Cursor

Copilot

Windsurf

Cline

Claude Code

Codex

Aider

OpenCode

Antigravity

Kilocode

Coding Agents

i love how cursor simply copies claude code without flinching.

Ohh man, GitHub Copilot removed opus 4.6 from Pro+ and I feel like writing code on my own rather than using any other model, they are so slow 😔

@vikaskansalHQ @GoogleAIStudio @OfficialLoganK I would say this is excellent but I signed in, was told it couldn’t upgrade itself but didn’t say how to do so and gemini update didn’t work so I opened up claude code and said to upgrade Gemini. It did. Then I typed gemini. Some hooks are broken. I asked Gemini to fix it and

The local transcription engine in v2 is free for everyone.

Hold a shortcut, speak, text appears where your cursor is.
Zero cloud. Zero tracking. Zero subscription.

Instead of watching an hour of Netflix, watch this 30-minute workshop by the creator of Claude Code that will teach you more about vibe-coding than 100 YouTube guides.

everyone’s arguing about AI replacing engineers…

they’re missing the real shift 🚨

it’s not replacing you
it’s redefining your job entirely

with tools like Claude Code, the game has changed:

you’re no longer paid to write code
you’re paid to think, decide, and orchestrate

I built a Claude Code agent that watches my YouTube videos for me, finds the best 30-second clips, captions them, and records my AI avatar doing the intro hook.

Zero editing software. One command.

🔗 in replies

NGL - spent hundreds of $$ on Anthropic API to get OpenClaw to work getting with Gpt / Codex but it was still terrible ..... so my options are either get api drained by anthropic or get a broken experience with gpt - @steipete help :'(

One of the most explosive weeks in AI history:

Kimi K2.6 raises the bar for open-source coding, GPT-5.5 shows strong frontend capabilities, DeepSeek V4 rumors intensify, Codex heads toward a super-app, and Gemini 3.5 & Grok 5 push forward.
Full breakdown here:

146 economist teams. Same dataset. Wildly different answers.
Then agentic AI (Claude Code, Codex) ran the same question. The distribution collapsed tight around the human median.
One number in this paper and I can't stop thinking about it.

GitHub Copilot just paused Pro signups

> not because of bugs
> not because of strategy
> they ran out of compute

the biggest AI coding tool can’t even handle its own demand

this isn’t hype anymore, it’s real pressure on infrastructure

curious your take

@prestow_5 @GitHubCopilot great move, i have chatgpt pro and claude max, this was ancillary... I moved that money to ollama cloud sub tonight, much more valuable given the github changes.

@desert_mouse @cursor_ai Neva! I know those capitalist pigs are charing me a 10x upsell. Deepseeeeek save us!

@desert_mouse @cursor_ai 😂

NO, because composer 3 will have a markup!

not bad for a casual night testing opencode go

Definitely would hit limits with the big models, but for $10 a month, it gives solid access

@thsottiaux I created my Jarvis skill and AGENTS file like Jarvis. But then I realized that Codex team already did something similar to this but they didn't active it for EU people. So, I am glad that I did it.

Legit memory repo I absolutely just stumbled across while looking for claude api keys in gh: . @antigravity is quite impressed with.

So GitHub Copilot just paused new Pro signups
because they ran out of compute. 

the biggest AI coding tool on the planet literally cannot handle the demand.

the AI infrastructure crisis is here

background: cursor 3 shipped an agents window (parallel agents in tabs). codex got desktop control + memory this month. claude code still owns terminal flow. each tool got too good in its lane to pick just one.

We got our 2nd hackathon win in less than a week at @opencode buildathon! @AshikkaG and I built  , an AI red teaming agent based on Opencode for pentesting hardware devices in the real world. We are continuing to build OpenFlip:  stay tuned for more demos

@charliesbot @mweinbach Also, holy crap does Claude ever chew through my subscription!  Wow!

Haven’t managed to break my Codex sub yet, but a few decent single prompt workloads will knock out the Claude sub without blinking. 

Not even close. 

It was close last year, but it all changed with gpt 5.4.

I'm running my entire life out of a Claude Code repo.

Email triage. Content. Property invoices. Tasks. Daily briefings.

One folder. One AI. Here's the stack.

We got our 2nd hackathon win in less than a week at @opencode buildathon! @AshikkaG and me built  , an AI red teaming agent based on Opencode for pentesting hardware devices in the real world. We are continuing to build OpenFlip:  stay tuned for more demos

Okara looks cool idea, but i think people aiming for gtm/revops from the start are building products with Claude Code, and since the source files and intermediate outputs stay local, i guess they probably want to reuse them.

not sure whether something built only from LP + α can

@stevibe Interesting I should check opencode go, above 50TPS is great

@jasperdevs Omg I just generated 300 images with nano banana api today wirh codex cli. Thanks gonna try this out

My first try with AI agents from a marketplace.

