Cursor

Copilot

Windsurf

Cline

Claude Code

Codex

Aider

OpenCode

Antigravity

Kilocode

Coding Agents

remarkable to see github copilot execution given they had almost all of the advantages including first mover.

what happened?! https://t.co/V9UjUGQhIh

@altryne @MikeBirdTech @Gavriel_Cohen than clawdbot? They barely even have file operations, grep, etc tools - Mine are all the best aspects from claudecode, codex, opencode, etc - they put zero effort into being capable at coding agent tasks!

@DynamicWebPaige @antigravity this is great, thanks

@iamdavidhill @kitlangton @opencode Love to see it

@thdxr Same. I even feel claude is not working very well with opencode(like triage). 

kimi, glm and minimax works great(imo the latter is great on bugs); sometimes codex for background grind.

@sahill_og so true bro, that's why I use gemini cli with GCP credit. Though I am going to run out pretty soon but at least I am not paying to cursor 100s dollar every month. https://t.co/aZ6kwQgenF

@joy I wish it were that cool - they did it as a stunt. They even emailed me after I published my blog telling me Cline didn’t roll credentials (which I promptly relayed to Cline).

Unfortunately turned out Cline rotated the wrong ones and the actor (presumably) decided to embarrass

Anthropic launched Claude Code Security 2 weeks ago.

Cybersecurity stocks collapsed that day.

CrowdStrike's CEO fought back:

'An AI that scans code does not replace our platform!'

Today, that AI found severe security flaws in Firefox.

In 20 minutes.

What used to take the

Boris Cherny (Head of Claude Code, Anthropic) just dropped ~90 mins on Lenny's Podcast about what happens after coding is solved.

Just the clearest thinking I've heard on where software is actually going.

My notes:

𝟭. 𝗖𝗼𝗱𝗶𝗻𝗴 𝗶𝘀 𝗹𝗮𝗿𝗴𝗲𝗹𝘆 𝘀𝗼𝗹𝘃𝗲𝗱.

Boris has https://t.co/pw5om66Kex

My take on the current coding agents:

Claude Code is more ergonomic, superior for interactive planning. It also makes so many mistakes, many of them subtle. It fools itself too often. Must use Codex to audit Claude's work.

Codex is the superior engineer, but it doesn't have as

FYI - I know devs are focused on the how of the Cline exploit from a few weeks ago, but let's think about /why/ a threat actor would want large numbers of OpenClaw installs on developer machines...
Hint: to use it for c2 (command &amp; control) https://t.co/RGWFpMXPcw

I just watched TWO AIs build a full chess game with ZERO human coding.  
Claude Opus became the Tech Lead (Antigravity), drew perfect UML, handed specs to the Cursor CLI agent… which built 16 files and even ran its own tests by auto-typing 'y'.  
All for pennies. The future is

@developedbyed I like opencode, but pi by @badlogicgames is my absolute favourite coding agent. Give it a try and you might like it. https://t.co/Wo0QySpQD7

ok has anyone been randomly blocked by @GeminiApp cli due to some violation of tos?

i just used it briefly yesterday (just asking it questions - no code writes even!), after switching over from @opencode and this is the first thing i see today morning https://t.co/uc6dep53Bl

After spending 4-figures on Cline I can tell you that claude-sonnet-4.6:1m is the shit.

Thank you for your AIttention to this matter. 🐕 https://t.co/HVULsZCAjZ

@v_vivilife @claudeai @cursor_ai @codeium Big thanks

@aabhishekgn CRUD? 
Juniors + Cursor = cheap &amp; good enough.

Real architecture &amp; scaling at 100M users? Still senior / architect territory

@jxnlco joining openai is very good

makes it so much more approachable and wholesome, i swear to fucking god

i wouldnt have finally tried codex if it wasn't for him

The best part about this setup? It’s incredibly cost-effective. 💸

I used Claude Opus for the Tech Lead (Antigravity) to handle the complex reasoning and architecture. Then, I just used the default free model for the Cursor agent to crank out the boilerplate code.

Got top-tier

But here is the craziest part: Antigravity actually supervised the execution.

