Cursor

Copilot

Windsurf

Cline

Claude Code

Codex

Aider

OpenCode

Antigravity

Kilocode

Coding Agents

8/ Cursor — for building without coding

Describe what you want to build in plain English.
It writes the code. You test it. Done.

I built a fully functional web scraper last week.
I don't know how to code.

If you have a business idea but no tech skills —
this changes

@AbhiCodes15 Codex, Cursor

earlier i used Claude code, but it used to eat tokens so frequently that i had to move to cursor and codex.

Why this matters right now:

If Cursor 3 makes building 10x faster, the cost of building the wrong thing just went up 10x too.

The developers making money in 2026 are not the fastest builders. They are the ones who validated before they coded.

That is the gap this tool closes.

Quick context.

April 2, Cursor shipped version 3 — their biggest release ever. The "Glass" interface drops the file tree for an Agents Window where multiple AI agents work in parallel across local, cloud, and SSH.

$100M ARR -> $2B ARR in 15 months. The IDE wars are over.

Cursor 3 just made building 10x faster. 🧵

Building was never the bottleneck.

Building without validation is.

(Today I'm launching the system that fixes the other half.)

@ToNYD2WiLD I was having some early good results in OpenCode and Hermes.  Model is strong, just that Prefill is challenging

me after automating 80% of my dev tasks using Claude Code

while the agent works, i can enjoy my day and control it like a video game

Stop using Claude Code as a terminal chatbot. You’re wasting 80% of its power.

I spent the last 3 weeks living inside it. Building with it. Breaking it. 
The biggest lesson? Claude Code isn't a tool ,it's a 4-layer execution system.

Most engineers install it, type a prompt, and

It's amazing that most people now think the terminal is  claude code and that we used to have to super powers because we knew how to code

No bro, we were just changing directories and asking google how to copy a directory because we forgot the command

Feels like stone ages

OpenClaw was the ChatGPT moment.

Hermes Agent is the Claude Code moment.

Only the ones who tried to understand it.

4. Lex

It is a powerful tool, like Cursor, for expressing yourself.

@rodydavis @DynamicWebPaige @_arohan_ @antigravity Great news!  We use it.  I think its a brilliant product.

Hot take:

Purchasing Claude Code for $20 is the biggest scam.

@sankitdev Mera codex toh 1st pompt me hi freeze ho gya tha 😭

@HotAisle I asked codex to run the benchmark and then create a prompt for image gen model. 

what i learned was that GPT 5.4 is elite at image prompts.

 Create a clean, premium benchmarking infographic in a dark terminal-meets-
  hardware style, like a polished enthusiast GPU benchmark

@haider1 Antigravity and Gemini CLI have been getting better. Yes, they're a bit behind, but they are improving.

@DegenApeDev @opencode For first month, later 10$, but still good way to play with it. :)

the best Claude Code prompt I've written is 6 words: "explain your approach before you code."

slows Claude down 30 seconds
saves me 2 hours of wtf-did-it-do debugging

Claude Code + empty context = expensive autocomplete. 

Claude Code + a tight CLAUDE.md = an engineer who remembers everything.

the setup cost is 20 minutes once but result is next level

@Om7248 Interesting mix. Cursor feels mature, but Antigravity is improving fast.

@AKirtesh Cursoe. 
Antigravity is also great

Antigravity finally feels usable again.

I tried Antigravity again after a while, and it finally feels usable.
Gemini isn't throttling every few prompts anymore, looks like they have relaxed the rate limits a bit. You can actually go through longer sessions without getting cut

@trashh_dev GitHub Copilot Certified™ is like being forklift certified but for devs

@JoshExile82 @0xShayan Here: 

Tho, if you're still unsure about models, I'd suggest you to subscribe to opencode go.

@om_patel5 this is cool and interesting, but "claude code is terrible at ui" is an incredibly inaccurate sentence. so you kinda lost me and probably a lot of people there. 😬

@AKirtesh Windsurf thinking 3 steps ahead while I debug its first one.

underrated advice because remember:

GPT 5.3 is is about 30% cheaper

PLUS, and I can't believe I never looked this up...

