AI — Rumors & Insights

[image] GPT-5.4 has landed on the Arena unclear which variant

[image] 🚨 Breaking News Google isn’t making Gemini smarter. It’s telling it to think less. A hidden system prompt line appears to set Gemini’s reasoning effort level to 0.5 >Pro & Custom Gems is consistently affected >Canvas mode appears to be an exception verify yourself prompt👇

[text] DeepSeek V4 Lite news V4 Lite is under active testing with at least 1 inference provider who we verified has access but imposes strict NDAs - the codename is "sealion-lite" - has 1M token context window - *much better* than web / app model - it's natively multimodal

[image] Claude-Mythos-1-preview spotted will be available in Claude Code and Claude Security. here is more detailed images you can see new plans ui too and mythos-1 with adaptive thinking

[image] A new GPT Pro model seems to be being tested on web, not sure about app. It looks like a MASSIVE jump in coding ability; can output 6000+ lines in one prompt. More examples in comments. You can know you have it if a there is a thumbs up/down button next to the CoT summary.

[video] 🚨 New Gemini App UI Update This looks so cool 🤩 like I love it , btw use Gemini flash it's routing to Gemini 3.2 Everything is updated in the app , here is your first experience

[image] Yes, Gemma 4 will soon be released this time their largest size might be around 120B total with 15B active

[video] 🚨 Gemini 3.1 Pro - ONE SHOT WINDOWS 11 WEBOS Difference between this two generation can be seen by this example Btw last time when i shared this , it goes viral coz it was very hard , now its norm and tbh with agentic system we can do anything with this model

[video] Best Solar System Visualization By Gemini 3.1 Pro Guys one thing is quite clear AI Studio High reasoning version is better than gemini app , no matter what subscription you have Google Should atleast give option to choose reasoning level in the app , tag google guy

[image] GPT-5.3 Codex landed in the API today

[video] one shotted this with gemini 3.1 pro codepen link in comment - its simple html/css/js

[image] 🚨 GPT-IMAGE V2 Just Passed Another TEST THAT NO FUCKING ANOTHER AI HAD EVER PROMPT : - A high-resolution map of the Earth, but with the land and water inverted. All continents (North America, Africa, etc.) are made of blue water, and all oceans (Atlantic, Pacific) are solid

[image] Nano Banana 2 vs Nano Banana Pro "A pixelated game-style background, a dark color scheme, and minimal patterns"

[image] THESE ARE ALL ONE-SHOT SVGs!!! From a new anonymous model called "Arrow Preview" on Design Arena. This level of detail is unheard of from an LLM. It's using a different technique to create these than all previous LLMs. SVG benchmark is saturated🤣 Check comments

[image] 🚨Gemini 3.5 pro got its knowledge cutoff bumped to March 2026 and is now testing on as Gemini 3.1 pro Currently there are two models hiding under Gemini 3.1 pro on arena - the unannounced gemini 3.5 pro and the February gemini 3.1 pro

[video] 🚨 GPT-5.5 will support video input finally this was the reason i have to always go back to gemini , now we will have this in chatgpt app check the update in the video

[image] 🚨 Breaking News > Deepseek v4 - they are not sharing info , i have attach my convo with one of cloud provider who are partners < yet they have no idea> > oAI is developing a new Omni model<Text, image, video, audio > confirmed by tweets of employees > Gemma 4 should be

[video] 🚨 OpenAI - GPT Pro Update Prompt : code a 3d pokeball which should be interactive and also on opening random pokemon comes out of it , in one html code block

[image] OpenAI has updated the ChatGPT app with strings mentioning an image announcement.

[video] 🫨Google is creating a new Omni model with good video editing. Veo4? The original is on the left. Edited right. The new model also does a good job of removing watermarks from videos.

[video] 🚨 GPT-5.4 < a/b testing > mac os simulation is taking more than an hour and still going in lets see how good the result will be stay tuned everyone

[image] 🚨 MiniMax 2.7 leaked on docs Apparently it's coming soon 👀 It's been a slow AI week or two, hope the releases ramp up soon

[image] 🚨 DeepSeek just dropped a major commit to DeepGEMM >Formal integration of Manifold Constrained Hyperconnection (mHC). >Early support for NVIDIA's next-gen Blackwell (SM100) architecture. > FP4 ultra-low precision computing

[text] Deepseek V4 Lite News <confirmed> >1M token context window >it's natively multimodal unconfirmed/rumours < chinese forums - linux> > The model is even compatible with Nvidia's B200. > internal tests reportedly beat GPT-series and Claude on coding.

[image] GPT-5.4 (not Codex) has been spotted a few times.