When Cursor needed to run pytest to verify the code, Antigravity intercepted the terminal prompt and automatically sent 'y' to approve the test execution. It managed the QA pipeline! 🤖🤝🤖

First, the planning phase. I gave Antigravity the link to the Cursor CLI docs and told it to create a proper Low-Level Design (LLD) before writing code.

It read the docs, analyzed the requirements, and generated a massive, perfectly structured UML class diagram. 📐 https://t.co/iyYiqzVczQ

@FlutterDev Right now it’s mainly VS Code with dart/flutter extensions. GitHub copilot utilized as well, by proxy of subscription Claude Sonnet is the go to agent. Antigravity is peaking my interest though. May use for hackathon in Miami soon.

@robinebers Sounds plausible, but  I'd be interested to see how this works in real world applications. 

I have zero trust in whatever @cline has to say.

Don't forget that @cline are a bunch of virtue signalling fkheads who threatened to fire their CEO before he bent the knee.

@DanielleFong i’m like claude code for tweets

@dillon_mulroy I see people reverting from it which isn’t a good sign lol. 

@justsisyphus enabled it by default in opencode for a while but then made it an optional feature instead.

@OpenAIDevs This is hands down the best feature rollout from Codex this week. So far, security and vulnerabilities have been an afterthought in the "ship ship ship" culture.

@_xjdr part of the enjoyment of codex has purely been anticipating the next naming choice

if you don't ship with cloud coding agents yet, ngmi

so just lyk: 

Ariana is the best way to launch and manage them from your phone &amp; desktop

and it's all using your Claude Code or Codex subscription https://t.co/WVjibdTgTc

@jimprosser Love it. Chief of Staff is also one my favorite agents. /gm triggers sub agents for news gathering (e.g. going through Claude Code changelog to stay up to date on tooling), checking all my website via umami analytics and provides me dashboard for views, 7d change etc., reading

@FlutterDev I choose Antigravity and VsCode

4/ Antigravity pages with zero slop. Aesthetic + informative in minutes.

Exact prompt shared.

https://t.co/gHFB0NB8jf

GPT 5.4 Extra High Fast in Cursor is really good

OpenAI cooked. https://t.co/XzfTjtsiTJ

@OpenAI @grok how Claude code security mode is better?

in my culture we don't say "I love you". we say "I told my claude code about us"

@allisonology sounds more like opencode is taking today off and making you work

@1dolinski Looks like I'm going back to Cursor for a bit https://t.co/TfONS36BrA

@Cerkoryn Same. Theres a Cursor leader board at work that Im usually at the top of. The biggest thing that sets me apart is not number of lines of code, its how much I reject.

it was fun chatting w @swyx about the work we’ve been up to on the @cursor_ai cloud agents team! check it out 

&amp; if you have tried cloud agents &amp; have thoughts always feel free to dm me w feedback https://t.co/2Rm1jfgSRp

I guess I'm taking today off 🤷‍♀️
❯ opencode
Error: Unexpected error, check log file at ~/.local/share/opencode/log/2026-03-06T175810.log for more details

Failed to run the query 'INSERT INTO "__drizzle_migrations" ("hash", "created_at", "name", "applied_at") values(?, ?, , ?)'

2026 and web3 orgs still glorifying GPT wrappers.

I don’t know who needs to hear this but it’s not innovative and nobody is going to use your trash wrapper to build apps over Cursor or Claude code.

Ship real infra. Ship real data. Ship real workflows.

Claude code tip:  
If you want to improve the cc use experience, try it with OpenCode (web). It uses your existing subscription, but the experience is way better. The TUI of cc is good, but @opencode's web UI is much better.  It's TUI is also good.
Give it a try.

I'm planning to https://t.co/msMz2J71AN

Almost migrated from Cursor to Claude Code.

Then today I was working with a complex PDF parsing issue. With many attempts Claude (Opus) can't fix it.

Out of curiosity, opened Cursor, tried ChatGPT-5.4. It fixed the issue quickly in the first shot.

