Competition is great.
It looks like OpenAI and Anthropic are competing to see who can suck up to users harder, just to keep them from jumping off their train and hopping onto the competitor’s.
OpenAI went ahead and dropped API prices for SOL 5.6 by 20% for a whole THREE months.
Anthropic was like: fine, we’re extending the increased weekly Claude Code limits through August 31.
Users: message received, keep it coming. Ox Alpha is completely free anyway, so we’ll be waiting for even more concessions.
It looks like OpenAI and Anthropic are competing to see who can suck up to users harder, just to keep them from jumping off their train and hopping onto the competitor’s.
OpenAI went ahead and dropped API prices for SOL 5.6 by 20% for a whole THREE months.
Anthropic was like: fine, we’re extending the increased weekly Claude Code limits through August 31.
Users: message received, keep it coming. Ox Alpha is completely free anyway, so we’ll be waiting for even more concessions.
🎉286🔥280👍265❤243
MiniMax H3 Max: 36x faster — and the weights are going open
Regular MiniMax H3 in Minimax Design cloud vs H3 Max on fal.
Same settings: 15 sec, 720p.
H3 took around 6 minutes. H3 Max took around 10 seconds.
That’s roughly 36x faster, with comparable quality to my eye.
Important caveat: this is not a clean model benchmark. These are different production stacks. Still, 36x is almost certainly not just “more GPUs”.
Big news:
fal confirmed H3 Max weights will be released.
Asked whether Ref2V will also be open:
“yes, working on it.”
So we may get an open H3 Max checkpoint, and later Ref2V too.
The key question: how much of that speed is in the model, and how much in fal’s stack?
I wouldn’t expect H3 Max to run 36x faster locally.
The gain likely comes from fewer sampling steps, optimized attention, fused CUDA/Triton kernels, mixed precision, parallelism, VAE and memory optimizations.
Part of the speed is in the weights, part is in fal’s infrastructure.
That ratio will matter most after release.
Regular MiniMax H3 in Minimax Design cloud vs H3 Max on fal.
Same settings: 15 sec, 720p.
H3 took around 6 minutes. H3 Max took around 10 seconds.
That’s roughly 36x faster, with comparable quality to my eye.
Important caveat: this is not a clean model benchmark. These are different production stacks. Still, 36x is almost certainly not just “more GPUs”.
Big news:
fal confirmed H3 Max weights will be released.
Asked whether Ref2V will also be open:
“yes, working on it.”
So we may get an open H3 Max checkpoint, and later Ref2V too.
The key question: how much of that speed is in the model, and how much in fal’s stack?
I wouldn’t expect H3 Max to run 36x faster locally.
The gain likely comes from fewer sampling steps, optimized attention, fused CUDA/Triton kernels, mixed precision, parallelism, VAE and memory optimizations.
Part of the speed is in the weights, part is in fal’s infrastructure.
That ratio will matter most after release.
🎉361👍359❤350🔥319
THIS is WILD.
Not another candle, not another skincare bottle. I am building a full launch kit for a futuristic smart ring inside Lovart with Seedance 2.5.
Lovart is not just a video generator. It is an AI design agent, so I start with one brief: matte graphite smart ring, holographic charging case, neon blue interface, luxury tech launch, 30-second hero video, posters, and social ads.
Try Lovart here
Watch full video https://www.tiktok.com/@web3.world.yt/video/7680504360232701205
Not another candle, not another skincare bottle. I am building a full launch kit for a futuristic smart ring inside Lovart with Seedance 2.5.
Lovart is not just a video generator. It is an AI design agent, so I start with one brief: matte graphite smart ring, holographic charging case, neon blue interface, luxury tech launch, 30-second hero video, posters, and social ads.
Try Lovart here
Watch full video https://www.tiktok.com/@web3.world.yt/video/7680504360232701205
❤125🔥120👍114🎉110
🚨 OpenAI introduced GPT-6 Astra — and this feels like more than just another model upgrade.
Astra is built not only to answer questions, but to actually do work: coding, browsing, research, documents, and complex multi-step tasks.
Some standout numbers: 99.9% on ARC-AGI-3, 97.6% on FrontierMath Tier 4, 100% on ExploitBench, a 1.05M-token context window, and up to 128K output tokens.
It’s also OpenAI’s first model to reach Critical cyber capability, meaning it can potentially discover and exploit previously unknown vulnerabilities with the right tools.
The bigger story is the shift itself: AI is moving from “smart assistant” to something you can increasingly delegate entire workflows to.
Astra is built not only to answer questions, but to actually do work: coding, browsing, research, documents, and complex multi-step tasks.