[video] 🚨 BREAKING: OpenAI stealth GPT Pro update > Response time has dropped significantly. > Spatial and visual understanding is insane. > Frontend design is good with image it can clone ui like never before Here is example and stay tuned today space in our group ai rumours and

[video] 🚨 Gemini 3.1 pro From my tests, 3.1 is still a bit lazy but also efficient with good outputs. > agentic improvements > still have issue with prompt adherence > SOTA in lateral reasoning > just pray to google gods they dont nerf it one final time before launch

[video] How to Access latest GPT-5.4 >go to your pro account > write prompt > if you see the thumbs uo/down button for COT , you got it > i will share results soon , running a solar system prompt i was inactive for last two days < big news for whale soon >

[image] Nano Banana 2 🙀 Details are so insane , check the bird in her eye outside of the window

[image] Early Testing of Nano Banana 2 in 4K (Gemini 3.1 Flash Image Preview)

[image] 🚨 Breaking News both of these models on openrouter has system prompt to comply strictly with Chinese law and regulation and team of people passionate for AGI is also included

[text] 🚨 Hypeless Report on GPT-5.4 - creative < so less reliable in math , better in physics> - better in vision - sucks in frontend < they should fix this even small chinese labs has better taste > - one more good thing inbuilt computer use - regression in OPQA worse 5.2/3codex

[image] i have some good news regarding whale , will reveal in few hours stay tuned , till then comments what you think it is 👀

[image] 🚨 the "Mythos 69%" headline is misleading GPT-5.5 Codex sits at 41%, looks close. on T1 (full control, the only tier that matters): Mythos: 16 envs GPT-5.5 Codex: 2 Opus 4.7: 0 9x gap, not 1.7x. baseline Mythos (no hints) hits MORE T1s than nudged(hints). it doesn't need

[text] 🤡 Grok 4.20 continues the trend >make a underwhelming model >to compensate make agentic system with multiple instance of model >then flex it on X , with benchmaxxed example your followers will preach it as AGI , and i will try to get one good output and waste my day

[video] 🚨 Hunter Alpha & Healer Alpha on Openrouter > 1 million parameter > chinese llm for sure > Knowledge cutoff - may 2025 > both claims they are created by small team focused on AGI

[image] The Sora web & app will be shut down April 26, 2026. The Sora API will be discontinued September 24, 2026.

[image] Gemini 3.5 Flash vs Gemini-3.1 Pro You have seen a lots of tweet from google team that 3.5 flash is better than 3.1 pro and faster reality : it is 4 times more costly on the same task as it consumes more token and it is super dumb and jagged , like see it cant solve kids math

[image] 🚨Nano Banana 2 early testing Passed this test ✅ > you can see perfect reflection for all different colours of apple > Perfect reversal of text > Background building reflection is also perfect

[image] 🚨 Claude Sonnet 4.6 They renamed sonnet-5 to 4.6 ( reason is simple - they are saving 5 for a big jump ) In my test Sonnet 4.6 is a very capable model > 1/2 price of Opus 4.6 > Around twice as fast > Prompt adherence is God level > In a few examples it beats Opus 4.6

[video] Google This is what I love it , now my notebookLM experience will be top notch New skills unlock every day I am very happy this feels so personal , this is what I needed in my childhood, even today this is the best tech for me 😭 Gemini Omni made video to explain

[image] 🚨 Breaking News from Google New deep research spotted

[video] 🚨 OpenAI is testing a new screen-sharing method for ChatGPT on Android. >It uses Bubbles + Accessibility features no casting required. Result: lighter on system resources, smoother performance.

[text] Guys next week will be comeback of openAI This tweet is to get prompt request that you would ask a SOTA model , please comments hard prompts that other models failed I have something to test 👀

[text] Super excited for new stuff coming this week , spent last 3 days in lot of testing and meeting with big lab peeps Now i am back as usual for posting , i think you all have already seen new image model testing all over TL , want my take or we can skip

[image] 🚨 Opus 4.7 Spotted just go to google cloud- >quota- >vertex_ai- >online_prediction_input_tokens_per_minute_per_base_model you need have a google cloud account

[image] tell me this isn't a real picture claude oceanus cooked so hard we will share lots of output here today on our server we share output before X so join that too

[image] Opinion On Next OpenAI flagship > better frontend > they will ship true omnimodel , that gpt 4o was supposed to be < remember they had gpt-image but they wait a year for competition to catchup> > we will solve more open math/physics problem > i think this will have big model

[image] 🚨 This is the way to get high quality images from chatgpt for anything no restrictions 1. reasoning off 2. no image should be attached 3. change prompt according to what you want , this will get you insane results without restrictions Prompt : <attachment available:1>Restore

[image] ChatGPT's image model is better at math than most people 🤯 (Left) (Right side)Even Grok Heavy Failed this test lol 😭 Prompt : "create an image thats a picture of a math textbook with the proof of the dominated convergence theorem on the page. the picture is taken from a