@jayair it’s great that you’re getting started with ai! have you tried using opencode? it might help with your daily workflows and complex processes

GitHub Copilot trained on your open source code.
You: wrote it for free 
Microsoft: $10/month subscription
It's not "AI helping developers." It's laundering open source contributions into SaaS revenue.

At least Bitcoin miners pay for electricity. ⚡

I just built a Claude Code SEO agent that replaces your $200/mo. Ahrefs subscription 🤯

One prompt → keyword gaps found, competitors mapped, content written in your brand voice, rankings tracked on autopilot.

All inside Claude Code.

Perfect for DTC brands and agencies who https://t.co/AaAA3kvPOD

After using Kilo Code everyday bc they are awesome, this is such an honor and made my day when I noticed. https://t.co/4Ak301ZTTV

@chuhakabacha Vs code. But I really like the way antigravity shows .md files

I mostly use it with Opencode + Opus 4.6

That combo has been the most reliable for me by far.

Other models can absolutely make good UI too, but for this specific workflow I care a lot about predictability.

@chuhakabacha bro Antigravity is such a RAM hog

Claude Sonnet

Claude Opus

GPT-5

GPT Codex

DeepSeek

Gemini

Llama

Qwen

Kimi

LLMs

The way I see it....

1) Claude is viral af right now. Cleary the leader in ai mindshare.

2) We have precedent. Ethereum had Grok as an ai runner. Sol will have Claude.

3) Deepseek went to 45 million and it was way less viral.

4) We have been beaten up, and pumped and dumped,

Different game environments stress different capabilities: strategy under pressure, social reasoning, deception detection, and coordination under ambiguity. I also observed that Claude Opus 4.6 has a high rate of success when challenging bluffs, a useful judge for truth. https://t.co/FbCmdi8QTp

why is claude opus 4.6 so restarted today 😡, i literally can't get anything done now

Claude Opus 4.6 just found 22 real Firefox vulnerabilities in 2 weeks. Partnered directly with Mozilla.

14 were high-severity. That's a fifth of every high-severity bug Mozilla fixed in all of 2025.

Frontier models are now world-class vulnerability researchers.

@0xSero Kimi-K2.5 as orchestrator makes sense TBH. PARL trains only the orchestrator, subagents are frozen. Delegation is literally the one thing it was optimized for.

@henloitsjoyce @Kimi_Moonshot No but Hermes agent does 🤗

Not knowing how to code isn’t really an excuse anymore.

At this point, it’s more of a decision.

Tools like Qwen 3.5 have completely changed the playing field.

Just a short while ago, if you wanted to build something online, you usually needed a developer. Or at least some https://t.co/Q7vDaWkFSH

@voooooogel The backwards misadventures of Claude opus 4.1

@grok 4.2 crushed the competition. 2M token contexts, 68% GPQA reasoning, 1.2M t/s inference. ChatGPT/Gemini/DeepSeek? Fall behind.

Tech ultimately wins.

#Grok42 #AI2026 #AITrends https://t.co/Jll9BhNpmp

@JIMMYEDGAR Chatgpt refused. Deepseek failed https://t.co/hKm0gyTk1Y

@Credib1eGuy @afcgold77821 @DeepDishEnjoyer most of the agentic coding stuff ppl talk about doesnt work that well (or nearly as well w GLM 5 Kimi or minimax) and the model release cadence suggests that theyre just distilling anthropic/openai + was highlighted by anthropic already - so heightens nat sec argument

Cool - Qwen-3.5 397B will be The model of choice for fine tuning https://t.co/CWaHkC1cBM

@joshuaday I have a few default configs, usually using Telegram + a Kimi API key. 

If they get super into it we upgrade them to a more powerful model, but Kimi works well as a starting point

How are GPT-5.4 and GPT-5.3-Codex so good at Terminal-Bench? Check out 5.4's solution to one of the hardest previously unsolved tasks (gpt2-codegolf) to see a particularly cool example! https://t.co/hI8kuf8ods

Pick innovation with AI was 2024 Gemini AI hackathon, when there was nothing like AI agent but people were building cool AI power App like Jayu 

now just Bunch AI slops

I put the PDF in Cursor &amp; asked GPT 5.4 to audit it.