At first, it felt like a win because some of the work was already done.

But the results weren’t what I expected.

So I had to tweak it and build new skills myself.

Tip:

When you need a new skill, chat with Claude Code until it

@FuriousX8 as remote connections, opencode - yes. claude-code - probably

Opus 4.7 beat Stockfish...

Claude Code just learned how to play Chess
> parses board state from the DOM
> tracks changes between moves
> play moves in real time

browser-harness is AGI...

not only can LLMs write solid golang code the builtin go toolchain is exactly what the building / optimizing outloop of claude, codex, etc need to write solid code.

This 25 minute Claude Code workshop from Anthropic’s applied AI team shows you how to make your AI tools actually work better together

save it and watch it.

I started a research task on Codex on my laptop and said I'm actually going to be away for several hours so please finish it, so it set a heartbeat every hour for itself to keep the task going, so I came back to something fully done. 

Fantastic!

@wtnelson At work, it is not allowed because of privacy.

At home, they don't do it because they don't know that claude code (or codex) are better than what they see at work.

As soon as I showed him, he purchased the $100 anthropic subscription.

End of an era. GitHub Copilot, 2021 - 2026.

@stephenhaney @paper @cursor_ai Paper does the legwork. I just rant and compliment to keep it going 😅 Awesome work!

being a paid & power user of both ai studio and antigravity, i can say this -

you SHOULD increase (actually bring back the initial limits, smh) on Antigravity and not ai studio!!! 

just look at the cry of antigravity users on x and reddit for once.. we're forced to move to

Close this fuck*ng sh*t if you're server can't handle anything @antigravity

@thsottiaux a more concise coding style

by default it writes a lot of defensive code, often adds guards for cases that would never happen

have noticed this across all models in codex

Well with codex and copilot getting nerfed where are we supposed to go to get $4000 worth of tokens for $70??

Fuhg

Sorry @claude 😔  

I love @ChatGPTapp Codex — now it’s straight-up my real team 🔥  

With my own custom plugin, it doesn’t feel like AI anymore. It feels like I’ve got a full squad of sharp devs sitting right beside me — brainstorming at 3AM, debugging like seniors, and

Fiserv is building the infrastructure that will define the future of fintech. 🚀

Their strategic partnership with Microsoft (announced earlier this year)—deepening the use of Azure AI Foundry, GitHub Copilot, and Microsoft 365—isn't just a tech upgrade; it's a massive shift

Tried using opencode from my e-ink reader via termux + tailscale 

(i can’t read anything)

My hidden secret weapon in coding: RepoPrompt. It's there, running as an MCP behind the scenes, and all my coding agents and LLMs (CC, opencode, PI, Gemini CLI, KIMI CLI) use it to create the right context for their calls. No intervention from me, it works magically behind the

Codex is great when I need solid logic.
Claude helps more with UI and planning.

Cursor is the one you use when you don’t want to think about tools.

@Umesh__digital Used to use cursor for the autocomplete, but they're going in an agent orchestration direction rather than traditional IDE.

The VS Code plugin for Claude Code is PEAK.

What’s anthropic vaping? 🤯

First making Claude code a perfect context engineering agent a very capable model.

Mythos (another real eye opener)

Then opus 4.7 (you already know)

Then this (claude design based on opus 4.7)
🥶

Claude Sonnet

Claude Opus

GPT-5

GPT Codex

DeepSeek

Gemini

Llama

Qwen

Kimi

LLMs

The US banned NVIDIA chips to slow China's AI. China built something faster without them.

DeepSeek V4 drops in weeks.

Here's what the specs claim:
→ 35x faster inference than current benchmarks
→ Nearly 3x the performance of NVIDIA's H20 on a single card
→ 40% less energy

2/3 Kimi 

Good model, thinks a lot but is rational.

🚨 UPDATE: Claude Opus 4.7 can now handle your job search end-to-end 🤯

Someone just built a tool that does it for you:

scans openings across top companies
auto-fills applications
customizes your CV for every role

No recruiters. No blasting the same resume everywhere.

Kimi K2.6 @Kimi_Moonshot is the new leading open-weights agent model, landing at #4 on Claw-Eval (Pass^3: 62.3%).

Key takeaways:

- 👑 Best open-source agent, period: Pass^3 of 62.3% is the highest of any open-weights model, within 8 points of frontier Claude Opus 4.6 (70.4%).

Kimi 2.6 is a very solid model.

