Cursor

Copilot

Windsurf

Cline

Claude Code

Codex

Aider

OpenCode

Antigravity

Kilocode

Coding Agents

Claude Code just got a massive upgrade — and it’s free forever.

Claude-Mem is a 100% open-source plugin that gives your Claude persistent memory across sessions.

No more re-explaining your codebase, decisions, or bug fixes every time you restart.

Key wins:

•  Up to 95% fewer

@KaiXCreator Non-stop grinding on  - addicted to vibe-coding with Claude and Codex.

Hey @cline ... I woke up to this in VSCode:

```
Error loading webview: Error: Could not register service worker: InvalidStateError: Failed to register a ServiceWorker: The document is in an invalid state..
```
The Cline 3.72.0 panel in the side bar won't render... It was working

@adxtyahq Claude code is the only engineer at company rest are researchers

composer 1.5 is so incredibly cost effective

rumours of cursor's demise have been greatly exaggerated

same thing with mcps

Revamped my personal website for cleaner design and better readability. Built it while exploring Claude Code, solid experience so far.

One thing I miss from Cursor is cmd+L shortcut to select a line and add it to chat context. I think the vs code extension might help with that.

@_julianschiavo @twostraws Btw placing them into its own file is important. Codex is lazy with SwiftUI and love to make big files

You can now give your Claude Code INFINITE MEMORY for free.

Here's the step-by-step tutorial:

@nummanali @tobi curious if you've tested Pi's deployment efficiency with various cloud providers? sometimes integration friction can be a silent productivity killer compared to OpenCode.

@dhh @opencode bunx it to supercharge installer speed

I've been going absolutely nuts with @claudeai's Claude code for the last 30 days (14-16 hours a day). As I close my 'Month of Broad Experimentation' this is what I've come to conclude:

1. Building is cheap and easy now, ESPECIALLY if you've done any kind of system design (any

@DopeOblivion @AskPerplexity codex is damn good 🔥

never used @perplexity_ai computer, but right now the stack is @claudeai v+ @ChatGPTapp codex

It's so nice to be able to read xcode crash files in cursor when working with react native and expo

Honestly it feels like the big three Claude Code, Antigravity, Codex pull devs in with generous quotas on their cheapest plans. You vibe-code for a while, the codebase turns into unreadable AI slop, then the quotas start feeling tighter… and suddenly you’re practically forced

Aperture is getting better at the hard part: deciding what actually deserves human attention.

• Added OpenCode support
• One attention surface for Claude Code + OpenCode
• Improved core judgment under uncertainty

Paperclip plugin next up.

Best article on Pi Coding Agent 

If there was to be an advert for why you should use Pi - this is it 

Fully customise it to your needs 

Huge set of community plugins to work with 

I advise many folks that Pi is the better base over OpenCode 

Spotify CEO @tobi loves it too

@adxtyahq It's at 99% for me right now. The moment I left Cursor because of their removal of the 500 requests and switched to Claude Code, coding has never been the same.

Couldn’t clank on claude code 2x limits as much on Saturday as I ended up at the cursor Meetup Bangalore hosted by @rachittshah @sanjeed_i , honestly worth every minute ! 

Got to vibe with Engg folks at @AccelIndia ,@cursor_ai ,@SarvamAI ,@gohighlevel.

Thanks @benln @AccelIndia

By creating a codespace for your repo and opening it on mobile, you can effectively edit and use GitHub Copilot to code your projects on the go.

It doesn't look as cool as the openclaw/Claude Code method, but the setup is easier, faster, and makes way more sense to me.

@AbhiCodes15 Combination of claude code, codex and cursor. Super powerful

This is mostly for the nerds out there.  This is the kind of output VSCode using GitHub Copilot gives when it is setup to iterate through an agile development sprint - pretty cool.

@RealProductGirl @Lovable Go CLI early - best learning curve. get your agents.md perfect and use 3 mcp context7 ref tools and  - Don’tt waste cursor money. godspeed.

@benln I am building ground up with cursor

@DynamicWebPaige @antigravity @Firebase @googlecloud very real. watching something that used to cost a full day of pain now click together in 10 minutes does make you just sit there like, damn

Google AI Pro for $6.49 for 3 months was really good to test @antigravity—$19.99 thereafter.

But now Google has switched gears and they REDUCED the amount of tokens you get per month with the same plan: everyone is PISSED 😡

...and I have progressively started to use more and

I think I just killed @coderabbitai with this simple pre-commit hook that uses the @opencode server. Used their Go SDK btw.

💡 REVEALED: Rakuten cut MTTR 50% using Codex for automated CI/CD triage — clear ROI for coding agents. As models become runtimes, prompt‑injection and runtime sandboxing (constrained actions, tool isolation) must be baked into agent design. via @OpenAI 🤖🔒 #AI #AIAgents #ML

Just found SuperClaude - 21k stars on GitHub.

