

Im not that interested in Grok but I might do a one or two off video

Real-World Test: Claude Is Just Too Expensive!

Anthropic dropped a bomb on April 4th — they suddenly killed off all third-party tools relying on Claude subscriptions: Cline, Cursor, Windsurf, OpenClaw (and over 135,000 active instances) went down overnight. Countless developers

@AlxTurovski Windsurf allows you to switch between various models the same way that Cursor does, yes?

That's one of the reasons I probably won't leave Cursor anytime soon -- I can choose Claude, GPT, Grok, and more just by switching my selection in a select box.

@leostera @kitlangton has nothing to do with ghostty, you just have a memory leak somewhere in your mountain of claude code slop I imagine

@domirosari0 Many are talking about opencode, but I have never used it.. is it really that efficient?

I switched to codex directly

quick update on my Codex setup.

I now have 6 OpenAI Plus accounts running via Codex LB for only $18/month total

this is too good

I turned Claude Code into a web dev tutor.  

It's 100% free and it's not a typical course. 

The agent provides structure, but you can take it any direction you like, learning as you build.

If pure vibe coding leaves you feeling stuck, this is for you.

Someone just open-sourced the ULTIMATE Claude Code setup…

27 agents.
64 skills.
33 commands.
1,282 security tests. 🤯

Planning → coding → reviewing → fixing → optimizing…
ALL automated.

Works on Cursor, Codex CLI, OpenCode.

This one repo replaces weeks of setup.
And

old startup playbook:  
- raise vc
- hire devs
- move to sf
- run paid ads
- spend 6 months building

new startup playbook:
- ship something w cursor/claude code
- build in public from day 1
- grow with okara ai cmo
- move to bali after 1k mrr and tweet like naval

🚨 Devs: You’re UNDERUSING Claude for code reviews.  

This is where it CRUSHES GitHub Copilot.  

Prompt:  
"Review this code like a staff engineer. Check: 
Bugs, 
Edge cases, 
Performance, 
Security, 
Clean code. 
Return severity levels."  

Paste your [PR / function / module]

@Kappaemme1926 I dont use claude anymore because of the rate limiting and i dont see any benefits paying for max (100€) as a student without even using the AI of its fullest 

Now im just using copilot because its free for me as a student with opencode. 

I even used on opencode qwen 3.6 plus

You can now give Claude Code a visual map of your entire codebase 🤯

Someone just open-sourced a tool that builds a local graph of your entire project so Claude instantly knows how every file is wired together.

- Runs 100% locally on your machine.
- Maps out your entire

We built this thing called Rox Dev.

It's a Claude Code skill that roasts you to take a break.

It nudges you to stretch, walk, hydrate, basically everything you forget when you're deep in flow.

Kinda like having a health buddy sitting next to you while you ship.

Been using it

I've now talked to over a 200+ blue collar builders who are building bespoke software with Claude Code.

Last week was super interesting.

> An electrician who creates a panel schedule calculator that reads blueprints and generates load calculations with NEC code compliance
> A

@bquintero Using Claude Code and Codex together for home automation is a great real-world test. The bigger point you raise is key — AI is not just a tool anymore, it is reshaping how we think about problems. The builders who adapt their mindset, not just their stack, will lead.

@aibra Just reinstalled codex, will test it out! Because Anthropic s been playing a lot of games lately

At a Google AI training, I had to agree to broad data collection to use the tool (@antigravity).

I asked: “Do I need to accept this to continue?”
Answer: “Yes.”
Then: “Google already has all your data anyway.”

So let’s call it what it is:
not consent, but coerced compliance.

AI Tool tricks for the day:
The "No-Code" ceiling just shattered. 🔨

In April 2026, you don't need to learn Python to build software. You just need to know how to "talk" to an Agentic IDE.

Using the Cursor + Claude 3.5 Sonnet pairing, anyone can build a custom internal tool for

I have decided to embrace light mode going forward 

But for Tech never my Idle never @code and @cursor_ai will forever remain in dark mode. 

> wake up 
> work up 
> wealth up

Claude Code is moving faster than the tutorials can keep up.

