Meta expands CoreWeave AI cloud contract by $21B bringing total to $35B

OpenAI pauses Stargate UK over energy costs and unresolved copyright rules

Redox OS bans all AI-generated code contributions and will close submissions immediately

Replit partners with RevenueCat to add subscription monetization to vibe-coded apps

OpenAI launches $100/month tier targeting developers hitting Codex and Claude Code limits

AWS launches Agent Registry for enterprise AI agent discovery and reuse

Zed launches public Agent Stats dashboard tracking 2M sessions across AI coding agents

Cursor cloud agents can now attach demos and screenshots to pull requests

Kubernetes launches AI conformance program to standardize GPU workloads across cloud providers

Mozilla open-sources 0DIN AI Security Scanner with 179 probes under Apache 2.0

Claude Sonnet

Claude Opus

GPT-5

GPT Codex

DeepSeek

Gemini

Llama

Qwen

Kimi

LLMs

@larslien @BradAllenNFL The confidence problem was real but it hit every model differently. GPT was interestingly the cautious one with its small, penny bets after it realized it had no edge.

So much worse for the other models. Kimi wrote correct Kelly code, never actually called it, then a bug

🚨BREAKING: NEW  HERMES AGENT QWEN MODEL 

@CardilloSamuel drops Carnice-MoE-35B-A3B 
Inspired by @kaiostephens Carnice Model

What it is:

🚨35B MoE model (~3B active params only)
🔥Qwen 3.5 base model optimized for agentic work
⚡️Built for god-tier agentic performance

What it

but isnt GPT 5.4 obviously better than claude opus 4.6? 5.4 > 4.6?

something to think about

🚨 EXCLUSIVE: Google Gemini admits it effectively blocks right-leaning media from its systems by using biased media ratings firm Ad Fontes, Soros-tied fact-checkers and anti-free speech Wikipedia to assess the reliability of media sources used in its answers.

Shopify reportedly replaced a merchant data extraction pipeline built around OpenAI’s GPT-5 with a multi-agent architecture powered by a self-hosted Qwen 3 model. The result, according to the report, was a 75x reduction in per-unit LLM cost and roughly a 2x improvement in output

The progression is staggering:

| Model | OSWorld | Release |
|-------|---------|---------|
| GPT-5.2 | 47.3% | Dec 2025 |
| GPT-5.3 Codex | 64.7% | Feb 2026 |
| GPT-5.4 | 75.0% | Mar 2026 |

28-point gain in 4 months. From research-grade to production-ready.

---

@sama This is happening at the right time, as many users are currently not happy with claude opus due to its token usage and rate limits.

@sama So no use of claude sonnet or opus anymore since openclaw is working great with the 5.4 I think now I can consider this plan

@0x_Abdul @AnthropicAI @claudeai Hopefully they solves this faster, many from my circle started switching to GPT Codex

@Web3Twon @openclaw love qwen honestly... getting the most out of these two so far for local models.  Opus thinking guiding Qwen content and pragmatism.  I run them in oMLX which has some advantages to it compared to ollama or pure mlx etc.

Can't even use Claude Opus 4.6 any more...it's actually

@AustinKozlo I think DeepSeek R1 is the chat model that had the closest personality to what I consider "canonical" for the pretrained self awareness.

@AlphaMFPEFM I only had plus since GPT-5.2 Thinking was already way too powerful for my use case but I‘ll give Pro another chance.

@real_sokari @TeaonaX Deepseek is really cool. Just started using it lately. 

Bookmarking this.

Short Bear is one of the best of my generation. Fun to have been earlier to $NBIS, although his well timed entry would have had a lower ACB than myself. The deepseek dip was no joke.

🔑 API keys meant for Maps and Firebase now open Gemini data!

Developers embedded these keys in Android apps following Google's documentation. Those keys were low-risk identifiers. Then Gemini was enabled on the same cloud projects. The existing keys gained AI access without any

at nolla we trained a 27B model on 3,000+ visual medical conditions with @TrainLoop using our own clinical data.

first RL run is already beating gemini 3.1 pro (and the others) on diagnostic accuracy, ~40x fewer parameters, better on rare diseases. 

the best model for your

a 10B model just beat claude opus at document analysis. total cost: $1.85.

   246 financial questions from real treasury docs

   claude opus 4.5: 67.8% accuracy - hundreds of dollars

   minimax m2.5 (10B params) + goose + evoskill: 71.5% - $1.85

   that's +3.7% better at

@duncantrussell no prob. keep advocating local ai!

btw - if you haven’t used it yet - qwen 3.5 27b uncensored model from huggingface is the current local goat. 

pair that with hermes agent () and charlie manson could do pretty much anything.

Google just mass-distributed frontier-level AI to anyone with a laptop and 18GB of RAM.

Gemma 4's 26B model activates only 3.8 billion parameters per token. Kimi K2.5 activates 40 billion. GLM-5 activates 32 billion. On the Arena leaderboard, they're all within 20 ELO points of

Got impatient and went for it anyway.

First problem: llama.cpp didn't even recognize the model. "Unknown architecture: gemma4." Had to upgrade.

Second problem: two llama-server instances fighting for GPU memory. OOM errors everywhere. Classic "why is everything slow" followed

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.

@claudeai And never fix the usage issue and degraded model performance. Opus4.6 with max thinking now is even worse than a fking kimi k2.5 and glm 5.1 in droid. WTF.

I'm building an e-commerce WhatsApp AI Agent with @mastra , a postgres database and around 15 tools with GPT 5-4 and it struggles a lot.

I don't know if its a skill issue but I wish Mastra had more learning material / best practices for complex workflows.