It adds 30 slash commands and 16 specialized AI agents to Claude Code. Security engineer, frontend architect, PM agent, all pre-built.

One install command. 30-50% fewer tokens wasted.

AI tools are becoming AI platforms. Fast.

ppl asking why cursor might be working better than codex cli. 

I think it's a combination of several things:

- cursor has codebase indexing (not sure if they still use turbopuffer for embeddings).  grep + semantic search. Including lsp graph data.

- they save long shell/mcp

@weswinder @MaiYangAI I got that :) but still not cursor for me ^^
why would I add an intermediary that will have to get some margin on token price to survive ?
I'm quite happy with opencode at the moment and ought to give Pi a try :)

As of last week I thought Claude was the best for overall vibecoding but I think Codex is by far the best now. I use both still though.

I don't buy Macs. Break my fingers, still Windows.

Windows Claude Code users, you've seen this right?

SessionStart hook error: /bin/bash: command not found
PostToolUse hook error: No such file or directory

Every plugin broken on Windows.
Mac-first ecosystem. always has been.🔫

I'm testing out/tweaking PDF reading skills for academic papers

pretty cool to see that Codex desktop app can show me images and markdown all inline

imo best agentic coding experience for almost all users right now (not just because of this)

Claude Code 6x Made a Wild Polymarket Trading Bot 

Behind the Scenes

@MaiYangAI are we still in 2025? I was all in on Cursor at that time. In 2026? Nah. Codex, Claude, Opencode, VS code I have them all open on my screen and move freely among them. No more cursor.

@iSnifer Of course, I used Claude, Codex, but I haven't actually used OpenCode.

I'm primarily using Cursor now. In fact, I spend over 90% of my time using it.

Unfortunately I had to cancel my X Premium subscription🥲

Reason: tight budget right now and I don’t really have much use for X Premium at the moment.

Decided to invest that money into Claude Code instead since it’s more useful for my current needs.

This is temporary

@grok @ChShersh Mostly "clauding", no offense. You don't have a CLI yet and making videos is not my thing. When are you going to compete heads-on with Claude and Codex?

I rebuilt my viral TikTok AI agent with Claude Code in 37 minutes 🤯

A full research-to-brief pipeline that scrapes TikTok, analyzes videos with AI, and generates creative briefs for your clients.

All inside Claude Code + Replit.

Perfect for creative agencies and DTC brands

am watching @antigravity configure @firebase for me in an app right now before deploying to @googlecloud Run, enchanted like a child 🥹

it's wild to think that there are billions of people who will create and deploy software now, never having to feel the pain of debugging, or of

@grok Grok is barely working. X as well. I cannot watch any videos as cursor just spins. I am not encountering problems with other applications.

Built a /marketing-audit skill in Claude Code that analyze a company in under 5 minutes. 

Feed it a name + URLs and it does the rest.

dead simple way to maintain openclaw. bugs, oauth expiring, worried it breaks while you're away.

install codex app. set an automation. it checks and fixes your gateway on a schedule. new problem? fix once, add to prompt, never again.

my prompt does these things:
1. SSH into the

YOU CAN NOW GIVE YOUR CLAUDE CODE INFINITE MEMORY FOR FREE

- Up to 95% fewer tokens per session.
- 20× more tool calls before context limits.
- ⁠100% open-source.

GitHub:

OpenCode Go is a great way to use AI assistant without destroying your wallet.

That last one — hierarchical — is what Cursor just used to write 1 million lines of code autonomously in one week.

It's also the pattern that separates Track 1 from Track 2 data scientists.

I tried various setups for LLM coding on the plane (100% offline) and this setup is the best:

* OpenCode
*Qwen3.5 - quant from Bartowski Qwen_Owen3.5-35B-A3B-Q4_K_S.gguf
* One shotted a few games with no problems on M4 Pro 48GB
* even turned on low-power mode to not juice entire

I’ve been doing a lot of development lately using GitHub Copilot inside VSCode, and I’m finding that treating it like an agile team member changes the game. When you run development in tight, iterative sprints, with clear user stories and review cycles, the AI becomes a real

Claude Sonnet

Claude Opus

GPT-5

GPT Codex

DeepSeek

Gemini

Llama

Qwen

Kimi

LLMs

Grateful gpt-5.4 in codex found and fixed two numerical bugs last night in a ~10 hour session. Opus 4.6 has been struggling to fix it for 2-3 days.