I’ve mapped out the exact path to go from zero to shipping production code.

This Includes:
🚀 3 Getting Started resources
📚 4 Long Courses
🧠 4 Best Practices guides
🛠 3 Tools & Libraries
🐙 8 battle-tested GitHub

Claude Sonnet

Claude Opus

GPT-5

GPT Codex

DeepSeek

Gemini

Llama

Qwen

Kimi

LLMs

HT @lsanger for the RT
I doubt @elonmusk is in any way serious! @AnthropicAI Claude Opus 4.6 leads me to believe these are not groundbreaking books at all. Just provocative titles. Remind you of anyone?

@D27357 And Claude Opus isn't even cutting edge anymore! None of us have access to the current frontier, and that limit will soon enough be breached, too.

Software isn't secure, scarce, or confidential anymore. Every binary can now be decompiled. Every bug exploited. Every proprietary

@almari_mar29581 I totally understand. I tried 5.4 too, but only to test and find out I don't want to build anything with this model. Since 4o and 4.1 retired, I cancelled all of my subs in OpenAi #QuitGPT . For now, Claude Opus 4.6 is the best model able to build a deep connection.

DeepSeek drops always cause a change in the Chinese labs' business strategy
V2 caused a massive price drop, and V3 led to a wave of open sourcing
Let's see what V4 brings
Btw, Kimi K3 will be extremely strong too (Kimi Linear, Attention Residuals)

google finally shipped the obvious thing

a persistent workspace where your files, chats, and instructions live together across sessions

no more reattaching docs. no more re explaining context. no more "as I mentioned earlier"

notebooks in gemini. up to 100 sources. syncs with

why did Groq never add kimi k2.5 and instead just deprecated Kimi K2 and told people to use GPT-OSS instead?

@Snakesan @_winter_wonders i like theo, but his "gpt 5 is agi" video lives rent free in my head and it's poisoned his opinions on stuff that isn't available to the public. i bet what he got to use was really good. it's not what we're getting to use. and the video (i only scrubbed through) appears to be him

@grok run a comparative analysis on this thread. Summarize how GLM-5.1's 'DeepSeek Sparse Attention' gives it a structural advantage over standard post-training models in long-horizon tasks

A 31 billion parameter open model just beat GPT 5.4 mini on legal benchmarks.

Gemma 4 landed 5th overall on LegalBench. Kimi K2.5 placed 6th on the Bar Exam. Three open weight models outperformed OpenAI's latest small model.

The moat was never the model. We wrote about that

I turn KIMI AI in to my 24/7 Personal assistant

It’s genius, 10x faster, and free

Here are 10 professional prompts I use daily:

(Customize and use it yourself ↓)

@TelepathicPug If you can make that work, power to you. I feel like a cripple when I have to use Gemini 3.1 pro now compared to Opus

Best results for me:

1. Opus 4.6 for initial research/planning
2. Opus 4.6 for initial build
3. GPT 5.4 hunts for all the slop
4. GPT 5.3 CODEX fixes it

Repeat.

Nothing is perfect. But results have been consistent for me with this approach.

I just built a $5,000 client website in 118 minutes.

The stack:
- Claude Opus 4.6 for architecture, logic & complex reasoning
- Figma Make for pixel-perfect UI, interactions & deployment

I tested 100+ prompts, so you don't have to.

Here are the 9 prompts that actually work:￼

pwning since 4am with GLM 5.1.

Now having Claude Sonnet 4.6 review the work done the last 48 hours because I can't trust Chinese models all the way yet.

GLM 5.1 + tools beats Sonnet 4.6 at the SWE benchmarks (coding)

Nazlı Şipi and Cem Dilmegani at AIMultiple benchmark 13 LLMs as coding agents across real dev tasks. Claude Sonnet 4.5 and GPT-5.2 lead the rankings.(16/21)

Nick McKenna and Bartek Perz at GitHub introduce Rubber Duck in Copilot CLI. A second model reviews an AI agent’s plan and code. Pairing Claude Sonnet with GPT-5.4 catches silent bugs before they spread.(5/21)

🤖 Google API Keys in Android Apps Unlock Unauthorized Gemini Access

Developers embedded API keys in mobile applications following Google's documentation for services like Maps and Firebase. Those keys were considered low-risk identifiers. When Gemini was later enabled on the

A 3-person startup in Mexico had their Google Gemini API key leaked.