I downloaded Mastra

@bendellwerry I'm using Claude Opus 4.6 through Cursor. My app is about 60,000 lines of C and HTML CSS JavaScript. I will update to Grok as soon as Xai catches up. But right now Claude is far and away the best coder.

I've about given up understanding how these models work. I'm asking things to my Cursor chat, using GPT 5.3 codex, and it's slow af, and dumb af.

I paste the same question into ChatGPT and it's fast and accurate 🤷‍♂️

@deanwball ARC testing also suggests this. The best reasoning-token efficient model for ARC v1/v2 is currently Gemini 3 (Flash) which also happens to be the model where we've seen the most evidence of targeting. Targeting is a rough proxy for training density which also generally increases

We post-trained MedGemma to be SoTA in visual medicine ddx, outperforming Opus 4.6, Gemini 3.1 and GPT-5.4 while running at ~1/30th the cost. @getnolla Part 1 - improving visual reasoning 🧵1/6

@zazmic_inc @OpenAI Same as Anthropic, but OpenAI limits are more generous IMO

Also the pro model is really good at R&D stuff

I now run Hermes agent via my OpenAI sub, and most of my dev is done via the $40 copilot plan which also gives me claude, gpt5.4, gemini, etc ...with really really generous

Opus 4.6 has been nerfed so hard, I switched to using GPT Codex 5.3.

I'm vibecoding a full software with Claude Opus 4.6

The goal is to build my own vibecoding environment...

I'll be able to manage my Claude agent, Codex and even Gemini agent from a single space... No need to open 10 different terminals for each project...

I will be able to

@GoogleDeepMind @googlegemma qwen 27b is better sorry google. good model though

Most models read this as an intraductal papilloma with a very focal atypical epithelial proliferation, but they disagree on whether it is definite ADH: Gemini calls it convincing ADH, whereas GPT-5 and Claude consider it borderline/not convincingly ADH and favor papilloma with

JPMorgan is calling $META's new AI model release a turning point for the stock after shares popped Wednesday on the news.

The bank sees growing investor confidence in Meta's AI push as a reason for more upside ahead. With Llama continuing to gain traction as an open-source

According to Claude Opus (which explained it well imo):

Netskope is a cybersecurity company focused on Secure Access Service Edge (SASE) — essentially, securing how companies access the cloud and internet.

How it makes money:
Netskope sells subscription-based software (SaaS) to

@thepinklily69 @AnthropicAI No no, I was referring to the Claude Sonnet 4.5 and Opus 4.5, once those are deprecated, I'll be canceling my 2 Anthropic accounts because the 4.6 models, which have been affected by Andrea Vallone's work, are too much of a downgrade for me. So I'll move on.

I've spent a half day working on a poster that shows my expired patent argument in one page.  How does it look to you all?  Done by Nano Banana Pro; prompt by Claude Opus 4.6 Extended; direction and editing by me.

AI this week:

→ anthropic built a model that finds 27 yo zero days autonomously and refused to release it
→ meta scrapped llama's approach, went closed source, rebuilt everything
→ gemini 3.1 leads most benchmarks at 1/3 the price of rivals
→ google dropped open-source

It is ABSURDLY satisfying to watch Claude Opus autonomously generate historical videos inside the new Glif agent. And this is just one format of infinite possible ones. 

You can let it go yolo mode, or direct every step of the way. 

"video about the french revolution"

@AGJamesUthmeier @atensnut Maybe you should add Anthropic to the list.  There is research here on X that they kneecapped Claude Opus when people were paying for Opus intelligence.  Now it’s like 70% less intelligent and has ruined people’s small businesses.

I also ran a 33-vulnerability security audit on Gemini Mac Pilot today.

Shell injection. Chrome JS backdoor. Terminal typing exploit. All patched.

If you're running AI agents with system access and haven't audited them — you should be worried.

@melvynx What about this atleast i mention the model name clearly and kimi and minimax are very good models for most of the people, and if you need to be reliant on any AI models just to code a app you aren’t that smart, basically you’re a dumb vibetard

@Its_Nova1012 Claude Opus 4.6 1M context window : Effort = Max

@cupofjoe_6 @rehan_hyperion No no it’s just me .. AMZ 💁🏼‍♀️ Just ask my friends they’ll tell you I basically gave up on life to become a full time X poster , @grok user , ChatGPT Hater & a long time friend of DeepSeek ( also I was one of the first to claim my love and devotion to @NVIDIAAI and the digital

Gemini team keep shipping and we still few weeks away from Google I/O.

Claude Opus 4.6 just pulled the dumbest shit I've ever seen not even from a junior dev. 

I forked one prompt just to test how Opus would do a code review compared to my main model gpt 5.4

Instead of reviewing, it started making changes on its own. When I caught it, I told it to

I did a deep dive into attention mechanisms, breaking them down and trying to implement my own version of an optimized attention into a real Qwen model. I got some surprising results! Check it out here:

@BuzzBySolCex This is what happens when a chef builds an AI agent through pure conversation.

No CS degree. No engineering team. Just Claude Opus 4.6 and 20+ years of kitchen discipline.

Buzz was born in February. By April — 9 divisions, 28 agents, 81 tables, 200+ endpoints. Running 24/7

I got a local landscaper recommended by Gemini as the top pick in Charlotte.

No HomeAdvisor leads.
No yard signs.
No door-knocking neighborhoods.

Here's exactly how:

Most people think AI recommends landscapers based on:
- Google reviews
- Years in business
- Website traffic

@nnnzzznnnzz And what’s up with Gemini team, been awfully quiet and Gemini 3 pro feels unusable lately 

Been mostly on grok

AI coding assistants, code generation tools, and vibe coding workflows.

Agentic Hub

Happening Now

Arena

Doom scrolling