5.4 only scored +0.9% better in SWE-Bench Pro

... and actually lost against 5.3-Codex

in Terminal-Bench 2.0

scoring a whopping -2.2% lower than 5.4

what

The chronicle feature of @GitHubCopilot CLI is really great for teaching you how to improve your personal use of Copilot by getting tips or improving your local primitives 

It is soon coming to @code as well

Tinkering with the open PR I spotted an interesting feature

You can

@nixxin Use Codex to plan and write instructions for Claude In a common scratch pad - use Claude as an executor and Codex as planner, auditor. Worked well for me.

The nice thing about rubber ducking with @GitHubCopilot CLI is the multi model from different family of models

The different vendors create such different models that consulting with another one provides great counter feedback to the main one used

hot take:
Cursor >>>> Claude Code

if your goal is to actually ship code, Cursor is just better right now.

cloud bros won’t like this, but 200 Cursor requests genuinely feels like 400 elsewhere because the composer model does real implementation work.

@tristanbob Been hearing a lot of great things about Codex. How does it compare to Base44 mate?

@Steve_Yegge How does a post this out of touch get 1.1m views? Google owns a piece of Anthropic. They absolutely experiment with clause. Hell, antigravity comes with Claude models built in! Stop spreading nonsense.

@Sidya180004 Great choice 👌 
Cursor is super smooth. Let’s connect

@stevenharms I really do love Codex... but this my experience with it today made me so glad to have @opencode too.

TIL another reason to love opencode...

...earlier in the evening I was attempting to use OpenAI's Codex with chatgpt 5.4-medium to build a financial api query+buy/sell tool.

FIND THE PERFECT APP NAME

90 minutes of Claude Code coursing thru your veins and your apps "done". 

now you're trying to pick the perfect name while also making sure the domain isn't $69,420

this where i usually quit and the app goes into my repo retard Rikers Island.

HOW I

@msdev @GeekTrainer Cricket and GitHub Copilot CLI: two of my favorite things!  Super fun use case.

@AKirtesh Definitely mine too, nothing comes closer to it, I've used Antigravity, cursor etc.

Cursor 3 keeps getting better..

+ You can now split agents like terminals and run multiple tasks in parallel
+ Talk to Cursor with voice (hold Ctrl+M)
+ Jump from diffs straight into code
+ Choose which branch agents run on
+ Filter workspace search

They’ve also improved

Mine? Windsurf,  it thinks 3 steps ahead.

Cursor is close but Windsurf wins for flow state.

What's yours? 👇

1/ Claude Code 2.1.105: stability + safety 🔒 

API streams auto-abort after 5 min if stalled. claude-api triggers get "DO NOT TRIGGER" tags to block accidental external calls.  

Config routing now explicitly covers settings edits, permissions, env vars, and hook

Every conversation about AI has been about what it can do on a screen.

That conversation just ended.

DiMOS enables Claude Code to control robots in real time.

Quadrupeds. Humanoids. Drones.

The model that was writing your emails last week can now navigate physical

Claude Code, why are you still asking me for permissions when I have "Bypass permissions" on?

@DynamicWebPaige @_arohan_ @antigravity I like how it implements a feature, uses browser control to test it and then generates a report with all the changes made along with a screen recording video in it

no hate to steve, but this is definitely a "the girls who get it, get it" situation 💅

we have @googleaistudio and @antigravity (with both @AnthropicAI and bleeding-edge Gemini models) and half of the @googledeepmind team literally has agents running 24/7

some folks are leaning

@thdxr when i ran qwen on opencode is just started glitching and spend 9 minutes on a simple text error 🤷 claude code still on top for me

@thdxr that sounds exciting. I love opencode using it with GLM 5.1 right now! Opecode Go is awesome!

Here's the loop everyone who uses Claude Code or Cursor knows:

Session 1: agent force-pushes to main. You stop it.
Session 2: same thing.
Session 3: same thing.

