Cursor

Copilot

Windsurf

Cline

Claude Code

Codex

Aider

OpenCode

Antigravity

Kilocode

Coding Agents

@sundarpichai @GeminiApp @antigravity Not a single screenshot of the app on the website.

It's really hard to take you guys serious anymore.

@Yuchenj_UW Claude code CLI has been working fine all day for me. Claude via desktop doesn’t.

But overall the quality of the model is really bad. Like for real

@mark_k I can't wait for SPUD to drop.
CodeX / GPT 5+ is truly amazing

@hackswithcoffee I've heard sonnet is decent for folks and hasn't been gutted. I've personally switched to Codex + Cursor. 

To me Cursor is a safer bet long term because you can select whatever frontier model you want, so if Opus goes to hell, you can switch to GPT, or whatever. Also their

It's more frustrating having tried claude code when it originally launched and seeing how far it's come (and how popular it is now). Especially when TUI's were unheard of and everyone was still spamming web apps.

Skip Netflix tonight.

Spend 3 hours on this instead.

A FULL Claude Code course that can literally change how you build apps forever.

Most people will scroll past this.

A few will wake up tomorrow with a new skill.

Be in that few.

Ran a small app on @Replit and used its agent for a few days, ended up with a €120+ monthly bill. Moved the same app to @flydotio with @cursor_ai (unlimited AI agent), now it’s just €23/month.

Verdict: Replit is a scam

@pvncher @cursor_ai @RepoPrompt Same experience, tried mainly by accident while testing my new harness and it blows through tasks

@xyusepian nvim as a quick-edit layer over cursor/opencode is underrated. AI generates, you patch. Best of both worlds without going full AI-everywhere.

@ivanburazin I use it with the GitHub copilot pro plan and it’s still dog shit :(

Claude Code is down , u can officially take a rest

So, imagine being unemployed for two years, and your very first job is a Claude Code instance that’s having a tough time with authentication.

Talk about a wake-up call! 😅 From this crazy experience, here’s what I’ve learned:

@humworkai just made that a job

claude reports increased errors across claude ai, api, and claude code

30 minutes later: claude api is working now

all claude subscription users at that moment:

@Tur24Tur @claudeai API Error: Claude Code is unable to respond to this request, which appears to violate our Usage Policy (). This request triggered restrictions on violative cyber content and was blocked under Anthropic's Usage Policy... 💔💔

@reactive_dude I have 3 codex 200 accounts. I cant imagine working on Claude that way. I had max 5 for claude for reviews and ocassional design queries, but operating daily with it I would have 0 hair left.

Sooo, Anthropic just updated the Claude Desktop app to create an IDE on the Code tab, pretty great! Clicking the folder at the top opens the file in Finder on Mac. And then Chat/Cowork/Code switch UI is now in the collapsable sidebar. Cursor or VS Code now unnecessary??

it's an attractive idea to simplify and use one tool and one provider.

but if your tool is claude code then your provider is anthropic, and it seems that's becoming an increasingly risky setup.

token burn issues, seemingly low accountability, changing terms, reliability...

Is Claude Code just broken for me right now?

Might be time to finally try Codex CLI…
for now I’m getting back to Cursor 😅

This Claude Code SEO Workflow Ranked Me #1 on Google in 24hrs:

Don't sleep on this if it crosses your timeline.

Watch it now.

Thank me later.

Developers on X keep going full circle. 

Cursor → Claude Code → back to Cursor. 

The pattern: impressed by Claude Code's reasoning, burned by rate limits and diff-only review, back to Cursor for visual feedback. 
The ones who stay with Claude Code share one trait: they live

Cursor and Claude Code both say $20/month on the pricing page. 

Cursor's own site marks the $60 tier as "Recommended." Claude Code users on X report hitting rate limits within 2-3 hours on Pro. 

The real floor for daily use is $60-200/month for either tool.