Some standout numbers: 99.9% on ARC-AGI-3, 97.6% on FrontierMath Tier 4, 100% on ExploitBench, a 1.05M-token context window, and up to 128K output tokens.
It’s also OpenAI’s first model to reach Critical cyber capability, meaning it can potentially discover and exploit previously unknown vulnerabilities with the right tools.
The bigger story is the shift itself: AI is moving from “smart assistant” to something you can increasingly delegate entire workflows to.
🔥212❤202👍198🎉196
GPT-6 Astra: messy rollout, resets, and Plus segregation
Three updates: one okay, one very good, one very bad.
OpenAI has revealed the limits. $200 Pro gets 200 Astra messages/week. The new $100 Pro gets 50/week, shared with GPT-5.6 Sol Pro. Business Standard gets 15/month, Premium 50/week. So Astra isn’t unlimited even on $200 Pro.
Good news: OpenAI says it will compensate rollout delays. For every day Astra is unavailable on a paid account, users get one banked reset to use later
Bad news for Plus: OpenAI’s latest help article says, “It is not included with ChatGPT Plus in Chat.” Astra appears in regular Chat as GPT-6 Pro only on Pro $100, Pro $200, Business, and Enterprise
Meanwhile, the launch page still says Astra is coming to Plus, Pro, Business, and Enterprise. Where Plus gets access is unclear. Maybe Work or Codex?? Current docs confirm Astra there only for Pro
So: Astra limits are known, but Plus users waiting for it in regular ChatGPT appear out of luck for now. Codex access for Plus is unclear too
Three updates: one okay, one very good, one very bad.
OpenAI has revealed the limits. $200 Pro gets 200 Astra messages/week. The new $100 Pro gets 50/week, shared with GPT-5.6 Sol Pro. Business Standard gets 15/month, Premium 50/week. So Astra isn’t unlimited even on $200 Pro.
Good news: OpenAI says it will compensate rollout delays. For every day Astra is unavailable on a paid account, users get one banked reset to use later
Bad news for Plus: OpenAI’s latest help article says, “It is not included with ChatGPT Plus in Chat.” Astra appears in regular Chat as GPT-6 Pro only on Pro $100, Pro $200, Business, and Enterprise
Meanwhile, the launch page still says Astra is coming to Plus, Pro, Business, and Enterprise. Where Plus gets access is unclear. Maybe Work or Codex?? Current docs confirm Astra there only for Pro
So: Astra limits are known, but Plus users waiting for it in regular ChatGPT appear out of luck for now. Codex access for Plus is unclear too
❤362🎉347👍343🔥332
This media is not supported in your browser
VIEW IN TELEGRAM
So, how are the modelers doing with the new horsepower?
Or rather, with the new Astra.
Blender MCP + Astra + a bit of Computer Use: feed it a sketch of a steam locomotive, and a few minutes later you get 3,295 objects.
https://x.com/tomkrcha/status/2095756085890310311
I think that’s it… we’re done.
Or rather, with the new Astra.
Blender MCP + Astra + a bit of Computer Use: feed it a sketch of a steam locomotive, and a few minutes later you get 3,295 objects.
https://x.com/tomkrcha/status/2095756085890310311
I think that’s it… we’re done.
❤127👍125🎉117🔥110
This media is not supported in your browser
VIEW IN TELEGRAM
Viggle is back with an open model — Viggle-Animate — and it’s pretty damn capable.
For the first time, they’re releasing weights + code.
It replaces characters in video using MiniMax H3. Take a video, grab one frame, manually redraw the character as anything you want, and the model propagates that replacement through the entire clip while preserving motion, camera, and timing.
No pose skeleton, segmentation, face tracking, depth, or text prompt. Just the video + one modified frame.
Under the hood: a 33.1B MiniMax H3 ref2va fine-tune + DMD2 LoRA. Distillation brings it down to 4 sampling steps.
It’s 6× faster than Wan2.2-Animate-14B.
The catch: hardware. BF16 needs ~62 GiB for the transformer and peaks at ~80.1 GiB VRAM. They recommend 96 GB+ or CPU offload. It reportedly runs on a 32 GB RTX 5090 via NVFP4, but that quant isn’t public yet.
Also, you can’t use a normal character portrait. The reference has to be a frame from the original video, already redrawn as the target character.
https://viggle.ai/h3
For the first time, they’re releasing weights + code.
It replaces characters in video using MiniMax H3. Take a video, grab one frame, manually redraw the character as anything you want, and the model propagates that replacement through the entire clip while preserving motion, camera, and timing.
