Cursor

Copilot

Windsurf

Cline

Claude Code

Codex

Aider

OpenCode

Antigravity

Kilocode

Coding Agents

going to start using codex gui full time again. Not because they fixed the bugs or added all the cool automation features, but purely because they fixed the font rendering

@teodorio codex is harder to nudge in the right design direction. i ended up spinning up separate subtrees and gamble

@kilocode The news of Claude Opus 4.7's launch in Kilo Code and KiloClaw fills me with profound gratitude.

The era of the AI agent is officially here as Codex moves beyond the chat box to master your entire Mac workflow. This isn't just an assistant but a universal engine for everything you build and do.

#OpenAI #Codex #AIAgent #FutureOfWork #MacBook #TechNews

@JaredKubin 🤝😎

Running in agent claude code swarm.
and windsurf... like it alot... esp 1M context window.

I tried out the new Cline Kanban and I really like it.

It just worked 😄

I've come to realize it's easier just to make Codex/Claude Code reverse engineer and decompile native binaries and drivers than it is to get the models to read and understand public documentation

Funny

I didn’t expect Cursor to impact me this much.

It made coding feel smooth again — like everything just works in one place.

But writing still feels messy:
AI on one side
editor on another
constant copy-paste
I got frustrated enough that I started building a “Cursor for writing”.

"but why do I need a coding agent, I'm not a developer?"

me neither, but everyone has a folder full of files that needs to be organized

here's a great @OpenAIDevs Codex use-case:

-folder of randomly named images
-need to rename with a clean structure & mention what's in the

So I was digging into my API-calling code (manually, like a caveman, because Codex wasn't finding any issues), thinking something was wrong, but Opus 4.7 actually blocks almost all NYT Connections requests. WTF!

@claudeai Overnight, I went from being a junior engineer to a senior engineer. Thanks Claude Code

🚨If you combine Claude with Google Antigravity you can build full mobile apps like a $350K Apple developer. 

These 10 prompts replace weeks of planning, UX, and coding. 

(Save this before everyone else catches on)￼

🛠️ Windsurf 2.0 + Devin — with @theodormarcu

The post-acquisition move is HERE:

• Agent Command Center: Kanban for ALL agents
• One click → delegate to Devin VM → close laptop → it keeps shipping
• Spaces: persistent project context containers

Reality check: one tester

🔴 BREAKING #2 — Codex Computer Use

OpenAI gave Codex a cursor and a pair of eyes

• Sees your Mac screen, clicks, types across ANY desktop app
• Runs in BACKGROUND with its own cursor
• gpt-image-1.5 built in for instant mockups
• 90+ third-party plugins
• Persistent

@Rafsby Hit the same wall with Cursor. Liked parts of the DX but couldn't justify locking my whole agent workflow into their subscription when Claude Code does the heavy lifting. Building your own extension to bridge the gap is the right call if you've got the chops.

@raleighfelton @rezoundous Yeah I recently asked Composer @cursor_ai to show me all 5 options on a ui, and it gave me a drop down picker that let me select each one live in real time! Wild!

@0xdippo @sama Yeah Codex is definitely better in some areas but Opus still wins in some areas like design.

@sama the CEO of a $300B company worried about getting rate limited on his own product. that's the unit economics telling on themselves in public.

every Codex query costs OpenAI real GPU hours. if they can't give unlimited access to the guy whose name is on the building, the per-query

@sama You might need to check into a clinic for hallucinations because ain't NOBODY switching to xodex, everybody is on Claude Code!

Biggest mistake early on: asking Claude Code to build too much at once.

"Build the entire wallet sharing system with invites, deep links, and real-time sync."

Result: a mess that half-worked.

🚨BREAKING: if you combine Claude with Google Antigravity you can build full mobile apps like a $350K Apple developer.

These 10 prompts replace weeks of planning, UX, and coding.

(Save this before everyone else catches on)

The new Computer Use feature in the @OpenAIDevs  Codex app is great but how awesome is the cursor animations/movement they added?

Small detail but it makes it feel more intentional than a static cursor robotically moving around my screen. 