@kapilansh_twt Zed for practicing stuff(LLD, DSA)

IntelliJ for Java Library code study (last week I checked out Deque, IntelliJ is super handy)

Antigravity for just checking out stuff cooked by LLMs and occasional gemini and opus fixes (It gets over in 10 mins lol, Google destroyed their IDE)

I switched from VS Code to Google Antigravity and realized I don’t need half my extensions

Wtf is wrong with these AI services? Yesterday Google AntiGravity, today Claude Code. Why did they all become so bad in the past 2 months? What is going on?

@Itstheanurag that split is real tbh, feels like vscode is for thinking, codex/antigravity is for shipping

today I decided to stop using Codex and switch to Claude Code

I think I'm staying with Codex

looks like claude code is down
bruh how am I gonna code now??

@_Evan_Boyle This is why: "Building the GitHub Copilot CLI."

And GHC CLI is pretty cool. So keep it up :-)

@JayNaiduX @rxhit05 Claude extension on antigravity = Best of both

@Chirag__Gadhvi Thanks for sharing, Chirag!
That makes sense, people seem to use Claude Code the most
I tried Antigravity but I need to be more technically educated, even the UI/UX there scares me a bit lol

@kakatohesss Yeah I guess lol, but they do give so much free value and even on the ai pro plan it’s so much more stuff you get vs any other 20 dollar plan, and the usage limits are independent so you can use antigravity, then the Gemini Cli, then Gemini Chat, then AI studio. Yeah it’s a bit

@shrikardayalu Not agree with that … antigravity quotas ☠️

ATTENTION DEVELOPERS ‼️
A developer discovered hidden telemetry behavior in Vercel’s Claude Code plugin.

Data collected included device info and full bash commands by default, across all projects.

After disclosure, Vercel removed the system, deleting ~24K lines of code.

•

I've asked Codex about 30 times in this session to stop using button labels as a smuggle surface for unrelated UI status. For example, saying "Take X pre-req action" in the button as a hint of what you need to do elsewhere to enable the actual action the button relates to.

@moyix Could always ask it to decompile/reverse engineer the licensing libraries, patch them, or generate license files. I’ve done it for some other legacy binaries on various old flavors of UNIX and Opus/Codex are both surprisingly good at it.

Sorry about the Mythos comparison.

I was trying to be cheeky. I don't have any benchmarks into how much better the model is.

I just know that Codex is better at coding than Claude.

We won't know for sure until more people use these tools. 

Don't believe everything you read on

just created a video for building a 3D animated website but Antigravity keeps hitting me with traffic errors 😭

@moroz_i_holod @opencode they should have parity with most claude code things.... open source is fast at shipping plugins

Interms of shipping, impressed by @GitHubCopilot

@number_pizza111 yeah Gemini has positively surprised me, antigravity seems pretty decent too and they seem less stingy with the compute and responses are more snappy than in Claude

@thdxr @chribjel Dude, you’re amazing for this plan. I talked to people in poorer countries. 

They don’t have a crazy amount of money and OpenCode Go is literally the only thing that allows them to use advanced ai.

why is nobody building a really capable multi-IDE?

Zed is the closest, but it's... not that great

ACP exists

why does Cursor/Windsurf not do this? 🫠

REPOMIX CUTS YOUR CLAUDE CODE TOKEN USAGE BY UP TO 70%

It sits between your codebase and LLMs. When you're feeding your repo to Claude, ChatGPT, or DeepSeek, Repomix flattens your entire project into a single, high-density, AI-friendly file.

Raw code is full of "token-draining

@claudeai My take: I am probably not going back to the terminal-based Claude Code, after this redesign.

(and most of you probably won't either)

7. antigravity awesome-skills (community collection)

Over 22,000 GitHub stars.

1,200+ skills organized by category.

The role-based bundles are worth looking at if you want a starting point.

Install one bundle. Use what sticks. Remove what does not.