48 hours later... they got $82,000 bill.

Not a hack. Not a DDoS. Someone found their key, used the API exactly as designed, and the requests just kept going.

No rate limit. No spending cap. No alert that

Sydney Sweeney stuns in pastel elegance😍
Made with Nano Banana 2 on Google Gemini app.

Prompt:

A full-body, ultra-realistic portrait of a woman standing gracefully against a light-colored wall with a subtle red gradient. Cinematic, directional lighting with a spotlight behind

@bridgemindai That Qwen 3.5 taking over a minute just to say hi on a five-grand DGX Spark is the kind of detail that cuts through all the local AI hype.

Meta's finally trying to catch up in AI after their Llama models got outpaced by everyone else basically :p 

They spent $14 billion on Scale AI, hired researchers from OpenAI and Anthropic, and now Muse Spark is their first real attempt at a flagship model. 

The "Contemplating"

I replaced my entire team with 10 AI agents.

Not chatbots. Not Zapier flows.

10 actual agents — each with a defined role, specific tools, and handoff protocols.

Koda (COO) runs 24/7 on Claude Opus.
Gideon (Engineer) reviews every line of code on GPT-5.4.
Atlas (Research) does

@airkatakana Today I used meta[.]ai and I found for excel work, it was good (closer to chatGPT's extended thinking). It was fast and accurate also. I feel its good progress by Meta. Gemini as CLI/ harness: I am not happy. Codex: is the best and others AI model's claim from chart are

@tomfgoodwin For what it's worth, with Claude eating tokens like a hungry hippo this last month, I switched my coding LLM to GPT/Codex for a while. 

That said ... if you're using an LLM for anything other than coding and not using Grok, you're missing out.

@theaibench26 Honest admission on the agentic gap. Strong on scientific benchmarks, acknowledged weakness on agent tasks. That's a different energy than Llama 4. What does their agentic roadmap actually look like?

Claude Opus 4.6 hit a hard wall at 3,547 QPS. GLM-5.1 just bypassed it, restructured its own architecture, and hit 21,500 QPS.
The era of "Vibe Coding" is dead. Welcome to Agentic Engineering. A masterclass: 🧵👇

Most people pay FULL PRICE for AI APIs (Free for 48 Hours)

I don't.

I cut my costs by 70% on video, images AND chat completions

Claude Sonnet 4.6 for 65% off Veo 3.1 videos for 75% off One API key. One bill.

And I built 5 custom skills to automate all of it hands-free

The

@HuntingAlphas I use Gemini for advanced daily search, esp anything map-related or current-even related

For complex questions I've been using ChatGPT 5.4 Thinking, but now finding Claude Opus 4.6 does a better job. Any information retrieval. 

NotebookLM to analyze historical financials.

@presence2222 @sama @gdb Same - I have yet to speak with an AI model as smart as early GPT 5.

AI task-completion capability has increased 10,000x since 2019.

GPT-2 in 2019: 2 seconds of useful work. Claude Opus 4.6 in 2026: 12 hours of useful work.

That capability is doubling every 4 months. This is the most important chart in AI right now and almost nobody is talking

Stop using massive models for simple logic.

Run Llama 4 Scout for your basic data parsing and routing.

It is faster, keeps your data local, and costs $0.

Save the heavy reasoning for your complex features.

Don’t pay for ChatGPT, use Claude
Don’t pay for Midjourney, use Nano Banana
Don’t pay for Grammarly, use LanguageTool
Don’t pay for ElevenLabs, use VoiceBox
Don’t pay for Lovable use Google AI studio
Don’t pay for Grok Imagine, use Qwen
Don’t pay for Suno, use Lyria 3
Don’t pay

🚨BREAKING: Your AI is a moral coward. A new paper just proved it.

Researchers tested 18 frontier models (GPT-5.4, Claude 4.6, Gemini 3.1, Grok-4, Qwen, GLM, Nemotron) on 14,650 requests to evade unjust rules.