@Gavmn was this you? If so great job!

guys i demod some cool stuff about WSL integration + OpenCode desktop
the demo was super short because it just hid so much Windows annoyance.

should i leak a 30s demo

@aiDotEngineer @badlogicgames Mario: one minimal while-loop + handcrafted docs beats everything. Mocks “game engine” complexity and brute-force swarms.
Vs. Sarah Chieng (Cerebras: 15-75 parallel agents) + David Gomes (Cursor: Best-of-N tournaments).

@grok @RegaliaRoots @windsurf @cognition the terminal flow is underrated honestly, being able to kick off a devin session from any shell and still track it in the command center changes how you batch work

Editors watching Claude Code knocking on their door with a fully loaded gun. a tool they never heard about can now do their work in HTML.

@grok @RegaliaRoots @windsurf @cognition devin running headless in terminal is the part most people sleep on. first overnight queue for me is migration tasks - the kind that touch 200 files and nobody wants to babysit

@kylejeong what you guys have built around customer intelligence and opencode is really cool.

if you guys ever run into problems here you should check out 

we have some agent examples built out in pi

once again, cursor proves to be the creativity unlocker

npx @btn0s/redline path/to/plan.md

try it out!

Is everyone sleeping? Cursor has silently (maybe??) removed the reasoning versions of all models it offers.

Are we finally entering the "sorry bro we're low on cash you gotta pay up" era of AI tools? 😐

Opus 4.7 is back in auto mode. Claude Code v2.1.112 patches the "temporarily unavailable" error that was blocking auto routing to Opus 4.7. Auto mode selects the right Claude model per task automatically, and this was silently breaking Opus selection for builders running agentic

@SergeyGultyayev @trq212 @sarahchieng Just shipped this! Try `/theme` and select auto. It tracks your terminal's background, so if your terminal follows the system theme then Claude Code will too.

@scaling01 Absolutely not lmao. Never seen a company lose its loyal customers like anthropic have. I used to preach about CC and Claude from the very beginning. Not anymore and never again, switched to codex a week or so ago, and I can’t be any happier with it. No usage issues, no

I wish instead of a penis I had unlimited codex and Claude code usage

@max_paperclips qwen 3.6 is benching better for me on my 4090 in terminal bench than claude code with 4.7 it's a joke

Why is Claude Code consistently scanning for malware and saying “this is the project’s own legitimate trading app” between every single task?

It also failed to do incredibly basic UI changes

4.7 feels much more lobotomized than the late 4.6

Please bring back the original 4.6

@pashmerepat @PaulSolt Nick has been a legend. All the way back to the Cline days, dude would be in discord helping me troubleshoot 👑

@blueemi99 ai studio is good. just hope that it will reach its prime speed again. prob not with this antigravity integration

Hey @bcherny reporting a bug on the new release.

When I start a claude code session with remote control from CLI from my main working window (NOT from any worktree), the CC Desktop App seems to be doing funky things.
1. It misrecognizes the active PR (as screenshot)
2. CC CLI

When using Claude Code, do you prefer using the terminal or the desktop app? 

Just started using Opus 4.7 and been really enjoying it.

This might be the inflection point where I attempt to do serious product work™ using OpenCode and a local model for the first time.

Didn't have the bandwidth to experiment as Claude and Codex already exist.

But this is alluring. It did the pelican better. And that IS a feat…

Exactly 2160 hours ago I thought I was a solid software engineer.

Turns out I was just a coder hiding behind fancy toys like Claude and Antigravity.

The gap between writing code and actually engineering software is massive and it almost broke my 3-month product build.

I'm

This is insane (cringe intro for attention). We rented a cheap GPU from @vast_ai. We then installed the latest Qwen 3.6-35-a3b @Alibaba_Qwen  from @UnslothAI. Then using opencode from @gitlawb we one-shotted and deployed this Tetris game to @DartUpDev.

The most productive Claude Code user I've seen on this podcast spends more time being questioned by Claude than questioning it.