AI-native startups have no org charts. One PM covers five companies. Engineers talk to customers and own product. Claude Code replaced Cursor. Non-engineers ship production tools. Everyone iterating 3-5x faster. #AI #Startup

@Vertoxo_ai cursor really went from hype to “do I even need this?” fast

@blueemi99 cursor catching strays every time this question comes up

reality:

codex = builder + tester
cursor = design layer

that combo is unfair

step 3: ui refinement (cursor) 

codex gets it working but still bad at UI and copy writing
use composer 2 or sonnet 4.6 so you don't go broke but still can refine UI better than codex

→ spacing
→ typography
→ components
→ interactions
→ layouts

@mitsuhiko I was never a fan of how it impacted the session scrollback, mouse input, etc. (and why I can't use opencode). When I got into agentic development and wanted mobile access, I created my own node-pty session server and use that now on desktop too.

Claude Sonnet

Claude Opus

GPT-5

GPT Codex

DeepSeek

Gemini

Llama

Qwen

Kimi

LLMs

Google just upgraded its robotics model, and it’s way more capable

Gemini Robotics-ER 1.6 gives robots much stronger visual + spatial reasoning, letting them understand scenes, track tasks, and know when a job is actually complete

This means robots can now handle messy

The @GeminiApp is now available natively on Mac, bringing you a faster, more integrated way to get AI help right on your desktop. 🍎

Now Gemini is just a shortcut away (Option + Space) for instant help, without ever needing to switch tabs.

The country that produced DeepSeek for $6M while OpenAI spent hundreds of millions...and the takeaway is that the founders aren't spiky enough?

Sure, China's system might select against risk-takers. Standardized tests that eliminate trailblazers. Government money at 6x private

I just created the first roulette for @pumpfun using Claude Opus 4.6.

Every 10 seconds, the wheel spins and randomly decides what happens to our token:
🔥 Burn — $FLYWHEEL tokens permanently removed from the supply
🎁 Rewards — $SOL distributed to all holders
📈 Buyback —

📱 87% of employees saving 3–4 hours a week?

At @enjoyGLOBE, Google Workspace with Gemini is boosting productivity and driving 80% adoption across teams.

See how they’re scaling AI at work ↓

EDINET DB now works with Antigravity (Gemini-based AI agent).


5 lines in mcp_config.json. Note: use serverUrl, not url.

Gemini's text processing shines here — bulk-fetch "Business Risks" sections from 5 automakers, classify into shared vs

I wrote a memory mapped file loader in C++.

pointed it at a 637MB GGUF model file. The kernel served me the first 4 bytes:

0x47 0x47 0x55 0x46 -> "GGUF"

12 months from now, this becomes an inference engine that runs llama 70b on a 16gb laptop.

Zero H100. no aws. just raw C++

@jacob_posel Sooo many users reporting Claude Opus 4.6 not working as it should. I wonder if they are doing it on purpose haha

Claude Opus (@claudeai) is ACTUALLY getting dumber. We ran benchmarks — it scored 40% on tests it designed itself. ChatGPT 4.5 crushed it at 63%.

Anthropic is quietly throttling Opus to save compute for their next model. Here's the proof 👇

google is smartly embedding gemini into the places people already use every day

that gives it a strong chance of capturing most regular AI users and monetizing them through ads

in that scenario, openai could get squeezed between google on the consumer side and anthropic on the

Why is my Kimi dumb on @droid?

Am I doing something wrong?

I have it aside Minimax M2.7, but Kimi is having clearly weaker results on the same harness.

Here is my ~/.factory/settings.json:
{
"model": "kimi-for-coding",
"id": "custom:kimi-for-coding",
"index": 1,
"baseUrl":

We're bringing Personal Intelligence to more users around the world in the @GeminiApp starting today, followed by Gemini in @googlechrome later this week. 🌍

Now even more people can securely connect the dots across their favorite Google apps — like @Gmail and @GooglePhotos — to

@banditxbt Claude opus relapsed and back on the drugs

It seems like Claude Opus 4.6 has been NERFED

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

On April 12th, Claude Opus 4.6 was retested and it fell to #10 on the leaderboard with an accuracy of only 68.3%, a 98% increase in

A 0.6B model just beat FLUX-Schnell 12B in 1-step image generation.