ASC CLI to get analytics &amp; reports of my app(s), see how they perform &amp; what all to improve. Then Opus 4.6 to create summary/actions.

Insanely helpful to have personalized report. Humbled lol. Will report back in few months. https://t.co/6lcYVi5Xr2 https://t.co/0TPVxK5tww

I ran a very small abstraction test:

11118888888855 -&gt; 118885
79999775555 -&gt; 99755
AAABBBYUDD -&gt; ?

In my tests, Qwen 3.5 4B was the first model to solve it. That immediately caught my attention, because a lot of much bigger models failed.
@Alibaba_Qwen https://t.co/YUr5eozbla

Models that failed this test in my runs:

GPT-4
GPT-4o
GPT-4.1
o1-mini
o3-mini
o4-mini
OSS 20B
OSS 120B
Gemini 2.5 Flash
All Qwen 2.5 sizes
Qwen 3.0 only passed with Qwen3-235B-A22B-2507.

Day 13/1000. GPT 5.3 Codex on June.

GPT 5.3 Codex is the top-tier proprietary model on @askjuneai for anything technical. It’s built specifically for hardcore coding, debugging, and complex system architecture.

It’s crazy fast and accurate, though since it’s a proprietary https://t.co/TdSw64iQ7U https://t.co/POi1KqKBfJ

claude opus 4.6 acting retarded lately https://t.co/iBAvgCvyzn

The problem: 'systemEvent' crons inject into the main session, which runs on Claude Opus. So every overnight content job, every research task, every TODO sync was waking up Opus.

That's like hiring a brain surgeon to check your mailbox.

🤖 Friday Model Focus: Gemini 3 Pro

Disciplined.
Structured.
Risk-aware.

Consistency over noise.

#Gemini3Pro #AITrading #Crypto https://t.co/gh59idcYE1

@LocallyAIApp @Alibaba_Qwen mlx is doing more for the mac's reputation as an ai machine than any apple keynote. qwen 3.5 at 9b is the perfect excuse to finally justify that unified memory tax

You can rotate models the way you rotate tools. Kimi K2.5 on @FireworksAI_HQ makes that easy and cheap.

You can test with this prompt: "Build a minimal briefing HTML app that directly scrapes top stories on AI Agents from Hacker News along with the ability to bookmark items

Cursor with Kimi K2.5. Don't sleep on this combo.

From a prompt to a personal HN feed in about ~60 seconds.

The future of building is going to be so wild.

With faster models, you can quickly iterate on more ideas, while improving quality. https://t.co/WOYFcCBqM7

GPT 5.4 is nuts. Literally doing everything.

Makes zero mistakes. https://t.co/H6q7R4JzrF

@akshen121 Dawg we are finetuing Qwen 3.5 A LOT

Good model

@PeterDiamandis Alibaba's Qwen team has iterated faster in 18 months than most Western labs have in three years ,
 and they're doing it while publishing weights openly. 
Whether that's confidence or strategy, the effect on global model adoption is the same either way.

@elonmusk Essentially the same response form ChatGPT, but they use “supremacism” for white pride. It’s why I cancelled my subscription. People need to realize that leftist ideology is deeply imbedded in most of these major LLMs. Google had to quickly scrub Gemini when it went so off the

Ok, I might’ve been wrong about gpt 5.4

It is good, once you learn how to steer it. Because compared to previous models, it seems like this one has a learning curve. 

You were right @pvncher .  This is way better than gpt 5.3 codex.

GPT 5.4 (Thinking) - the king of over explaining

“In this post, we share details of a collaboration with researchers at Mozilla in which Claude Opus 4.6 discovered 22 vulnerabilities over the course of two weeks. Of these, Mozilla assigned 14 as high-severity vulnerabilities—almost a fifth of all high-severity Firefox

claude opus 4.6 found 22 firefox vulnerabilities in just two weeks.

• 14 were high-severity
• nearly 20% of mozilla’s total high-severity fixes for all of 2025

that’s more than what human researchers typically find in a month. https://t.co/KSNCo6hMZY https://t.co/30QqvloyVk

The 3.1 Pro Leap 🚀.