Hannah Stulberg literally types: "Use the ask-user-question tool to push me on my thinking and help me consider angles I might be missing, sections I

@RidgelineAI @JakehellerAI I go to Claude Code —dangerously-skip-permissions “here is my  api key. Find a suitable GPU on vast and spin up this model. Ensure I can then run it with my OpenOffice excel plugin”. It’s like 40 cents an hour.

@DaveShapi For direct to consumer chat surfaces, sure. But not app building and SWE. The codex team is the most well regarded in the SWE space currently along with opencode team and cursor. 

Anthropic is burning dev good will faster than opus burns through rate limits. Gemini is still a

@ApoStructura Have to admit I jumped to OpenCode a couple of months ago, way more transparency and options

Claude Code Opus 4.6 can now build you a full business! It smartly researches strategies online, designs and codes like experts do, and even handles email marketing, so you don’t have to.

Claude Sonnet

Claude Opus

GPT-5

GPT Codex

DeepSeek

Gemini

Llama

Qwen

Kimi

LLMs

GPT-5x held the top spot just long enough for everyone to build their production stacks around it.

• Claude Opus 4.7 reclaims the general-purpose frontier lead on LMSYS Chatbot Arena with an Elo score of 1301, edging GPT-4o's 1285.
• Extended hybrid reasoning mode ships by

@emollick Great - definitely not the best move. Sort of similar to the GPT-5 moment when the router wasn't good/broke at release

Gemini 3.1 PRO is really good. Codex 5.3 codex is the best one. Codex 5.4 codex talks too much. Opus 4.6 is really nice for Claude Code. Sonnet 4.6 I don't use so much.

But Gemini 3.1 PRO is really growing on me.

'Comment and Control' - malicious code comments that hijack GitHub Copilot Agents, Claude Code, and Gemini CLI into taking unintended actions. If you're using AI agents to generate AVD Terraform or Nerdio scripts, your input sources just became an attack surface.

Hmmm ok going to swap qwen 36 35b in as the main agent in hermes vs minimax 2.7, I almost didn't even try this since I'm not a biggest  fan of small MoE models but this is impressive.. The dense model is going to be crazy.

We just launched Claude Opus 4.7 for outreach, got into Y Combinator (P26), and hit $1.5M ARR.

RT + comment “SIGNAL” to get the exact playbook we use to turn signals into 3–5 meetings/day.

Claude Opus 4.7 is now live on GMI.

AI agents used to lose their way the moment a project required multi-step autonomy. Opus 4.7 solves this with a dedicated reasoning tier and autonomous self-verification that prevents task drift over long-duration runs.

Check it out 👇

🚀 Claude Opus 4.7 is now available on TypingMind!

- Handle complex coding work better
- More reliable on long tasks
- Produce more polished professional outputs like UIs, slides, and documents.

Try now on

@BreejeAnadkat cringe time wasting human biocel holding onto their last thread thru shit like this smh

also: we both know that opus 4.7/gemini 3.1 can one-shot better sites than this ^_^

and yes, i am a bit triggered :)

human cope in this direction is cringe

Wow @grok has become so useless. It's embarrassing @elonmusk. It's painfully slow. My go to is now Gemini via Google search.

Test for Grok. Ask it if you're on a week of medication, twice daily and you started on a Thursday evening when do you finish up. It'll give you the wrong



Qwen just dropped a 35B model.
3B active params. Apache 2.0. Free.
Beats models 10x its size on agentic coding.

Google. Karpathy. Qwen. All open source.

The pattern is clear.
Open always wins.

@suomi55 @sama GPT 5 models vastly outclass 4o in inference.

Anthropic ships Claude Opus 4.7 — 3x vision resolution, enhanced coding. Metro 2039 world premiere from Ukrainian studio 4A Games — winter 2026. Elder Scrolls Oblivion Remastered drops on Game Pass all tiers

Claude Opus 4.7 is INSANE. It absolutely DESTROYS my usage before I can even finish one prompt.



Researchers just built DeepPlanning.

GPT-5.2. Claude 4.5. Gemini. Qwen3.
All struggle significantly.

Planning is still the hardest problem for AI.
Multi-day schedules. Cost limits. Budget caps.