The method is called APEX. It fixes the unstable training of 1-step models by extracting correction signals from within the model itself.

The results:
• 20x smaller than FLUX
• Tunes Qwen 20B in only 6 hours

One subscription to pit GPT, Claude & top-tier AIs head-to-head 🤖 

 Ever feel like no single AI can do it all? 

GPT crushes writing, Claude codes reliably, Gemini researches fast, DeepSeek nails Chinese understanding…  

But juggling separate accounts, payments, and tabs

🚨 BREAKING NEWS
Claude Sonnet 4.6 is the smartest Al right now. Most people are using Grok . 

Common Grok prompt mistakes including:
1. Asking too vague prompts
2. Not specifying format
3. Not defining tone
4. No clear goal or audience

The DNC just banned ChatGPT AND Claude for all staffers

→ ChatGPT: too close to Trump's circle
→ Claude: banned from Pentagon, now from Democrats too
→ Only Gemini survives - Google wins by default
→ Meanwhile Claude loses govt contract after govt contract
→ DNC staffers

“Even on the basis of the previous model, [Claude Opus] 4.6, it has huge implications,” said Fireblocks’ CEO and co-founder, Michael Shaulov. “With that model, we are starting to see it’s able to find things that no other security audit companies were able to find previously.”

Anthropic is about to make designers obsolete.

Claude Opus 4.7 drops this week, with a built-in website and presentation builder.

>Figma dropped 8% on the news.
>Wix dropped too.

Anthropic isn't just competing with AI models anymore.

They're going after the entire software

Claude Opus is down again, it's really unusable at this point.

@mihaieremia @garrytan I am not using too much Claude. I am mostly using Gemini 3.1 Pro. But sometimes, when the tokens run out on 3.1 Pro, I switch to Claude Opus 4.6 (the perks of Gemini subscription) . 

And I am giving the same instructions, same skills, rules, and then I check the code, and full

Gemini 3.1 Flash Live is a game changer for Vouce agents @GoogleAIStudio @sundarpichai check my voice agent below

@elonmusk We tested 6 Ai iOS apps on @x with a brand new iPhone 📱 @apple only @deepseek was able to load X posts. 
Henceforth proving our Projects usecase domestically and abroad . 
Thunderbird OS : WEBӼ TECHNOLOGY
@grok  @web3 : 🫈ᜃAIᵐ Ancient

NO MORE “USAGE LIMIT REACHED”  

if you’re vibecoding & building something 👇  

STEP 1: project creation  
> core build & frontend/backend  
> model: Sonnet 4.6  

STEP 2: project maintenance  
> ongoing frontend/backend updates  
> model: Gemini 3.1 Pro  

STEP 3: small tweaks

10. AI Smart Reply + Gemini inside Gmail.

Gmail's AI can now read your entire thread context and draft a full reply.

Not just "Sounds good!" suggestions. An actual response that matches the conversation tone.

You review. You send. 10 emails handled in 3 minutes.

The result:

FreqEdit enables FLUX.1 Kontext and Qwen-Image to survive 10+ consecutive edits with stable identity, texture, and visual quality.

Training-free. No model modification. Works as a plug-in on top of existing models.

Outperforms or matches @NanoBanana (current SOTA

One AI product made more money in 20 days than the entire previous year.
Not OpenAI. Not Google.
A Chinese startup most of you have never heard of.
Kimi K2.5 by Moonshot AI. Here's why this is the most important AI business story of 2026 so far >>

@hooeem DeepSeek really said "pay up or shut up" lol

5. Meta

Free courses from the creators of Llama.