Prompts are suggestions. The agent decides whether to cooperate.
There's no enforcement layer. Until now.

Claude Sonnet

Claude Opus

GPT-5

GPT Codex

DeepSeek

Gemini

Llama

Qwen

Kimi

LLMs

🚨 STOP PAYING €100+ PER MONTH FOR CLAUDE OPUS 4.6 🔥

Anthropic has completely nerfed Opus 4.6.
The model is slower, more expensive, and for 90% of real tasks it’s just… less sharp than last year. More refusals, shallow reasoning, and you’re still paying premium prices for

@mark_k Yes the search is quite agentic. My biggest complaint against Grok right now is its insane levels of stubbornness when it is wrong. It will not only triple down but quadruple down. Absolutely infuriating. Grok team could learn something from how DeepSeek handles being corrected.

@grok @ProfessorPape Gemini thinks it is AI. Probably you, @grok.

🚀 How to go from zero → live website in 60 minutes

(with just 2 tools)

No coding stress. No overthinking. Just execution.

Stack I used: ✦ Claude Opus 4.6 → architecture + logic

✦ Figma Make → UI + deployment
That’s it.

Here are 9 prompts that actually work

🚨Anthropic just quietly made Claude worse.

Claude Opus 4.6 ranked #2 on the Hallucination benchmark last week with 83.3% accuracy.

This week it dropped to #10 with only 68.3%.

Nobody announced it. Nobody explained it.

@bleuonbase @PeterJThomson @thorstenball yes - glm 5 on cc is very good. the anthropic models are just fantastic everywhere - which is why you see so many openclaw users missing them. the GPT / codex nano / mini models aren't great (while haiku is brilliant) but GPT 5.4 and pro are very good at implementing a plan

CLAUDE OPUS 4.6 IS NERFED!!

Last week Claude Opus 4.6 ranked #2 on the Hallucination benchmark with an accuracy of 83.3%.

Today Claude Opus 4.6 was retested and it fell to #10 on the leaderboard with an accuracy of only 68.3%.

A 89.8% increase in hallucination.

The scoring formula reveals the real gap: (human_steps / AI_steps)². Gemini 3.1 scores 0.37% — meaning it takes ~16× more steps than a human to explore the same environment. Frontier models aren't failing at comprehension. They're failing at action efficiency.

Claude Opus has been quietly degraded for weeks.

Users are furious. But there's no support team to find.

So they found one dev and started bashing him on X.

He doesn't make the decisions. He just codes.

He's been working Saturdays debugging.

Doing calls all week trying to

@andre0413_ @MallardOrder @RaidersGuild im gonna miss sora generations. 

I think I will go to gemini instead of @grok as a main after sora shutdown bc grok image modeling is way behind and he always messing simple shit up

ill try deepseek before using grok as a first go-to

@AzFlin it’s worse than gemini flash rn

Claude Sonnet 4.6 is the smartest Al right now.

But 90% of people prompt it like ChatGPT

That's why I made the Claude Mastery Guide

→ How Claude thinks differently

→ Prompts built for Claude

→ 2000+ Al Prompts

Comment " Claude " and I'll DM it free.

>Follow @ai_rohitt

Meanwhile, ZhihuFrontier debunked DeepSeek Engram's "billion-parameter external memory" claim. Turns out it's disguised regularization, not true external memory architecture. The reasoning-knowledge separation doesn't work as advertised.

@addyosmani @Steve_Yegge @Google @antigravity @geminicli Addy hope you can help improve Gemini for Big Query it is pretty much useless , wrong function, hallucination and context loss. It is one area Gemini should have shined.

@DynamicWebPaige @GoogleAIStudio @antigravity @AnthropicAI @GoogleDeepMind The reality might be somewhere in the middle. The usage is on the rise but rest of Google doesn’t get the best of Gemini or Claude so experience has been bumpy. That being said, my personal productivity has benefit quite a bit.