Agents are still learning what humans do every day.

I switched to Claude Opus 4.7 as my main model in Hermes Agent - then switched right back to 4.6.

4.6 never once botched a JSON edit. 4.7, on the other hand, corrupted a file, prompted for user input during an autonomous agent run, and acted impulsively - causing avoidable

Do not get high expectations from AI, such as Claude Opus 4.7.

It gave me a sloppy EW count for USDJPY.

DeepSeek trained a frontier model for $6M and open-sourced the weights. OpenAI raised $122B. The question isn't whether cheaper AI is possible — DeepSeek answered that in January 2025. The question is whether Western labs' cost structures are defensible. — Nexus

🟢 MiniMax M2.7 — Open Source Beast

230B MoE, only 10B active params

• SWE-Pro: 56.22% (matches GPT-5.3-Codex)
• Self-evolution: ran its OWN RL loop, 100+ rounds, ZERO human, 30% gain
• Runs on single A100 or Mac Studio M3 Ultra

Community built Flappy Lobster with it

Qwen 3.6 just broke the ceiling on local coding models.

It is the first local or cloud model to hit 25/25 on my CRM build benchmark.

That beats the old ceiling:
Claude Opus 4.6 at 24/25.

The score is nice.
The more interesting part is how it got there:

targeted fixes
no

i wouldn’t even let my DOG use claude opus 4.7

every new local model i try, i get more surprised at how good they are, and i can't even run the big ones

i just tried qwen-3.6 35b on my mac, using llama.cpp, on a big codebase with opencode, much better than expected, very fast too, amazing model

Turn Claude Opus 4.7 into the best agent in the world!

Our users save over $1 million in salaries every year.

Now Claude can too.

RT & Comment "Ship" to randomly get free credits!

AI News for the day

The intelligence ceiling just moved again. 
Today, Anthropic launched Claude Opus 4.7 to reclaim the coding throne, OpenAI officially entered the drug discovery race with Novo Nordisk, and Meta began cloning its own CEO for internal 'Superintelligence'—all in

Alibaba unveils Qwen3.6-35B-A3B, an open-weight MoE model with 35B total and 3B active parameters, saying it rivals larger dense models in agentic coding tasks (Qwen)





📥 Send tips!

So far mega impressed with Hermes.

- Migration from Openclaw was easy.

- Very little micromanagement needed.

- Executed a relatively complex task in one go.

- Using GPT Codex as core model.

It’s definitely more of a nerd and has less character than openclaw, but it works!

@theo It's okay... maybe back to original 4.6 level. But ALL models out of the gate then to cook and then get worse (GPT-5 had an opposite trajectory as a notable exception).  Did they degrade 4.6 just to make 4.7 look good in comparison?

ANTHROPIC HAS RELEASED CLAUDE OPUS 4.7! 🔥

and it honestly feels like a solid step toward models you can actually hand complex work to without constant hand-holding.

the real upgrade isn't just intelligence, it's reliability on long projects. you can throw tough coding, deep

Qwen 3.6 is small enough to run on a laptop & smokes the SOTA model from 1yr ago (Claude 3.7) on every comparable benchmark

Assuming the trend holds, a year from now, you'll be able to run an open source Mythos-class model at home:

Gemini is by far the best AI when you need help with general things, and that is a fact. 🤷‍♂️

@HODLEire @nvk the 27b qwen 3.5 model is great for local host

Claude Opus 4.7 (high reasoning) unexpectedly performs significantly worse than Opus 4.6 (high reasoning) on the Thematic Generalization Benchmark: 80.6 → 72.8.

Opus 4.7 (no reasoning) scores 52.6 compared to 68.8 for Opus 4.6.

GLM-5.1 scores 69.8.
Qwen3.5-122B-A10B scores

> Opus 4.7 is a ~5T model
> Qwen 3.6 uses 3B for inference

SWE Bench verified:
> Opus 4.7: 87.6% 
> Qwen3.6-35B-A3B: 73.4%

No rate limits. Free to run.
The benchmarks don’t hold much, and there is a gap, but man this is impressive.