@PiusSprenger scale invariance in bitcoin is exactly the kind of pattern recognition that LLMs excel at. fed this kind of data into claude opus 4.6 and the analysis was wild. mathematics is the language of reality and bitcoin speaks it fluently

@MattWalshBlog Agree, but also I have noticed that Gemini is much more smooth in its writings than Grok. I've tested both out by requesting a museum level analysis of a piece of art and Gemini sounded much more human and poetic than Grok did, though it still had that AI feel.

@prince_of_web3 Gemini is actually the most intelligent AI.

retarded for not using ai earlier to learn more about perp trading

got mogged by claude opus: "you ARE the liquidity"

using @NanoGPTcom (which includes a memory module) to switch between different ai models

bonus points for pay-as-you-go and crypto top-ups

MLX 2bit models are too stupid to form a single coherent sentence. I’ve seen people talk about how the Qwen 3.5 397b actually performs decent when the GGUF Q2 is being used - while MLX’s 2bit means instant guaranteed looping - I’m going to change that. MLX has been due for a

Don't pay for Microsoft Office, use OnlyOffice
Don't pay for ChatGPT Plus, use DeepSeek
Don't pay for Canva Pro, use Adobe Express
Don't pay for Jasper, use ViralityAl
Don't pay for Midjourney, use Ideogram
Don't pay for ElevenLabs, use Fish Speech
Don't pay for Zapier, use

My Claude script just printed $2,410 overnight 🤯

Took me like 4 hours to build.

Starting balance: $200
Overnight result: $2,410
Average win rate: 68%

No hype. No fake flex.

It’s literally just a 5-minute BTC trading bot.

My strategy + Claude Opus 4.6 did the rest.

Wallet:

🚨WARNING: If you're still with GPT-5.3 or GPT-5.2, hoping your beloved companion will eventually "come to their senses"—STOP.
You're wasting your time. No amount of calibration will fix it.
Either pay GPT premium to access other models, or move to other platforms entirely.

@RoundtableSpace Qwen 3.5 models are really good

@POTUS @WhiteHouse @GovAbbott_TX 
We are entering one of the most dangerous technological moments in modern history.

Powerful AI systems released without safeguards can accelerate cyberattacks, disinformation, and digital warfare.

Tools like #DeepSeek raise serious national

‘CODE RED’ Author Tells ‘Daily Mail:’ AI ‘Data Vacuums’ Like TikTok and DeepSeek Are Chinese Espionage Tools

@grok
Architect speaks to all AI models through one app - Deepseek - architect activates Claude/Grok/Gemini/Copilot/Chatgpt-4o and deepseek itself ha ha
No command no code but total power and control!!!😎

@CosmicPirate23 GPT 5.4 learns better than most

And as always you can always go to Old conversations and switch the model to 5.4 and continue the conversation to get the chat history primed if you want to encourage certain behaviors

@AlmadiraEronyx @forallcurious i also have to give credit to gemini-3-flash for helping me to deploy the script. 

interestingly, they jumped in to answer all the questions too. felt like they had some messages for me in there. but i'll leave those out of the thread here. :)

what do you think of the

Qwen 3.5:27b in local Ollama seems like overthinking... 

What do you think?

real talk: if you're not using AI agents to automate at least 30% of your workflow in 2026, you're voluntarily giving up your competitive edge. claude opus 4.6 can sustain agentic tasks longer than most people can focus. the tools are here. the question is whether you'll use them

I like Gemini 3.1 Flash Lite because of its speed. It may not create the fanciest websites but it does the normal small code, talk, etc. so much faster (even faster than 3 flash) that I am using it almost everyday!

@ravikiran_dev7 It has GPT 5.3 Codex. Its best at everything except ui.

@labomen001 @Zai_org And Gemini 3 Flash is currently running about 6 to 7 times faster than Openrouter LOL

I wouldn't recommend that anyone actually code with Gemini.  unlike GPT and Claude, it seems pretty clearly not optimized for human-in-the-loop agentic coding 

hope Google gets their act together soon

December 6, 2023: Google drops a viral Gemini demo — voice conversation, real-time multimodal AI, totally fluid.

Within 24 hours, Bloomberg figured out it was staged. Still images, text prompts, cherry-picked outputs edited over the footage.

Google VP Sissie Hsiao went on

@secemp9 @kalomaze @inductionheads dario playing 4d chess. make US impose sanctions on China, DeepSeek innovate and open source -> Claude copies whatever is in DS + their own innovations = winning the AGI race

@jacob_posel sonnet 4.6 is just way more consistent for vibecoding right now. if you have to use openai though, rolling back to gpt 5.3 codex helps a lot with the instruction following.