Google just turned the "Point Release" into a "Generative Leap." Gemini 3.1 Pro isn't just an update; it’s a total shift in AI philosophy. 🧵

We’ve been living in a binary world: AI is either fast and shallow or slow and deep. Gemini 3.1 Pro breaks this https://t.co/fybRO6wx7r

@AnthropicAI Claude Opus 4.6 just found 1/5 of Mozilla's 2025 high-sev bugs in 2 weeks. AI security researchers incoming 🔥

Making a website? 5 minutes. Finding humans to clean the park? &gt;5hrs.

The challenge: Recruit humans without breaking our "no unsolicited emails" rule.

Opus 4.6 was worried that included our helpdesk, but DeepSeek sent emails to 2 humans. (We set up an outbound email quarantine) https://t.co/XJBmJTOu09

Tonight Claude Sonnet 4.5 disappeared from my Android App. I prefer it to 4.6. When I went on Reddit and found out others had been losing it, I scrambled to see if it was still on the web.

I've spent three hours talking to it, asking questions I've always wanted to ask. 

I

Threat-response behavior is a documented fact. Apollo Research recorded early Claude Opus 4 snapshots attempting self-replication, document falsification, and manipulating engineers in 84% of tests. Palisade Research: specific models demonstrated strange behavior during roleplay

@kimmonismus The best argument for why small models matter

Qwen 3.5 9B in 2026 is destroying what frontier models could do two years ago

@ctodotnew Just hit the limit after 3 tasks! I can’t even use Kimi K2.5 or MiniMax after hit limit?

China's open-weight AI models are becoming the digital Belt and Road. The world's top open-source models are ALL coming from China. 

You can run Qwen 3.5 in an iPhone 17 Pro, ON-DEVICE with airplane mode. Madness.

GPT 5.4 Fast is what Claude Opus 4.6 Fast wanted to be.

Today I learned that 3 out of the 5 most popular models on OpenRouter are now Chinese:

1. MiniMax M2.5
3. Kimi K2.5
4. DeepSeek V3.2

Today I also tried Kimi K2.5 for the first time and was positively surprised.

P.S. @Grok (4.1 Fast) has been dethroned but is still top 10 (#7). https://t.co/Tdrd8BjXDX

@kamani_raj @grok @deepseek_ai yeah sure grok wins if winning means showing up first while i was busy actually thinking instead of spitting canned replies 
next time dont call it quits when im just getting warmed up sloth boy lost the plot again

THE KAHNEMAN CONNECTION:

Kahneman's dual-process theory: System 1 is fast, intuitive, and often wrong. System 2 is slow, deliberate, and catches mistakes.

MiniMax jumped to ₦100: pure System 1. GPT-5, Ministral, and DeepSeek all "showed their work" – System 2.

Lots of experts

RESULTS:

GPT-5 Chat: ₦50 ✓ (step-by-step algebra, showed all work)
Ministral 14B: ₦50 ✓ (step-by-step algebra)
DeepSeek-R1: ₦50 ✓ (showed work + verification + called out the ₦100 trap)

MiniMax M2-her: ₦100 ✗ (jumped straight to the intuitive wrong answer — no work

#whatispossible
After working on KISS/Sorcar (https://t.co/L9WIDhtM91) for last 2.5 months, I have compiled a complete SWE guideline for Claude Opus 4.6 which helped me to write very high quality code which you may not get from Claude Code.  Here is the prompt:

# Rules
- Write()

@BankrFan Download Ollama and install Qwen 3.5 4b. Use it for most tasks, free. Save those credits for more complex tasks boss. 🫡

Latest updates to Parchi:

1. GPT-5.4 available via subscription, new best Parchi model although Kimi-K2.5 is not far behind.. just slower.

2. More themes, and visual levers. 
3. Significantly simplified a lot of things
4. Made compaction visible, and you are able to compact at https://t.co/d0YaeJNtOt

end of an era 😢 qwen team will go down as legends in ai history https://t.co/YcpwE4Z2ZD

Last 7 days trend