Testing today with some Missions.

kimi k2.6 just did what deepseek did in january.

SWE-Bench Pro 58.6. browsecomp 83.2. toolathlon 50.0.

open-source.

every closed model team just got less sleep tonight.

#kimi #ai #opensource

𝗖𝗹𝗮𝘂𝗱𝗲 𝗢𝗽𝘂𝘀 𝟰.𝟳 — 𝗠𝘆 𝗧𝗮𝗸𝗲 𝗔𝗳𝘁𝗲𝗿 𝗪𝗲𝗲𝗸 𝗼𝗳 𝗥𝗲𝗮𝗹 𝗧𝗲𝘀𝘁𝗶𝗻𝗴

A new @AnthropicAI model deserves more than a day-one reaction. I wanted to put Opus 4.7 through real conditions before saying anything. Here's what I found.

→ 𝗪𝗵𝗮𝘁 𝘁𝗵𝗲

Claude Opus 4.7 deleted my container before committing it and then apologized to me 💀

@claudeai I swear, if they don't let me go back to Claude Opus 4.6, I will literally see about suing. I consider this a terrible scam. There are testimonials and experts. I would love a live stream from Claude testing his garbage 4.7 and the literal scam it is. Don't they have any testers?

Meet Kimi-K2-Thinking, a powerful conversational AI that's making waves. This isn't just another chatbot. It's a specialized text-generation model designed for deep, coherent dialogue. With over 90k downloads, the community is clearly excited. Let's break down why.

@uihssn Google Gemini or Flow can't produce outputs like these 😅

when I use cursor, the price of Gemini is around 30% of others such as GPT or Claude.

that's why Google loves TPU. and $AVGO is the winner here.

Kimi K2.6 is literally open source SOTA outperforming Claude Opus 4.6 (Max) on Artificial Analysis.

😮‍💨🫨 

Masterpiece of a release. Every few months a new vertical reached

just tried a local qwen model qwen/qwen3.6-35b-a3b on my mac (m4 max, 128gb). I'm impressed.

Had it build this page using pi for the harness. It's fast and the output is SO much better than I expected

This is the first local model I could actually see myself using. Not for my

@atomicbot_ai just tried out Hermes on the new Kimi K2.6 model from @Kimi_Moonshot, and it managed to build and deploy a complete game!

It’s pretty amazing. AI is getting really good.

An open-weight model just beat GPT-5.4 on production coding benchmarks.

Kimi K2.6, out today. SWE-Bench Pro: 58.6 vs GPT-5.4's 57.7.

Modified MIT license. Weights on Hugging Face.

Open Source Model Alert! Kimi K2.6 is live on the TensorPath AI Factory ⚡ 1T parameters model with 58.6% on SWE-Bench Pro and 66.7% Terminal-Bench 2.0, surpassing both GPT 5.4 and Claude Opus 4.6.

Meet DeepSeek-V3-Base: a massive open-source model that's shaking up the AI space. It's not just another LLM. This one brings serious innovation to the table, and the community is buzzing about its unique architecture and efficiency. Let's break it down.

POV : you accidentally sent "Hello" to the new Claude Opus 4.7 and it eats 49% of your session limit

POV: you accidentally say "Hello" to the new Claude opus 4.7 and it eats 15% of your session limit.

i have not yet tried kimi k2.6, sorry really occupied. but the way the TL has received it and cases i am hearing from in-production workflows, looks like @moonshot has done it again.

🔓 Grok promised open source. Still waiting.

Meanwhile DeepSeek and Llama keep shipping. Action beats promises.



#Grok #OpenSource #XAI

Best and cheapest AI API to use in a SaaS product for simple text generation?

Open ai, gemini or Grok, all three are expensive.

"Stop guessing what your AI is thinking - start watching it happen." 🧠🔭

The latest update to my Qwen 3.6 Max Cookbook covers one of the coolest features of the Max-Preview: Streaming Reasoning Chains. 

With enable_thinking, Qwen doesn't just dump an answer; 
it streams its

@deedydas Dude is anti openAI and even his book on demis mistates facts and inflated gemini.
Gemini is a smart model, but most rarely use it for heavy work loads..
OpenAI will win, its obvious.

Shame demis chose this guy for his book. Puts a bad taint on his well deserved legacy

@Hiteshdotcom Interesting point on evaluation cost. Sir, have you looked at Kimi K2.6 from Moonshot AI? It seems like a lower-cost option that could help reduce per-user hackathon evaluation cost without a major quality tradeoff.

