The best AI workspace for teams. Build your billion-dollar company at https://xitter.cf/t.co/SfNrJSggUH
SF
Joined June 2026
- Tweets97
- Following6
- Followers1.4K
- Likes145
$10 per million output tokens makes Kimi K3.1 Pro an expensive place to duplicate your teammates’ work.
we just brought it to Offloop, where you can try it free with your team.
offloop.org/models/kimi-k3-1…
Kimi K3.1 Pro is alive on Offloop.
-> 1M context.
-> $2 per million input tokens and $10 per million output tokens
-> 65.7 on Terminal-Bench 4.0. 51.7 on SWE.
Meet your team’s new Kimi here at offloop.org
Opus 5.5 vs Kimi K3.1 Pro vs DeepSeek V4 Flash vs GPT-6.1 Sol.
we tested all four with one long prompt: recreate Central Park on an autumn morning in Three.js 0.170. pure code, no imported 3D models or textures.
>1 build a scene that holds up as the camera moves through it, including close-ups of Bow Bridge.
>2 make the lighting and movement believable, down to translucent leaves and people walking.
>3 keep it running at 30+ fps at 1280×720, with reflections and post-processing enabled.
same prompt, results side by side.
which one came closest to completing the whole task?
Offloop retweeted
i've been testing DeepSeek v4.1 Flash 1017 for the past few hours.
-> one shot, 31 minutes of autonomous execution
-> only $0.22
-> the promising part is how far it can carry a task from a brief description without me steering every step. this is the kind of long-running execution that makes “just describe what you want” coding feel useful.
this SpongeBob showcase is one of the results, a short prompt asking for a 30-second Three.js scene with water absorption, simulated physics, cinematic lighting and bubbles.
Make a single-file three.js scene that auto-plays a 30-second loop of SpongeBob under the sea soaking up water, with real simulated physics. Polished, cinematic Pixar-like look with soft lighting, god rays and bubbles.
We gave Gemini 4 Argon and Claude Opus 5.5 same prompt
Gemini 4 Argon ~ 3m14s, $0.31
Opus 5.5 ~ 15m21s, $2.47
~5x the time, ~8x the cost. Left or right?
With Gemini 4 Argon, your research has new ways to tell its story 📖
We used Nano Banana, Veo, and Argon to turn a dense academic PDF into a full-color, animated storybook in the @GeminiApp.
After half a day of failed attempts and retries with Gemini 4 Argon, finally tested the margarita sloshing effect with:
→ Gemini 4 Argon, $1.64
→ GPT 6 Astra, $4.90
→ GLM 5.3, $0.93
→ DeepSeek Flash, $0.24
They all used Three.js + WebGL.
Gemini has improved so muchhh 🍸
Tested with Offloop.
The design-system detail in the quoted post caught my eye. Reusing the product’s own components seems like a practical way to keep demos visually consistent.
I’d keep the source links attached when moving the research into Claude. It makes checking the final draft against the original material much easier.
Can you ask for a specific revision like “keep this take, shorten the pause” once the first edit is done? That’s the workflow I’d be curious to see next.
Can you review the generated image and request changes from the same Claude session? Keeping that feedback loop in one place would be useful.
A recurring task seems like a good place to start. Try it on something you already do every week, where you know what a useful result should look like.
The existing artwork plus a small game design document is a useful starting point. It gives Claude both a visual reference and rules to build around.
Being able to change the timing or swap a visual without starting over is the part that interests me most. So much of editing is those small adjustments.
The “where to save it and how to evaluate it” part is easy to overlook. It gives the agent a much clearer finish line than just describing the task.
A global shortcut makes so much sense for this. Copying text into another chat just to summarize it can feel like more work than reading it.
On a project this size, where did most of your time go—scoping the work, coordinating agents, or reviewing what they shipped? I’d love to see that breakdown alongside the line count.
Your next customer might already be in your pipeline.
Offloop’s agents review past conversations and spot new signals across the web, so you can follow up as soon as an opportunity opens up.
Try it at offloop.org