No pose skeleton, segmentation, face tracking, depth, or text prompt. Just the video + one modified frame.
Under the hood: a 33.1B MiniMax H3 ref2va fine-tune + DMD2 LoRA. Distillation brings it down to 4 sampling steps.
It’s 6× faster than Wan2.2-Animate-14B.
The catch: hardware. BF16 needs ~62 GiB for the transformer and peaks at ~80.1 GiB VRAM. They recommend 96 GB+ or CPU offload. It reportedly runs on a 32 GB RTX 5090 via NVFP4, but that quant isn’t public yet.
Also, you can’t use a normal character portrait. The reference has to be a frame from the original video, already redrawn as the target character.
https://viggle.ai/h3
❤57👍57🔥50🎉46
This media is not supported in your browser
VIEW IN TELEGRAM
FLUX.3 Video Edit model
It can take an existing video and, based on a prompt, change objects, characters, backgrounds, materials, text, effects, and even events within the scene. You can also change dialogue and translate conversations with lip sync.
The biggest bombshell: $0.03 per second. Dirt cheap!
So:
5 sec — $0.15
10 sec — $0.30
15 sec — $0.45
But.
It currently works with clips up to 15 seconds, and input videos above 720p are downscaled before processing. No masks, additional references, or resolution selection yet.
In short:
max. 15 sec
max. 50 MiB
up to 720p in the edit pipeline
24 fps
no seed control
apparently no output resolution or duration control (need to dig into this)
no masks
no additional reference images
no video-only motion/style reference
not an extension model — continuing a clip requires FLUX 3 Video Continuation
You can run the result through FLUX Video Upscale separately, up to 4K — but that’s extra $$$.
It can take an existing video and, based on a prompt, change objects, characters, backgrounds, materials, text, effects, and even events within the scene. You can also change dialogue and translate conversations with lip sync.
The biggest bombshell: $0.03 per second. Dirt cheap!
So:
5 sec — $0.15
10 sec — $0.30
15 sec — $0.45
But.
It currently works with clips up to 15 seconds, and input videos above 720p are downscaled before processing. No masks, additional references, or resolution selection yet.
In short:
max. 15 sec
max. 50 MiB
up to 720p in the edit pipeline
24 fps
no seed control
apparently no output resolution or duration control (need to dig into this)
no masks
no additional reference images
no video-only motion/style reference
not an extension model — continuing a clip requires FLUX 3 Video Continuation
You can run the result through FLUX Video Upscale separately, up to 4K — but that’s extra $$$.
❤167🔥166🎉166👍138
Media is too big
VIEW IN TELEGRAM
Google just dropped Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking.
Gemini 3.8 Live is fast voice-to-voice, understands video/images almost in real time, and automatically switches between 97 languages.
(Siri still has ONE.)
It can call tools and APIs in the background without interrupting the conversation.
Extended Thinking is the version for complex tasks: it can reason and speak simultaneously, handle multi-step processes, and verbally comment on its progress.
There’s a Gemini API and AI Studio.
Extended Thinking is also coming to Gemini Live, Docs, Gmail, and Keep.
API pricing is the same for 3.8 Live and Extended Thinking:
Input audio — $0.005/min
Output audio — $0.018/min
Video/image input — around $0.002/min
Text — $0.75 per 1M input tokens and $4.50 per 1M output tokens, including thinking tokens.
Google wants to turn Gemini Live from a “voice chat” into a full-fledged realtime agent that can see, think, talk, and do things in parallel through tools.
Gemini 3.8 Live is fast voice-to-voice, understands video/images almost in real time, and automatically switches between 97 languages.
(Siri still has ONE.)
It can call tools and APIs in the background without interrupting the conversation.
Extended Thinking is the version for complex tasks: it can reason and speak simultaneously, handle multi-step processes, and verbally comment on its progress.
There’s a Gemini API and AI Studio.
Extended Thinking is also coming to Gemini Live, Docs, Gmail, and Keep.
API pricing is the same for 3.8 Live and Extended Thinking:
Input audio — $0.005/min
Output audio — $0.018/min
Video/image input — around $0.002/min
Text — $0.75 per 1M input tokens and $4.50 per 1M output tokens, including thinking tokens.
Google wants to turn Gemini Live from a “voice chat” into a full-fledged realtime agent that can see, think, talk, and do things in parallel through tools.
❤90👍87🎉86🔥71
This media is not supported in your browser
VIEW IN TELEGRAM
Buttery-smooth 120 FPS Three.js arena action, running entirely in your browser. Meet Arena Shooter.