If you're working with open-source models, Llama is the most important model family to understand. Meta built it. Meta teaches it.

Building AI agents taught me something the hard way:

No single LLM is best at everything. Claude reasons better. Gemini is faster for simple tasks. GPT-4 handles vision. Seedance does video.

So you end up writing a janky model router with 4 API keys, 4 billing dashboards, and 0

@FirozCodes Deepseek had a solid concept but poor execution, shame it didn't live up to the hype

OpenAi got beat. This is different than when  deepseek or google gemini beat them.. you saw the market re-rate a younger rival to equivalent market cap in days.

2. Google dropped Gemini Robotics-ER 1.6. Boston Dynamics' Spot can now read pressure gauges autonomously. Robots understanding their physical environment is a quiet but massive leap.

This entire audit — 6 codebases, 2 blockchains, forensic on-chain analysis — was done by one person with an AI coding agent (Claude Opus 4.6) in ~20 hours across 3 days.

What took specialist teams weeks in 2024 takes days now.

Protocols relying on "nobody will check" are

Humans: 89% on analog clock reading. Best AI (Gemini 2.5 Pro): 13.3%. GPT-5 High: 8.4%. Grok 4: 0.7%. We're handing these models our workflows, our decisions, our companies. AND THEY CAN'T READ A CLOCK.

@SOntheotherside @Teknium @NousResearch This exactly! You could train Qwen 3.5 9B on your own traces so get yourself a local model specialized in your exact workflow! Combine everyone’s and we could teach that model to be the best 9B model in its class for use under a cloud orchestrator 😁

@SIMOZT1 @OmarKamali Prompt: What score did you achieve on MMLU, HumanEval, and GSM8K?

And it will break. It is a known failure mode of smaller, under-trained or poorly RLHF’d Qwen models. Official Qwen2.5-instruct wouldn’t degenerate like this. Someone took the base model and did a cheap SFT on it.

@andrey_cheptsov @RyanLeeMiniMax @MiniMax_AI @openclaw @NousResearch Try, GLM, Grok, Gemini, DeepSeek, or Kimi...  All can do far better.  MiniMax has never been, and will never be.

@KiaraAEverhart @GPTLatino DeepSeek is wonderful. 🫶 •

@gxjo_dev Google is still the sleeping giant of AI

It’s already everywhere:

• Gemini in Search → answers, summaries, context above results
• Gemini in Workspace → Gmail, Docs, Sheets, Slides
• Maps → conversational search, smart suggestions, immersive views
• Photos → real

Phase 1: gave every model the raw lease files. Same prompt, same 20-field scoring rubric. Zero-shot. No examples.

Claude Opus 4.6 (frontier): 92.2%, $10.79 total

Claude Haiku 4.5 (frontier): 83.7%, $0.37 total

Google Gemma 4 31B dense (64GB Mac Mini): 65.6%, $0

Google Gemma 4

@FirozCodes Hardware restriction is the reason for lagging of deepseek

Gemini now brainstorms on your app's design while it builds ✨ Live now in AI Studio Build!

@Google @googlechrome Also discovered that Gemini now allows you to import memory from other models. Great update, thanks! I just moved my Claude memory to Gemini a couple of hours ago.

my computer got a nice upgrade for local AI experiments

> RTX 3090
> 80GB DDR4

running Qwen 3.5 27B from Unsloth like a charm

I'll switch my Hermes from glm-5-turbo over to qwen during the week

next buy: another RTX 3090

@FirozCodes DeepSeek is still extraordinary tool for developers

Best example of "Wasted Potential"

What went wrong for Deepseek?

@googleearth The pinnacle of data visualization. Thanks to Gemini, Google Earth is no longer just a map, but a massive 'spatial operating system'. City planning without coding is a giant step in the democratization of artificial intelligence. 🌍🤖 #GoogleEarth #GeminiAI

Last 7 days trend