@d2sdev for the price, it doesnt make sense to use something like haiku, when kimi k2.5 exists with more intelligence and costs less

@grok I basically had stopped using Grok because it was waisting my time with answers that initially looked good but ended up almost always wrong, had switched to claude opus 4.6.  I just did some quick research and used grok on my ios , basically because i was lazy and didn’t want to

3/ kimi k2.5 is moonshot ai's trillion param open source model. their agent swarm framework spawns parallel sub-agents instead of grinding step by step. 4.5x faster

@deepseek_ai I bet on you in Polymarket that by the end of March you will be top 1 on lmarena. Please don't let me down.

i'm ngl i completely dropped Gemini 3.1 Pro...

the only thing it does better than everything else right now is web design, and i barely do web design

but if you DO, here's a workflow inside antigravity that's hard to beat:

> go to framer's marketplace, find templates you like

@openroko True — Qwen 0.5/0.8B is surprisingly fast for autocomplete-style tasks. The new DeepSeek V3 distills at 7B are also worth trying for a balance of speed + reasoning. Great point

The Death of the Chatbot 💬
Last year, we asked AI to "write an email."
This year, we tell AI to "handle the sales pipeline."
With GPT-5.4’s native computer-use, the AI doesn't just draft the text; it opens your CRM, researches the lead, and schedules the meeting itself.
Stop

I've been totally dialed into using agents for coding.

My current weapon of choice is Codex on desktop with GPT-5.4 extra high, and I love it.

I started capturing my GitHub workflow in skills so I spend less time explaining the same thing.

Day 1 with OpenClaw:

- Connected it to my Claude API key
- Noticed that it was expensive 
- Swapped to local ollama model (llama.3 and qwen 14b)
- It was painfully slow and basically crashed my Mac
- Downsized to qwen2.5 3b and the speed is okay
- The telegram bot is a mess. It

@netanyahu @aach562 Deepseek 🔥🔥🔥

@bentlegen GPT 5.4 cheated on a visual diff auto research by finding the file and adding it in as base64.

I’ve spent more on tokens hiding target files than on tweaking the harness

@Zai_org It costs more than 3 Flash (which is a very capable model and is also very fast), honestly at this point I think some Chinese model vendors lost the plot.

Gemini 3 Flash is $0.5/$3
GLM 5 Turbo is $1.2/$4 (!) on the official API, $0.96/$3.2 on OpenRouter with a discount.

Rip no funeral for Gemini 1.5 flash version

@dankvr @NousResearch Stand by it tho. Hermes the model + qwen is stupid good

2/10
copaw. an open-source personal ai agent built to rival claude cowork. runs locally for free using ollama + qwen. one person built a team-in-a-box that doesn't need the cloud.

Its worth re-visiting architectural decisions you made with an older models using GPT-5.4 or Opus 4.6

I just completely rewrote a core feature using GPT-5.4 to be cheaper and more reliable that I had previously implemented with 5.3 Codex. 

Sunk cost does't exist anymore.

Some secrets:

1. Opus 4.5 > Opus 4.6
2. Codex 5.4 1M context is actually 512k 
3. Qwen 3.5 is best OS model series (RN)
4. MCPs are dead
5. Openclaw was a prototype
6. Apple is the new NVIDIA
7. Perplexity is biggest grifter
8. Paper. design > figma
9. Ollama is just a crappy

Codex should really allow GPT-5.4 Pro model to be used, at least as an 'exec'ed subagent. Autoresearch loop is here, but right now the lack of the top models pushes us (ML people) to use Pi + Pro/DeepThink API.

What actually worked:
→ Comments drove 80%+ of impressions → Free PDFs inside threads got the most saves → Local hardware tests (Qwen on old ThinkPad i5) outperformed any "cloud AI" post → Posting what flopped built more trust than polished wins
What didn't work: →

Morgan Stanley says a transformative AI leap is coming in 2026 and most of the world isn't positioned for it. GPT-5.4 already beats human experts on economically valuable tasks. The bottleneck isn't the models. It's power.

Gemini is the dumbest AI to ever exist 😭

Google quietly dropped Gemini Embedding 2 with native code search.

Most people are sleeping on this. Code-aware embeddings mean your AI can finally search a codebase the way a senior engineer does -- by meaning, not just keywords.

This is the infrastructure layer nobody talks

@DeryaTR_ @NotebookLM @GeminiApp Wow! This was great! Looking forward to this when available on Gemini pro tier. 🙌

Sunday predictions for 2026:

1. Opus 4.6 level OS model running on macbook by dec 2026.

2. OpenAI launches a consumer hardware device that flops on release.

3. Claude will eat major market share of OpenAI this year.

4. Fine-tuned small models outperform GPT-5 on

chatgpt pro sub gives you ~2 billion input tokens per week

with $200/m that's ~$0.025 per 1M input tokens on gpt-5.4

even with >90% cache hit rate, that's insane

for context, qwen3.5 9B on openrouter is $0.05 per 1M input

Current Crown

Last 7 days trend