4.6 BILLION Astra tokens burned.
20+ vehicles to capture, zombies, day/night combat, 20+ upgrades & perks, and LLM bots trying to destroy you.
No signup. No install. Just open and fight.
Try https://vortex.channel/vr/soloplay/
4.6 BILLION Astra tokens burned.
20+ vehicles to capture, zombies, day/night combat, 20+ upgrades & perks, and LLM bots trying to destroy you.
No signup. No install. Just open and fight.
Try https://vortex.channel/vr/soloplay/
🎉197❤195👍190🔥186
🚨 Qwen-Image 2.1
Alibaba just dropped Qwen-Image 2.1 — a unified text-to-image + image editing model with only 7B parameters.
🔥 Native 2K
🔥 Up to 10 reference images
🔥 Better identity & consistency
🔥 Masks/scribbles for local editing
🔥 Native RGBA + transparency
🔥 ComfyUI support from day one
The really interesting part: RGBA editing and object extraction from regular RGB images. Need to test this one.
💾 VRAM-wise, 12 GB looks like a realistic target with quantization/offloading. 16 GB should be very comfortable.
The previous Qwen Image was ~20B. This one is just 7B.
🔗 https://modelscope.ai/models/Qwen/Qwen-Image-2.1/summary
Alibaba just dropped Qwen-Image 2.1 — a unified text-to-image + image editing model with only 7B parameters.
🔥 Native 2K
🔥 Up to 10 reference images
🔥 Better identity & consistency
🔥 Masks/scribbles for local editing
🔥 Native RGBA + transparency
🔥 ComfyUI support from day one
The really interesting part: RGBA editing and object extraction from regular RGB images. Need to test this one.
💾 VRAM-wise, 12 GB looks like a realistic target with quantization/offloading. 16 GB should be very comfortable.
The previous Qwen Image was ~20B. This one is just 7B.
🔗 https://modelscope.ai/models/Qwen/Qwen-Image-2.1/summary
🔥569👍565❤521🎉516
Opus 5.5 vs GPT-6 Sol & Luna — same day!
Anthropic dropped Opus 5.5, while OpenAI released GPT-6 Sol and Luna.
Opus 5.5: $4 / $20, AA Index 58 (#1), Terminal-Bench 66.4%, GDPval-AA 1846 Elo. Anthropic also promises less “Claude-style” fluff.
Sol: $2 / $10, 1.05M context, AA 48. DeepSWE: 68.8% at $2.74/task.
Luna: $0.10 / $0.50, AA 37. DeepSWE: 66.6% at just $0.22/task — almost the same score as Sol for 12x less.
The catch with Opus: tokens are 20% cheaper, but usage is much higher — 119K tokens/task vs 73K for Opus 5 and 31K for Sol. CodeRabbit saw 40–60% more tokens on code reviews.
On AA, Opus 5.5 medium vs Sol max: coding 52.5% vs 43.9%, documents +159 Elo for Opus, business processes roughly tied. Cost: $1.34 vs $1.06/task.
Bottom line: Opus 5.5 is currently the smartest model and 2.5x cheaper than Fable. Luna is the big story for high-volume classification and extraction: 10x cheaper than Haiku. Sol is the mid-range option for mass workloads. Sonnet 5.5 reportedly arrives in a couple of weeks.
Anthropic dropped Opus 5.5, while OpenAI released GPT-6 Sol and Luna.
Opus 5.5: $4 / $20, AA Index 58 (#1), Terminal-Bench 66.4%, GDPval-AA 1846 Elo. Anthropic also promises less “Claude-style” fluff.
Sol: $2 / $10, 1.05M context, AA 48. DeepSWE: 68.8% at $2.74/task.
Luna: $0.10 / $0.50, AA 37. DeepSWE: 66.6% at just $0.22/task — almost the same score as Sol for 12x less.
The catch with Opus: tokens are 20% cheaper, but usage is much higher — 119K tokens/task vs 73K for Opus 5 and 31K for Sol. CodeRabbit saw 40–60% more tokens on code reviews.
On AA, Opus 5.5 medium vs Sol max: coding 52.5% vs 43.9%, documents +159 Elo for Opus, business processes roughly tied. Cost: $1.34 vs $1.06/task.
Bottom line: Opus 5.5 is currently the smartest model and 2.5x cheaper than Fable. Luna is the big story for high-volume classification and extraction: 10x cheaper than Haiku. Sol is the mid-range option for mass workloads. Sonnet 5.5 reportedly arrives in a couple of weeks.
🔥189👍161❤160🎉150