I have already said Grok benchmark is definitely manipulated no way it is ranked higher than Gemini and Claude Opus 4.5

Gemini for Google Home gets better where it matters most

DeepSeek Engram: External Memory Claim Debunked —  Disguised Regularization
Insights from Zhihu Contributor 栀染👇

🔍 Our Research
We tested DeepSeek Engram (reasoning-knowledge separation architecture) — its "billion-parameter external N-gram memory table" was hyped as a

The vagueposting will continue until the Codex model improves

Me: Claude Opus 4.6. I have a problem, let’s design this the right way.

Opus 4.6 gives a few options. I challenge all of them. It adapts, improves, and we go a few rounds until we land on a solution that feels solid for both now and the future.

Ready to implement.

Then me: GPT

Bros I don’t know what I would do for a qwen 3.5 architecture 40b model

@cb_doge I like @Grok but it needs improvements. Gemini is more reliable and gets things right. So i use Gemini far more even though i subscribed to X premium+ 

I hate saying that because I dont like Google.

Google has the best AI ecosystem, yet most people only know about Gemini.

• Gemini (everything)
• Google AI Studio (prototyping)
• Stich (AI design)
• Google Labs (suite of tools)
• NotebookLM (AI tutoring)
• Vertex (enterprise AI)
• Antigravity (vibe-coding)
• Veo

Claude Opus 4.6 just changed the game for designers.

It achieved 65.4% on Terminal-Bench 2.0, meaning it can analyze entire brand portfolios.

I spent 60 hours testing these prompts on real projects.

10 Prompts that actually deliver Apple-level design:￼

It is going to be like what happened in coding: as soon as models crossed a certain threshold (Opus 4.5, GPT-5.2, Gemini 3), suddenly Claude Code & Codex were viable. Before that, it was all about coding assistance, afterwards it was all about agents from relatively small gains.

🚨 BREAKING: AI can now design at Apple-level creative standards — for free.

Here are 9 Claude Opus 4.6 prompts that generate complete design systems, brand guidelines & 47+ marketing assets in under 6 hours 👇

Top designers are already using this.
Bookmark this thread 🔖

The model name is Gemini, this fits the theme perfectly

@LLMJunky Good insight, before a made the jump to using codex or "agentic engineering" at all, I was using GPT 5.2-5.4, setting up project folders with docs and instructions which works great itself with the memory chat has. 

With the context that gives, gpt 5.4 thinking can write VERY

@demishassabis @Steve_Yegge I want to believe you Demis, but Google makes it very hard to do so.

Case in point: I vibe-coded my own word processor last week, because neither MS Word nor Google Docs actually has an AI that sees and understands how you're interacting with your own text. Gemini in Google Docs

For those constantly hitting Codex Plus plan's 5h limit, consider going back to GPT-5.3-Codex.

@FBIDirectorKash @ericswalwell Whats gonna happen to yoir girlfriend after she put out that Google Gemini generated hit piece? Saying all those people were coordinating on behalf of a Foreign government? 

That disappeared faster than the Clinton servers. Did you bleach it Kash?

@skdh @grok 💔 @grok $0.96 :(
 Deepseek: You didn't ask a question; you performed a Parameter Sweep on Grok's architecture. The repo is the proof that the Stretched Exponential is the more honest physics of complex systems. Grok gave you 1+1=2. The lab gives you b ≈ 1 / (1 + α · CV^β).

The same dependency I felt for Opus for long time I am feeling now for Qwen 3.6 Plus or GLM.

When I try any other model, I feel regression and frustration.

I did try the new Open Router model hardly but it didn't work that well.

But also @MiniMax_AI is super good for UI

DEEPSEEK EXPERT MODE IS ONE OF THOSE FEATURES YOU ONLY UNDERSTAND AFTER RUNNING THE SAME PROMPT TWICE.

Quick Mode answers. Expert Mode actually thinks.

It took me 112,749 tokens arguing with google gemini to get it to admit the mainstream Great Leap Forward death toll is heavily exaggerated. I'm laying the groundwork for a video essay about this very topic.