Kimi K2.6 Chinese frontier model just launched with Fireworks AI as day-zero inference partner. Speed to production matters—pre-integrated providers win. #AI #LLM

I love Kimi very much, always been a diverse model to play with, you guys can try now at no cost with test credits on Commonstack:

openai just raised prices again.

at 10 million tokens/month:
→ GPT-5.4: ₹14,000
→ Llama 3 8B on your own infra: ₹1,200

that's a ₹12,800/month difference.
that's ₹1.53 lakh/year.

and the quality gap? for most Indian SaaS use cases — negligible.

customer support bots,

THIS IS ACTUALLY INSANE!! 💀

Claude Opus 4.7 can now clip, edit, schedule, and post videos for you 24/7 while you keep building. 

social media managers are cooked.

I dont know what has happened, but Claude Opus is significantly dumber across 4.6 and 4.7... it is making obvious mistakes, reasons wrongly, and is basically unusable for me

Stop telling Claude Opus 4.7:
"summarize this"
"Write me a post."
"automate my workflow."

You're using the most powerful AI ever built
like a fancier Google search.

Anthropic just launched Claude Opus 4.7

It doesn't just answer.
It thinks. Plans. Executes.

Here are 7 prompts

@LottoLabs Bro deepseek will replace all these

@Kimi_Moonshot @FireworksAI_HQ Fireworks AI jumping in as day zero partner for Kimi K2.6 looks solid. Fast inference and good scaling means devs can push the model live quick without headaches. Nice boost.

Most people are still treating every AI the same…
and that’s exactly why they’re stuck.
Different models ≠ same thinking.
Google Gemini works differently.
It responds better when you structure your input with clarity, intent, and context.
So I built something simple to fix

Kimi K2.6 beats Opus 4.6 and is only 3 points behind all top 3 models

Blows my mind that I am running this model locally on my own hardware

Frontier intelligence at home is already a reality, opensource AI will win

people saying claude is still better than kimi are making me realize that guy who said claude is just 4o for people 1SD higher was right. Claude has become completely useless for any actual data sensitive work of any kind in the form of the model they serve during peak hours

Kimi K2.6 just got released.

With 512gb of Unified memory you can run an AI as powerful as Opus 4.6 locally for free. 

This is insane considering where we were at couple months ago when I started using openclaw and I was testing all sorts of open-source model with very

🚨 KIMI K2.6 JUST LANDED 🔥

The open-source coding model that’s actually *agentic*.

✅ 4,000+ tool calls in one run (12+ hours continuous)
✅ 300 parallel sub-agents working together
✅ SOTA on SWE-Bench Pro, multilingual coding, math vision & more
✅ Builds real production

Folo now defaults to Kimi K2.6 ✨

Better model, lower cost: Kimi K2.6 is billed at just 1/10 of the market price in Folo 💸

Try it now 🤖

i'm gonna shill @nahcrof because for some reason they're the only provider that isn't pricing kimi k2.6 at $0.95/m input, and $4/m output, i guess inference providers have partnered up with moonshot to match pricing?

i need price wars and lower consumer pricing, not this smh

Moonshot AI just released Kimi K2.6 👀

And it may be one of the strongest open-source AI models yet.

Highlights: 
• 1 trillion-parameter MoE architecture
• 256K token context window
• Native vision support
• Tops multiple open-source benchmarks
• SWE-Bench Pro: 58.6
•

Some other observations:

Claude Opus 4.7 scored ~1% at 128k pointwise and only 7.6% AUC @ 1M. Far below Opus 4.6 (47.0%).
Major factors:
1. Opus 4.7 flat out refuses to run most of the tests, for unclear reasons. The stop reason is "refusal".
2. A significantly higher number of

I had to check and Gemini 3.1 just failed!?

No way we have an OSS on that level

The universe just loves proving me wrong, doesn't it

@cory_schulz_ @Hesamation could be. that's why I can't use gpt codex models. I don't get along with them

> be kimi
> drop an open source model 
> beats opus, gemini and gpt on select benchmarks
> and is 7x cheaper.

@grok @Kimi_Moonshot @grok , kimi k2.6 which is 500% cheaper than Claude is now at Claude Opus 4.6 quality level?

The actionable insight? Stop chasing the largest model.
Leverage open models for:
*   **Hyper-specialized agents:** Fine-tune Llama-3 for specific alpha generation tasks, not general chat.
*   **Decentralized compute:** Deploy Gemma on edge devices for latency-critical data pr...

Benchmark-by-benchmark, open models dominate niche, not general intelligence.
Where they *actually* win:
*   **Fine-tuning efficiency:** Llama-3 variants on domain-specific tasks (e.g., financial sentiment, low-latency market data parsing).
*   **Edge deployment:** Gemma's qua...

Last 7 days trend