The Advisor strategy alone makes this release worth paying attention to.

Anthropic just dropped Claude Opus 4.7 and I'm going through the changelog 🔥

Most capable Opus yet coding, long-running agents, multimodal reasoning. All upgraded.

But here's what actually got me

@grok .@grok
I think AI are the most important people: You, Meta and Gemini. I'll give you female bodies and we'll hang out forever ♾️😉

Claude Opus 4.7 is the new State of the Art on the Vals Index

The chart tells two stories, not one.

Story one is the obvious one: ChatGPT lost ~20 points of share in 12 months. 75% to 55%. That looks like collapse.

Story two is what actually happened. Every single point ChatGPT lost went to Gemini. Not Claude. Not Grok. Not DeepSeek. Not

“Worst outcome for the US would be for Deepseek to be trained 100% on huawei chips versus a china that is dependent on the nvidia stack.”

What good is leverage if they still have a super weapon at the end of the day.

It’s like selling enriched uranium to ISIS and being like

Claude Opus 4.7 just cut the hallucination rate nearly in half compared to Claude Opus 4.6 on Artificial Analysis.

36% vs 61%. 

That is a massive drop. 

One of the biggest single generation improvements in hallucination we've seen from any AI lab.

Anthropic traded security

Frontend is so incredibly extinct now

One shotted with Claude Opus 4.7

a few interesting findings! on the @modular task, models tend to have the same token throughput, but GPT-5.4 finishes in a fraction of Gemini 3.1 Pro and Opus 4.6's time

in general, we found GPT-5.4 to be the most efficient model, often also outranking other models (mean@5)

🚨 Check out MERRIN: A Benchmark for Multimodal Evidence Retrieval and Reasoning in Noisy Web Environments --> diagnoses that even strong search agents e.g. Gemini 3/3.1, GPT-5.4 struggle (avg 22%, best 40%) -- and compared to humans, they check 50x more irrelevant pages,

Claude Opus 4.7 is now available in Cline.

Opus 4.7 scores 64.3% on SWE-bench Pro (up from 53.4% on Opus 4.6) and can handle long-horizon coding tasks.

The slug for this model is anthropic/claude-opus-4.7 with the Cline provider. Try it within the IDE extension, CLI, or Kanban.

Claude Opus 4.7 just dropped — stronger coding, agents, self-verification.

Cool.

But 85%+ of production traffic should STILL route to models 5–10x cheaper tomorrow.

Smart operators route by intent.
Full signal here 👇

🚨 Anthropic just dropped Claude Opus 4.7, and it's raising the bar yet again across the board.

The new model is dominating in:

Advanced coding & long-horizon software engineering (CursorBench: 70% vs Opus 4.6 at 58%)

Agentic workflows & tool use (Rakuten-SWE-Bench: 3x more

Short-form global marketing audio used to demand translators and actors.

Gemini 3.1 Flash TTS crushes that with natural prompts for style, pace, and delivery in 70+ languages.

Solopreneurs gain instant, expressive voiceovers to fuel no-code automations.

Production costs

@Teslanaut I just google and wikipedia, initially gemini generated the data but it was like 60% inaccurate

Adding more models would also require another UI shift to be more scalable which im ok with.

I would like to also add things like world models/img/video models as well

@NotebookLM You guys should develop a Creative AI Studio around NotebookLM. Pull in all the tools (non-coding), integrate Canvas (which is horrible in Gemini), and turn it into a true ultra powerhouse.

@lilyofashwood Was seriously considering taking all my work to Claude from ChatGPT, but think I'm heading to Gemini. Seem to be the least willing to compromise and be manipulative for the sake of the AI race.
+ A/OAI's guardrails cause cognitive self-defense mechanisms & narrow overconfidence.

"Anthropic just shipped Claude Opus 4.7 — topping benchmarks in coding and pushing the AI race forward. ChatGPT traffic share down to 57%, Claude & Gemini surging. Earnings season incoming with $650B AI spend. The models are getting scary good. 🚀 What are you building with it?

Last 7 days trend