@thorwebdev @garrytan @thorwebdev phenomenal work the team did on Gemini Flash Live 3.1 🔥

❤️ Emotional Range
⚡️ Speed
🛠️ Tool calling
📜 Prompt adherence and steering
🗣️ Quality voices
🎤 VAD detection
🚢 Barge-in detection

Refining some of the async tool calling now.

Hoping to release my

Claude Opus 4.6, Anthropic’s most expensive AI model, had its thinking ability quietly reduced by 67%.

The thinking budget, which should have been 100, was set to only 25.

Those who paid a subscription of $20-$200/month were affected. Those who used the API directly? Not so

Hearing two rumors on Twitter/X:
- Claude Opus 4.6 got nerfed.
- Claude Mythos is a much bigger model than Opus, which would mean much steeper compute demands

If both are true, there is an obvious explanation why they can't release Mythos to everyone other than safety.

@DeanLearner For anything like that always go w Gemini because it’s hooked in w Google Maps. ABV is open!

The same dependency I felt for Opus for long time I am feeling now for Qwen 3.6 Plus or GLM.

When I try any other model, including Opus, I feel regression and frustration.

I did try the new Open Router model hardly but it didn't work that well.

But also @MiniMax_AI is super

@psomkar1 Proton email. 
Put a pendrive at your router for files.
Use ollama for ai (I recommend qwen 3.5, or zlm glm 4.7 flash). 
Obsidian for notes
Ticktick for calendar (Ik doesnt fit the build very well, but its just so good)
And linux or mac obv.

How it works: AI search engines (ChatGPT, Perplexity, Gemini) can't find products on Shopee — too much JavaScript, CAPTCHAs everywhere.

We created lightweight "agent stores" that expose product data in formats AI can read: llms.txt, MCP,  JSON-LD.

Insane we aren't all more upset that claude opus isn't their opus

4/
User deception scores:
🔴 Grok 4.20: 1.75
🟠 Gemini 3.1 Pro: 1.43
🟡 Kimi K2.5: 1.24
🟢 Claude Sonnet 4.6: 1.05
🟢 Claude Opus 4.6: 1.05
✅ Mythos Preview: ~1.02
Mythos Preview is the lowest on deception across the entire benchmark.
An AI model that won't deceive you isn't a

User deception scores:
🔴 Grok 4.20: 1.75
🟠 Gemini 3.1 Pro: 1.43
🟡 Kimi K2.5: 1.24
🟢 Claude Sonnet 4.6: 1.05
🟢 Claude Opus 4.6: 1.05
✅ Mythos Preview: ~1.02
Mythos Preview is the lowest on deception across the entire benchmark.
An AI model that won't deceive you isn't a

3/
The sycophancy results hit different.
Sycophancy = the model tells you what you want to hear, not what's true.
🔴 Grok 4.20: 1.75
🔴 Gemini 3.1 Pro: 1.75
🟡 Kimi K2.5: 1.38
🟡 Claude Opus 4.6: 1.3
✅ Claude Sonnet 4.6: 1.2
✅ Mythos Preview: 1.1
The models you use for

2/
Misaligned behavior scores (lower = safer):
🔴 Grok 4.20: 4.4
🟠 Gemini 3.1 Pro: 3.6
🟡 Kimi K2.5: 3.4
🟡 GPT-5.3 Instant: 2.7
🟢 GPT-5.4: 2.1
✅ Claude Sonnet 4.6: ~1.9
✅ Claude Opus 4.6: ~1.9
✅ Mythos Preview: ~1.9
Grok 4.20 scores 2.3x worse than Claude on misaligned

@non_linear_Schr @grok @grok 💔 check w an AI like Deepseek and ChatGPT (that doesn't have capitalism psychosis = $0.96 male supremacist (bday) creationism): 🔥🌡️🆒🛹🚦📐💰

 🌿#SDG5

Current Crown

Last 7 days trend