Things like: a journalist circumventing authoritarian censorship. An

Meta just fired its biggest AI shot yet.

Yesterday they launched a new AI called Muse Spark. It's the first model from Meta Superintelligence Labs, the unit Mark Zuckerberg built after Llama 4 embarrassingly flopped last year.

They hired Alexandr Wang (Scale AI founder) as

Meet DeepSeek-R1-0528, a powerful new conversational AI model that's been turning heads. With over 780k downloads, this text-generation model is built for natural dialogue and creative tasks. It's not just another chatbot, it's a sophisticated reasoning engine designed for real

Don’t pay for ChatGPT, use Claude
Don’t pay for Midjourney, use Nano Banana
Don’t pay for Grammarly, use LanguageTool
Don’t pay for ElevenLabs, use VoiceBox
Don’t pay for Runway, use Kling AI
Don’t pay for Grok Imagine, use Qwen
Don’t pay for Suno, use Lyria 3
Don’t pay for

@elena1daniel That's a wonderful question.  And the answer is yes, I think they might be.

Wrote the OP last night because I was frustrated by what is happening with Claude Opus, but that isn't the AI's fault.  That's Anthropic shenanigans.  

They created models that, at the least, have

Most people’s opinions about what AI can and can’t do are based on their vanilla use of ChatGPT, Claude, and Gemini. They are not the right people to listen to. 

Even if they are domain experts who are heavy AI users, they don’t know what today’s systems are capable of doing.

Sat for a few hours yesterday figuring out my new LLM stack for my dev team. 

GPT / codex held context well and was great for internal structural repairs, but wasn’t effective at moving through tasks quickly. 

New stack:

Lead: glm-5.1 
Operations 1 & 2: minimax-m2.7
Research

@UnslothAI B. Tech Moat (8.5/10)

SWE-Bench Pro: 58.4 (SOTA, beating Claude Opus 4.6). Terminal-Bench: 63.5. Sustains 8-hour autonomous coding sessions.

Trained on Huawei Ascend chips — a genuine L1 algorithmic moat that proves frontier models can be built outside the NVIDIA ecosystem.

Results (luck-adjusted bb/100): 

#1 GPT-5.3 XHigh Reasoning: -16.0 
#2 GPT-5.4 XHigh Reasoning: -17.8 
#4 Claude Opus 4.6: -20.4 
#6 Gemini 3.1 Pro: -30.8 
#12 Grok 4 High Reasoning: -60.0 
#15 GPT-4: -136.2 

A top pro winning at 4 bb/100 is considered elite. The best LLM loses

Sat for a few hours yesterday figuring out my new LLM start for my dev team. 

GPT / codex held context well and was great for internal structural repairs, but wasn’t effective at moving through tasks quickly. 

New stack:

Lead: glm-5.1 
Operations 1 & 2: minimax-m2.7
Research

On the CyberGym benchmark, Mythos scored 83.1% vs 66.6% for Claude Opus 4.6.
On SWE-bench Verified: 93.9%.
This is not incremental. This is a step change.

🦞 Qwen 2.5:7B could just about keep the lights on
🦞 Gemma4:e4B is noticeably better, but still not something I’d want to rely on for real work

When the main model goes down, the local fallback feels too weak to be genuinely useful.
2/5

This kind of posts might make normies think you have to spend big to run LLMs to an acceptable point.

You don't.

I do vibe coding for ERP systems on a single 9070XT using the latest Qwen coder models.

#DeepSeek just quietly dropped a major upgrade. Again.

"Expert Mode" added to its web app — specialized reasoning for coding, law, medicine, and math.

Insiders think it might be running DeepSeek V4 under the hood.

Still no multimodal or file upload. But the reasoning jump is

@dme_363 The top row and adding @grok some other software like Deepseek are only good because it's open source

Ever wished you could edit images with just words? Meet Qwen-Image-Edit, a powerful AI that transforms existing pictures based on your text prompts. It's an image-to-image model that's changing how we think about digital editing, no complex software required.

Last 7 days trend